Since 1987 - Covering the Fastest Computers in the World and the People Who Run Them
Since 1987 - Covering the Fastest Computers in the World and the People Who Run Them
MLCommons’ latest MLPerf Training results (v2.0) issued today are broadly similar to v1.1 released last December. Nvidia still dominates, but less so (no grand sweep of wins). Relative newcomers to the exercise – AI chip/system makers Graphcore and Habana Labs/Intel – along with Google again posted strong showings. Google had four top scores basically splitting wins with Nvidia in the “available” category. The number of submitters grew to 21 from 14 in December. Overall, the results of this round were about 1.8x better performance than the last round.
This may be what early success looks like for MLPerf, the four-year-old AI technology benchmarking effort created by and run by MLCommons. The latest round of repeat submitters included Azure, Baidu, Dell, Fujitsu, GIGABYTE, Google, Graphcore, HPE, Inspur, Intel-Habana Labs, Lenovo, Nettrix, Nvidia, Samsung and Supermicro. First-time MLPerf Training submitters were ASUSTeK, CASIA, H3C, HazyResearch, Krai, and MosaicML. The most notable absentee, perhaps, was Cerebras Systems which has consistently expressed disinterest in MLPerf participation.
“MLPerf represent a diverse set of real world use cases,” said Shar Narasimhan, Nvidia director of product management for accelerated computing. “Speech recognition and language processing models are key along with recommenders, reinforcement learning and computer vision. All of the [training] benchmarks remain the same as the previous round, with the exception of object detection lightweight, which has been improved.
“We had 21 companies submit into the various MLPerf benchmarks in this particular round and they had over 260 submissions. It’s a real increase in participation and speaks well for the benchmark as a whole. The 21 submitters used four different accelerators from Google to Graphcore to Habana and Nvidia. We’re excited to see that 90 percent of the MLPerf submissions in this particular round use the Nvidia AI platform,” said Narasimhan.
Google got into the act, reporting, “Today’s release of MLPerf 2.0 results highlights the public availability of the most powerful and efficient ML infrastructure anywhere. Google’s TPU v4 ML supercomputers set performance records on five benchmarks, with an average speedup of 1.42x over the next fastest non-Google submission, and 1.5x vs our MLPerf 1.0 submission. Even more compelling — four of these record runs were conducted on the publicly available Google Cloud ML hub that we announced at Google I/O.” One Google win was in the research, not the available, category.
Nvidia (and Google) had plenty to crow about – top performance in four categories – but the MLPerf training exercise is starting to have a very different feel – perhaps less dramatic but also more useful.
Because system configurations differ widely and because there are a variety of tests, there is no single MLPerf winner as in the Top500 list. Even when Nvidia was sweeping every category – driven by its formidable technical strength and lack of competitors – it was always necessary to dig into individual system submissions to make fair comparisons among them. Nvidia GPUs were really the only accelerator game in town. Now, the (slowly) growing number ML accelerators participating in MLPerf, their improving performance versus Nvidia, and MLPerf’s effort to regularly freshen its test suite are transforming MLPerf into a broader lens through which to assess different AI accelerators and systems including cloud instances.
The analyst community also seems to think MLPerf is carving out a durable, useful role.
Judging performance using the MLPerf benchmarks is still tricky. For example, in the top chart above from Nvidia, results for the absolute time-to-train a model by a system (or cloud instance) is shown. Google TPUv4 turned in the top performances on BERT and ResNet-50. Looking just at ResNet-50 results directly from the MLPerf spreadsheet, the two Google instances (tpu-v4-8192 and tpu-v4-6912) had 4096 and 3456 TPU chips, respectively, and delivered times of 0.191 and 0.230 minutes. An Nvidia DGX with 4216 NVIDIA A100-SXM-80GB GPUs delivered 0.319-minute time. A Graphcore system (Bow-Pod256) with 256 Bow IPUs took 2.672 minutes.
Obviously, differing system size and number of accelerators is a big factor in time-to-train. Often so is the software stack. For this latest training round, Nvidia took a stab at normalizing performance on a per-chip (slide below, *method explained at end of article) basis and says it had the fastest per-chip performance on six of the tests. The point here is it takes a lot of care to compare systems using MLPerf results. Fortunately, MLPerf has made the results spreadsheet readily available and relatively easy to slice and dice.
By way of background, MLPerf was launched in 2018 with training as its first exercise. It has been busily adding benchmark suites covering different areas (e.g. inference) with most being run twice-yearly (see chart below). The lineup now also includes four inference suites and an HPC training exercise. Currently, MLPerf is developing an AI-relevant storage benchmark it hopes to launch in 2022. How all of these different benchmarks fare long-term is unclear, but their very emergence is more evidence of the AI technology wave rolling through IT, including HPC.
For this round of training, MLPerf added a new object detection benchmark that trains the new RetinaNet reference model on the larger and more diverse OpenImages dataset. MLPerf says the new test more accurately reflects state-of-the-art ML training for applications such as collision avoidance for vehicles and robotics as well as retail analytics. “I’m excited to release our new object detection benchmark, which was built based on extensive feedback from a customer advisory board and is an excellent tool for purchasing decisions, designing new accelerators and improving software,” said David Kanter, executive director of MLCommons, parent organization of MLPerf.
Each set of benchmarks typically has two divisions: “The Closed division is intended to compare hardware platforms or software frameworks ‘apples-to-apples’ and requires using the same model and optimizer as the reference implementation. The Open division is intended to foster faster models and optimizers and allows any ML approach that can reach the target quality,” states MLCommons.
Overall, said Kanter, “If we compare the best results of this round to sort of the high watermark previously, there’s about 1.8x better performance. So that’s a pretty significant move forward.”
Within the divisions, MLPerf divides benchmark results into Categories based on availability as bulleted here:
(MLPerf organizers invite participants to submit statements describing their systems and any attributes or tuning that optimizes them for the AI suite under test. HPCwire has included those statements at the end of the article – they range from the useful to the usual market-speak so read them accordingly.)
Without doubt Nvidia remains the dominant player with ~90 percent of submitted systems in Training v2.0 using Nvidia accelerators. The company chose not to try to rush its recently announced Hopper GPUs into the latest round. Given the new Hopper Tensor Cores have the capability to apply mixed FP8 and FP16 precision data types, it seems likely that the next round of MLPerf will feature them.
Narasimhan said, “We wanted to focus on what was commercially available and that’s why we submitted on the A100. This is our fourth time submitting on the A100 Tensor Core GPU. We introduced it two years ago on MLPerf 0.7 training in July of 2020 and we’re now doing the fourth submission on the same GPU. It remains the only accelerator to submit on every single benchmark. Google submitted on four different benchmarks, and Graphcore and Habana both submitted on just two benchmarks. There were no CPU submissions in this particular round.”
Nvidia, Graphcore and Habana Labs all held separate briefings. Nvidia, perhaps predictably, emphasized the value of performing all of the benchmarks as proof of a system’s ability to handle typical AI workflows, which Narasimhan said often include use of many different AI models.
He also argued that scale matters, “Some of these submissions take place at scale, particularly those from Nvidia and Google where we’re using thousands of accelerators in a particular submission. That’s why you see some of the training times are really short for some of the Google’s submissions. Other participants were smaller scale and so they have significantly larger times. But the importance of submitting at scale cannot be understated. It allows us to train new, more complex models, especially some of the new large language models.”
Graphcore’s a Different Beast
Graphcore, which touts itself as the only differentiated processor in the MLPerf mix, promoted its improving performance and the fact that its software ecosystem – necessarily a little different lower in the stack – is mature and able to work typical AI applications. In March, the company launched Bow, its third-generation Intelligence Processing Unit (IPU) featuring wafer-on-wafer technology. The new device is manufactured using a new variant of TSMC’s 7nm process that enables wafer-on-wafer (WoW) packaging. (See HPCwire coverage, Graphcore Launches Wafer-on-Wafer ‘Bow’ IPU)
The IPU MIMD (multiple instruction, multiple data) architecture differs from SIMD (single instruction multiple data of CPUs, GPUs and TPUs. “The IPU offers true MIMD parallelism and has distributed, local memory as its only form of memory on the device,” says Graphcore.
Graphcore held a separate briefing on the latest MLPerf round. Matt Fyles, senior vice president, software, said “We have built a different processor than Nvidia, Google, and Intel in this kind of competition. I think it’s interesting for two reasons. One is we had to do quite a lot more work to make what is an essentially performance-optimized multiple instruction multiple data processor work in [SIMD] fashion. That’s really a testament to the way in which a lot of the models are working today; it won’t necessarily be the same as what happens in the future.
“Really, for us, MLPerf is about making sure we can deliver the same kind of applications that are that are common on other platforms, and also to show that we can also perform well…it’s useful because it puts us in the GPU club, we can show the same models running. When we’re looking at what we’re doing with our customers, there’s obviously a lot more to do, which is showing this differentiation, showing that we’re looking at doing different things. Because otherwise, it is just a you know, a race to the bottom against Nvidia and that’s not going to be a good race for a lot of companies”
He also noted, “It’s interesting how different platforms get different levels of kind of scrutiny because of being very different or not being very different. For example, I think in this round in particular, platforms that were more GPU didn’t require an audit, whereas you know, because we’ve done something very different people are obviously very skeptical about does it work? How have you done this and so we end up with a bit more scrutiny on what we’ve done, which I think is fair.”
Fyles presented slides showing Graphcore’s improved performance since the training v1.1 exercise and also took aim at Nvidia (slides below). Also noteworthy, Baidu had Bow-based instance submissions that used Baidu’s PaddlePaddle AI platform, roughly equivalent to Google TensorFlow platform.
Habana’s Gaudi2 Arrives Ready
Habana/Intel also brought a new chip to the MLPerf training exercise, Gaudi2, launched in May. (See HPCwire coverage, Intel’s Habana Labs Unveils Gaudi2, Greco AI Processors). Gaudi2 has 24 Tensor processor cores, up from 10, and the in-package memory capacity has tripled from 32GB (HBM2) to 96GB (HBM2E) and the on-board SRAM has doubled from 24MB to 48MB. “That’s the first and only accelerator that integrates such an amount of memory,” Eitan Medina, COO of Habana Labs, told HPCwire at the time.
The company issued a blog today on its MLPerf showing, “Compared to our first-generation Gaudi, Gaudi2 achieves 3x speed-up in Training throughput for ResNet-50 and 4.7x for BERT. These advances can be attributed to the transition to the 7nm process from 16nm, tripling the number of Tensor Processor Cores, increasing our GEMM engine compute capacity, tripling the in-package high bandwidth memory capacity and increasing its bandwidth, and doubling the SRAM size. For vision models, Gaudi2 has another new feature, an integrated media engine, which operates independently and can handle the entire pre-processing pipe for compressed imaging, including data augmentation required for AI training.”
Medina led the press/analyst pre-briefing and said, “The other aspect that we’d like to emphasize here is that this is a significant boost to the to the first generation, but in a key aspect to our software development, we are using the same commercially available software for the MLPerf submission that customers that use our SynapseAI software suite will be able to see similar results to what we submitted for MLPerf. We believe this is really important to make sure customers get out of the box performance that would be representative to what they get using our commercial software.
“With regard to submission, we haven’t yet used all the means and all the improvements we have to the architecture. For instance, we haven’t used yet, the new FP8 data type that introduced into Gaudi2. And with gaudi2 two, we’re able to, of course, expand that to the compute power of the architecture, with a much more powerful GEMM engine, tripling the number of the tensor processor cores, much higher memory bandwidth and memory capacity with 96 gigabytes of memory capacity, as well as the higher integration of the RoCE engines in the pipe.”
Interestingly, Habana has no plans to submit in the MLPerf HPC Training exercise, said Medina, “As a company, our mission is to really focused on deep learning acceleration. And I think that will still be the primary focus. We are looking at some intersection points between HPC and AI workloads to find out if accelerating AI can fit within an overall end to end workload, so I cannot say definitely how we will evolve. Right now, our primary focus is still on deep learning – the primary model families, vision models and language models [and] we’re expanding that to recommendation systems and even larger models.”
Google is, of course, long-term pioneer in driving AI technology. It’s more public participation in MLPerf is further evidence of the benchmarking organizations increasing value. Google issues a blog (Cloud TPU v4 records fastest training times on five MLPerf 2.0 benchmarks), written by Naveen Kumar, principal engineer, and Vikram Kasivajhula, Director, Product Management, ML Infrastructure.
“Our 2.0 submissions, all running on TensorFlow, demonstrated leading performance across all five benchmarks. We scaled two of our submissions to run on full TPU v4 Pods. Each Cloud TPU v4 Pod consists of 4096 chips connected together via an ultra-fast interconnect network with an industry-leading 6 terabits per second (Tbps) of bandwidth per host, enabling rapid training for the largest models.
“Hardware aside, these benchmark results were made possible in no small part by our work to improve the TPU software stack. Scalability and performance optimizations in the TPU compiler and runtime, including faster embedding lookups and improved model weight distribution across the TPU pod, enabled much of these improvements, and are now widely available to TPU users. For example, we made a number of performance improvements to the virtualization stack to fully utilize the compute power of both CPU hosts and TPU chips to achieve peak performance on image and recommendation models.”
Google also took aim at Azure. “Cloud TPU’s industry-leading performance at scale also translates to cost savings for customers. Based on our analysis summarized in Figure 3 (below), Cloud TPUs on Google Cloud provide ~35-50 percent savings vs A100 on Microsoft Azure (see Figure 3). We employed the following methodology to calculate this result,” wrote Kumar and Kasivajhula.
Lastly, it’s worth mentioning that one submission – from HazyResearch, Stanford University research group – is actually just a single graduate student according to Kanter, “HazyResearch is one of the most intriguing submitters because they are the smallest submitter. [It’s] actually the work of a single graduate student. One of the things to me that was really impressive is, you know, the MLPerf benchmarks are complicated, but even so a single graduate student can get them to run and make a good submission.”
Link to MLCommons/MLPerf announcement, https://www.hpcwire.com/off-the-wire/mlperf-results-highlight-more-capable-ml-training/
Link to Nvidia blog, https://blogs.nvidia.com/blog/2022/06/29/nvidia-partners-ai-mlperf/
Link to Google blog, https://cloud.google.com/blog/products/ai-machine-learning/cloud-tpu-v4-mlperf-2-0-results
Link to Graphcore blog, https://www.graphcore.ai/posts/baidu-results-underscore-big-mlperf-gains-for-graphcore
Link to Habana/Habana announcement, https://www.hpcwire.com/off-the-wire/intel-announces-mlperf-results-for-habana-gaudi2/
* Method for calculating per chip performance from Nvidia slide: “To calculate per-chip performance, this chart normalizes every submission to the closest scale of the fastest competitor. The fastest competitors are shown with1x. To determine the fastest competitor, we selected the scale common to most submitters.| Format: Chipcount, Submitter, MLPerf-ID | ResNet-50: 8xNVIDIA2.0-2069, 3456x Google 2.0-2011,16x Graphcore 2.0-2047, 8x Intel-HabanaLabs 2.0-2073 | BERT: 8x Inspur 2.0-2070, 3456x Google 2.0-2011,16x Graphcore2.0-2045, 8xIntel- HabanaLabs 2.0-2073 | DLRM: 8x Inspur2.0-2068 | MaskR-CNN: 384x NVIDIA 2.0-2099,1024x Google 2.0-2009 | RetinaNet: 1280x NVIDIA 2.0-2103, 2048x Google2.0-2010 | RNN-T 8x Inspur 2.0-2066 | 3D-UNet: 8xH3C2.0-2060,| MiniGo: 8xH3C2.0-2059 | MLPerf name and logo are trademarks. See www.mlperf.org for more information.”
Vendor Submitted Statements from MLPerf
ASUS is proud to have enrolled in MLCommons’ influential MLPerf Training 2.0 submission for 2022, allowing our products to be judged for their performance in the competitive field of AI training.
Specifically, we submitted the ASUS ESC N4A-E11 server because we consider it to be the perfect choice for diverse AI workloads in simulation and data analytics, including virtualization, medical diagnoses and voice recognition.
With ESC N4A-E11, ASUS has expertly leveraged the CPU and GPU architecture to maximize performance from powerful combination of the AMD EPYC 7003 and NVIDIA HGX platforms. Our optimized hardware design paired with ASUS-exclusive, firmware-based tuning technology speeds up overall efficiency and precision in AI training models. In fact, ESC N4A-E11 is designed especially for high-performance computing (HPC) fields, and has been deployed in numerous HPC projects globally this year alone.
The ESC N4A-E11 architecture, which is engineered for up to four NVIDIA HGX GPUs and a single, 64-core CPU, delivers leading results in multiple training models — including BERT, SSD, RNNT and UNET3D. These models provide a good path for deployment of training models into daily operations in enterprise or data centers.
As demonstrated by the latest MLPerf Training 2.0 benchmarks, ESC N4A-E11’s cutting-edge performance and software combine to create an end-to-end solution that can be deployed data centers, in the cloud or at the edge — empowering remarkable results.
ASUS as an integrated-solutions partner delivering leading hardware for the fields of supercomputing and data centers, supported by an extensive AI portal and AI software stack. The MLPerf benchmarks help our engineers to develop an ever-better understanding of AI applications and performance needs — and we’re excited to continuing delivering innovations to both shape and grow MLCommons.
Azure is pleased to share the results of our MLPerf training v2.0 submission. For this submission we used the NDm A100 v4 series virtual machines (VMs) powered by 8 NVIDIA A100 GPUs (80 GB), 8 NVIDIA 200 Gb/s HDR InfiniBand cards, 96 AMD Rome cores, 1.9 TB of RAM, and 8 * 1TB NVMe disks. The NDm A100 v4 allows for high-end AI training needs from 1 – 256+ VMs (8 – 2048+ GPUs).
These benchmark results demonstrate how Azure:
To generate the results, we used Azure CycleCloud to orchestrate the cluster environment of 16 VMs. In our previous submission for MLPerf training v1.1 we used it to orchestrate over 264 VMs. We used the Slurm scheduler configured with NVIDIA Pyxis and Enroot to schedule the NVIDIA NGC ML Commons containers***. This setup enabled us to deploy our environment in a timely manner and perform the benchmarks with strong performance and scalability. For more information on how to deploy this setup please see cc-slurm-ngc.
The NDm A100 v4 series VMs are what we and our Azure customers turn to when AI and ML training is required. We are excited to see what new breakthroughs our customers will make using these VMs
*** Special thanks to the NVIDIA team for all their support
Baidu has been working on accelerating large-scale models to benefit more industries. These real-world applications are more complex than training the MLPerf reference models, due to the huge number of parameters, heterogeneous infrastructure and other factors. With MLPerf Training 2.0, we are pleased to present the performance from PaddlePaddle on large-scale models.
PaddlePaddle has continued optimizing on X-MAN, the open computing system powered by NVIDIA A100 80GB GPUs. We submitted the BERT benchmark results using both PaddlePaddle and PyTorch, showing that BERT on PaddlePaddle ranks among the fastest frameworks tested on NVIDIA A100 80GB GPUs.
We are happy Graphcore made a submission with PaddlePaddle on IPUs with outstanding performance. As for BERT training performance on IPUs, PaddlePaddle is in-line with Graphcore’s PopART framework. It shows PaddlePaddle’s hardware ecosystem is expanding, and PaddlePaddle performs excellently on more and more AI accelerators.
These remarkable results come from both optimizations within PaddlePaddle and ongoing collaboration with hardware vendors. Driven by industrial practices, the distributed training architecture of PaddlePaddle fully demonstrates its key attributes of low-storage and high-performance training with heterogeneous hardware. It leads to several featured innovations: the world’s first general-purpose heterogeneous parameter server architecture, the world’s first 4D hybrid parallel training technology, and an end-to-end adaptive distributed training architecture.
These innovations enabled the release of PCL-BAIDU Wenxin: the world’s first Knowledge-Enhanced 100-billion-scale pretrained language model, and a series of big industrial models, such as State Grid-Baidu·Wenxin, SPD-Baidu·Wenxin and HELIX-Fold. Moreover, PaddlePaddle is collaborating with hardware partners to provide excellent performance and user experiences. PaddlePaddle v2.3 was released recently with support for both NVIDIA GPUs and Graphcore IPUs. As such, PaddlePaddle brings more options for training large-scale models to the community.
Baidu-PaddlePaddle looks forward to accelerating large-scale models to benefit the world with innovative technologies in distributed training and cooperation with hardware providers.
CASIA (Institute of Automation, Chinese Academy of Sciences) constantly focuses on researches of intelligent science and technology, which has sufficient innovation capabilities of chip R&D, system integration and algorithm design. We have rich experiences in the field of finance, transportation, medical treatment, etc. Meanwhile, our institute maintains long-term cooperative relationships with Dell EMC, H3C, NVIDIA, etc.
As a new member and first-time submitter of MLCommons, CASIA participated in 5 tasks of MLPerf Training V2.0, and successfully submitted 7 results across 3 configurations of our high-density AI server with excellent performances. We are one of the first to release the single node 16 NVIDIA A100 server named Xiangxue-3B in MLPerf developed by CASIA and Chip-hop Ltd.(www.chiphop.cn). It supports up to 16x NVIDIA A100 GPUs, dual 64 Core AMD EPYC 7773X processors, etc., and provides 1280GB HBM2E, 4TB DDR4 system memory and 1.5GB CPU cache. Xiangxue-3B has the flexibility of various PCIE topologies, which can satisfy general and extensive application needs. The results demonstrate its extraordinary performances in all of the 5 different tasks of Minigo, SSD, ResNet, Mask R-CNN and 3D U-Net, supplying referential performance data to AI users. With its high GPU expanding capabilities, users can quickly accelerate the speed of training tasks without building up complicated distributed systems. Xiangxue-3B supports various types of PCIE accelerators, and all the 16 cards’ IO interfaces are available on the chassis, making it suitable for both high-density computation and IO upgrading.
CASIA is a comprehensive research institute and has top algorithm experts, software and hardware engineers. As the developer and user at the same time, we have deep understandings on the actual needs in the academic and applicationable scenarios. It is our capability to provide various types of customized software and hardware products, which can be developed efficiently and further upgraded in fast speed. We embrace more opportunities of innovative cooperation and will constantly support the development of MLCommons.
Put innovation to work with the right technology partner.
For data-driven decision making, Dell Technologies submitted 42 results across 12 system configurations. Results are available for single-node and multi-node PowerEdge XE8545, R750xa and DSS8440 servers with NVIDIA A100 SXM 40GB and A100 SXM 80GB, A100 PCIe 80GB and A30 GPUs on MLPerf training models.
Dig into the Dell engineering test results. Test for yourself in one of our worldwide Customer Solution Centers. Collaborate with our HPC & AI Innovation Lab and/or tap into one of our HPC & AI Centers of Excellence.
Fujitsu is a leading company of information and communication technology systems. We support our customer’s business by providing robust and reliable ICT systems.
In this round, Fujitsu measured seven benchmark programs except for minigo with our latest server, PRIMERGY GX2570M6. The system is 4U rack mount and has an NVIDIA HGX A100 board that includes eight GPUs. As for storage, we configured RAID0 with PCIe connected NVMe SSDs to feed enough data during training. Compared with our past submissions, the performance is proportional to the number of GPUs.
Our purpose is to make the world more sustainable by building trust in society through innovation. We have a long heritage of bringing innovation and expertise, continuously working to contribute to the growth of society and our customers.
GIGABYTE is an industry leader in high-performance servers, and uses hardware expertise, patented innovations, and industry leadership to create, inspire, and advance. With over 30 years of motherboard manufacturing excellence and 20 years of server and enterprise products, GIGABYTE offers an extensive portfolio of enterprise products.
Over the years, GIGABYTE has submitted benchmark results for both Training and Inference. As well, the submitted GIGABYTE servers were equipped with various brands of accelerators (NVIDIA and Qulacomm) and CPUs (AMD, Ampere, and Intel) in configurations to showcase systems that target different markets (x86 and Arm).
GIGABYTE submissions for MLPerf Training v2.0: – GIGABYTE 4U server: G492-ID0 – Dual 3rd Gen Intel Xeon Scalable processors – Platinum 8380 – NVIDIA HGX A100 80GB 8-GPU system – Frameworks: mxnet, Tensorflow, hugectr, and PyTorch
GIGABYTE will continue optimization of product performance to provide products with high expansion capability, strong computational ability, and applicable to various applications at data centers. GIGABYTE solutions are ready to help customers upgrade their infrastructure.
In MLPerf v2.0, Google’s submissions demonstrated exceptional performance across all five of the benchmarks we submitted.
Four of Google’s submissions were in the cloud category based on our publicly available Cloud TPU v4. This means that all of Google’s ML infrastructure is available to ML practitioners and innovators around the world.
Two of our submissions this time around were at the “full TPU v4 pod” 4096 chip scale with each pod delivering 1.1 exaflop/s of peak performance. As we’ve noted before, each Cloud TPU v4 pod consists of 4096 chips connected together via an ultra-fast interconnect network with the equivalent of an industry-leading 6 terabits per second (Tbps) of bandwidth per host, enabling rapid training for the largest models.
Google’s TPU 2.0 submissions were made on Tensorflow and are a significant improvement over our 1.0 submissions last year, reflecting the significant investments in our software stack that we have made, including optimizations made to the TPU compiler.
Finally, it is worth noting that the Cloud TPU v4 pods powering our MLPerf results are located in our Oklahoma data center which operates at 90% carbon-free energy. In fact, with a PUE of 1.10, the Oklahoma data center is one of the most energy efficient data centers in the world. Additionally, the TPU v4 chip itself is highly energy efficient with 3x the peak FLOPs per watt of the prior v3 generation.
Since the last MLPerf Training round, Graphcore has focused on building out breadth and depth in our model garden of real-world applications driven by customers. These include emerging models, like GNNs and graph networks which are well-suited to the IPU’s highly differentiated, designed-for-AI architecture, large Transformer NLP models like GPT and more accurate vision models like EfficientNet.
With MLPerf Training 2.0, we are pleased to show continued significant performance improvements for MLPerf reference models BERT and ResNet and a new submission for speech transcription, Transformer-Transducer. Other highlights include the first PaddlePaddle submission from Baidu on IPU and the introduction of our new Bow Pod systems, straight into the available on-prem and cloud category.
In March 2022, we launched a new Bow IPU processor and Bow Pod platforms, delivering up to 40% better performance and up to 16% better power efficiency. We are delighted to show results for our new Bow systems, just a few months later, from Bow Pod16 through to our large Bow Pod256 system, designed for large scale distributed training. These systems all feature in the MLPerf ‘available’ category as they are already shipping to customers and are available in the public cloud.
Our software continues to go from strength to strength, with additional software improvements on top of those delivered by our new Bow systems, delivering an overall 58% increase in BERT performance and a 44% increase in ResNet-50 performance.
As our software matures, our software ecosystem expands as new partners are able to integrate with our open-source APIs and libraries in the Poplar SDK. We are delighted that Baidu has made a submission with its PaddlePaddle framework using Graphcore IPUs for the first time and has received outstanding results. The submissions for BERT on Bow Pod16 and Bow Pod64, achieve near identical performance to the results submitted for Graphcore’s PopART framework. This is a testament to the consistent performance that can be achieved across multiple frameworks on the Bow IPU platforms, and the efficiency in model development using PaddlePaddle.
We made a new submission for a speech transcription transformer-transducer model in this round. Working with our customer, Gridspace, we applied this 100M parameter model to accelerate training on real world data – over 700GB/10k hours of speech – to a state-of-the-art word error rate. By scaling the model to run on a Bow Pod64 training time on this dataset was reduced from weeks to days.
As a digital solution leader, H3C is committed to providing intelligent software management platform, diversified smart computing system and unified artificial intelligence platform for computing power scheduling, so as to achieve easier product deployment, more reliable operation, more powerful monitoring and more convenient troubleshooting, and escort the production of AI customers.
Complex deep learning models often consume lots of computing and storage resources. Different AI workloads have different requirements for hardware bandwidth, topology, single-point computing power and the like due to their diverse capabilities in data processing, data storage and data interaction. H3C introduces H3C UniServer 5500 G5 8HGX A100 server model (medium and large scale), R5300 G5 PCIe GPU server model (medium and large scale) and 2U R4900 G5 universal server model (small scale) for different AI workload scenarios, in order to meet customers’ demand for computing power resources in different scenarios.
In this benchmark test of MLPerf Training v2.0, H3C participated in the training evaluation for eight task with six types of models, including R5500 /R5300/R4900 G5, and has made excellent achievements in natural language processing, target detection, medical image processing and other fields.
We’re pleased to deliver Habana’s third training submission results, including the recently launched second-generation Gaudi®2 deep learning training processor, our new purpose-built AI processor that further enhances Intel’s AI XPU portfolio. Gaudi2 was launched in May and shows dramatic advancements in time-to-train, which translate to 3.4x speed-up for the vision model (ResNet-50) and 4.9x for the language model (BERT) relative to first-gen Gaudi performance. These advances can be attributed to the transition to 7nm process from 16nm, the Gaudi2 memory subsystem which contains 96 GB of HBM2E memories delivering 2.45 TB/sec bandwidth and additional architecture advances. The quick and seamless launch of Gaudi2 and our ability to submit Gaudi2 in this current MLPerf evaluation were enabled due to the continued maturity of our SynapseAI software stack.
In addition, we submitted performance of 128 and 256 accelerator configurations of our first-generation Gaudi solution that demonstrate great performance and impressive linear scaling. The performance of both generations of Gaudi processors is achieved without product enhancements or special software manipulations that differ from our commercial solutions available to Habana customers out of the box. As a result, customers can expect to achieve MLPerf-comparable results in their own Gaudi systems. Both generations of Gaudi have been designed at inception to deliver exceptional AI deep learning efficiency—so we can provide customers with the excellent performance reflected in these MLPerf results, while maintaining very competitive pricing.
In addition, we are pleased that Supermicro submitted Gaudi’s first OEM server results, based on the first-generation Gaudi training solution. We look forward to continuing to advance performance, efficiency, and end-user ease of use with the Gaudi platform and continue to support the MLPerf organization, our partners and our peers in leading the future of AI.
Transformer models (e.g., BERT, GPT) have emerged as the most widely used architecture in applications such as natural language processing. Transformers have grown larger and deeper, but equipping them with longer context remains difficult, since the self-attention module at their heart has time and memory complexity quadratic in sequence length. Many approximate attention methods aimed to alleviate these issues do not display wall-clock speedup against standard attention, as they focus on FLOP reduction and tend to ignore overheads from memory access (IO).
We argue that a missing principle is making attention algorithms IO-aware–carefully accounting for reads and writes to different levels of fast and slow memory (e.g., between fast GPU SRAM and relatively slow GPU HBM). On modern GPUs, compute speed has out-paced memory speed, and most operations in Transformers are bottlenecked by memory accesses.
We propose FlashAttention, a new algorithm that computes exact attention with far fewer HBM accesses. This requires (i)computing softmax without access to the whole input (ii) not storing large intermediate matrices for the backward. We apply two well-established techniques to address these challenges. (i) We split the input into blocks and incrementally perform the softmax (tiling). (ii) We store the softmax normalization from the forward to quickly recompute attention in the backward.
Our algorithm both runs faster (4x) and uses less memory (10x) than Pytorch standard attention, thanks to the massively reduced amount of HBM access. This brings substantial speedup and memory saving to many Transformer models (BERT, GPT2, ViT). Our BERT submission to MLPerf 2.0 is 11% faster than the best one-node result from MLPerf 1.1.
We look forward to future work on applying these general techniques to other accelerators. We’re really excited to see what new capabilities will emerge from being able to use longer context in Transformers.
HPE makes AI that is data-driven, production-oriented, and cloud-enabled, available anytime, anywhere and at any scale. We understand that successfully deploying AI workloads requires much more than hardware. That’s why we deliver a full complement of offerings that enable customers to embark on their AI journey with confidence. Award-winning HPE AI Transformation Services make some of the brightest data scientists in the industry available to assist with everything from planning, building, and optimizing to implementation, and we offer continuing support through PointNext.
Built upon the widely popular open source Determined.AI Training Platform, the Machine Learning Development Environment (MLDE) software and the Machine Learning Development System (MLDS) integrated hardware and software solution from HPE help developers and scientists focus on innovation by removing the complexity and cost associated with machine learning model development. Our platform accelerates time-to-production by removing the need to write infrastructure code, and makes it easy to set-up, manage, secure, and share AI computing clusters. With MLDE, customers are training models faster, building more accurate models, managing GPU costs, and tracking experiments, while HPE’s Greenlake provides seamless integration of cloud and on-premises data hosting and analysis.
Today we are publishing MLPerf Training results based on the HPE Apollo 6500 Gen 10+. For these results, dual AMD EPYC 7763 processors and eight NVIDIA HGX A100 80GB GPUs delivered leading results across multiple categories, including image detection, classification, and segmentation. Mounted within our newest internal cluster to enter the Top500, Champollion, the Apollo 6500 Gen 10+ is a highly flexible system that supports either modular or PCIe GPUs from multiple vendors and dedicated workload profiles that allow users to optimize for power or throughput. As a founding member of MLCommons, HPE is committed to delivering benchmark results that provide our customers with guidance on the platforms best suited to support a variety of workloads.
Inspur Electronic Information Industry Co., LTD is a leading provider of data center infrastructure, cloud computing, and AI solutions, ranking among the world’s top 3 server manufacturers. Through engineering and innovation, Inspur delivers cutting-edge computing hardware design and extensive product offerings to address important technology arenas like open computing, cloud data center, AI, and deep learning.
In MLCommons TrainingV2.0, Inspur made submissions on two systems: NF5488A5 and NF5688M6.
NF5488A5 is Inspur’s flagship server with extreme design for large-scale HPC and AI computing. It contains 8 A100-500W GPUs with liquid cooling. NF5488A5 system is capable of high temperature tolerance with operating temperature up to 40℃. It can be deployed in a wide range of data centers with 4U design, greatly helps to lower cost and increase operation efficiency.
NF5688M6 based on 3rd Gen Intel® Xeon® scalable processors increases performance by 46% from Previous Generation, and can support 8 A100 500W GPUs with air cooling. It accommodates more than 10 PCIe Gen4 devices, and brings about a 1:1:1 balanced ratio of GPUs, NVMe storage and NVIDIA Mellanox InfiniBand network.
In the closed division, Inspur performance of Bert, RNNT, DLRM, Resnet, MaskRCNN and UNet3D are improved by 18.15%, 13.84%, 5.95%, 1.24%, 10.40% and 7%, compared with the best performance Inspur achieved in Training v1.1.
Comparing “entry-level” options for training neural networks is of interest to many organizations with limited resources. This is why Krai, a staunch supporter of MLPerf Inference, is making a foray into MLPerf Training in this round.
Our ResNet-50 result of 284 minutes on two NVIDIA RTX A5000 GPUs consuming up to 460 Watts can be usefully compared with 236 minutes on two NVIDIA A30 GPUs consuming up to 330 Watts. Our result consumes 39% more power and is 20% slower, which translates into 67% more energy consumed. However, the A5000 is 2-3 times cheaper than the A30, making the less efficient option still attractive for small organizations. Furthermore, taking into account the power of other system components (say, 200 Watts), the total energy consumed may only be 25% higher.
Another interesting comparison is with 396 minutes on a single NVIDIA A100 PCIe 80GB GPU consuming up to 300 Watts. While our result consumes 53% more power, it is 28% faster, which translates into only 10% more energy consumed. However, a pair of A5000 GPUs is 3-4 times cheaper than a single A100 PCIe 80GB.
Seymour Cray once famously quipped: “If you were plowing a field, which would you rather use? Two strong oxen or 1024 chickens?” We at Krai are not arguing with Cray: we are not expecting to deliver training on a thousand Raspberry Pi’s any time soon. Still, our motto is “horses for courses”: using horses may be just fine if oxen are hard to come by (or buy), and elephants live on an entirely different continent.
MosaicML, a startup on a mission to make ML training efficient for everyone, is pleased to share our MLPerf Training 2.0 results. Our results demonstrate the ability to accelerate training through software and algorithms, and the ease with which enterprise ML teams can realize more efficient training.
Our algorithmic efficiency methods achieve a 4.5x speed-up on training the Image Classification benchmark compared to our own baseline. Image Classification benchmark (open division) is trained in 23.8 minutes on 8x A100 NVidia GPUs, compared to our baseline without efficiency methods taking 110.5 minutes.
Our submission uses our open source training framework Composer, built on PyTorch, instead of heavily customized benchmark-specific code. Enterprise ML engineers can easily apply these same techniques to their own datasets. With just a few lines of code, train faster on existing hardware, or use the MosaicML Cloud, our purpose built platform for efficient ML training, for an optimized experience.
We believe that algorithmic efficiency is key to making training of large scale models more accessible to the broader community, and our submission to MLPerf is a first step in fulfilling that goal.
We submitted to the Image Classification benchmark in the Open division with two configurations:
MosaicML believes in lowering the barrier to MLPerf benchmarks, and our software includes an MLPerf logger plug-in to automatically generate compliant submission logs. We invite the broader community to use our provided tools to further optimize the algorithms and make their own submissions to future MLPerf benchmarks.
Overall, NVIDIA continued to show great performance across every MLPerf Training 2.0 workload in this most recent round of the industry-standard benchmark for AI training.
Our NVIDIA A100 Tensor Core GPUs again delivered exceptional performance and efficiency results across all tests, as demonstrated in our own submissions as well as those by over a dozen of our partners. These include submissions by cloud provider Microsoft Azure, as well as system makers ASUSTek, Baidu, CASIA, Dell, Fujitsu, Gigabyte, H3C, Hewlett-Packard Enterprise, Inspur, Lenovo, Nettrix, and Supermicro.
Our submissions on all eight tests ran the gamut from single-node all the way up to 4,216 GPUs across 527 nodes – the largest at-scale submission seen in this round. Our platforms delivered excellent per-GPU AI training acceleration, while our at-scale submission showcased not only our multi-node GPU scalability, but also our full datacenter platform including our InfiniBand networking technology.
Software plays a large role in delivering results both in single-node and at-scale. Submissions this round brought up to 13% more performance than our same system submitted in the previous round, which highlights the ongoing benefit customers get from our continuous optimization of our software tools. Beyond MLPerf, NVIDIA accelerates many additional models, which are freely available to developers both from NGC as well as our GitHub repository.
We congratulate our 13 partners on their outstanding submissions, and commend MLCommons for their ongoing work on these industry-standard benchmarks. We look forward to upcoming rounds of MLPerf testing for AI inference, HPC and training. MLPerf’s broad set of tests spans a wide array of today’s most popular AI workloads and scenarios. These diverse results help IT decision makers towards data-driven platform investments best suited to their particular needs.
This is Samsung’s second participation in MLPerf Training with better performance than the previous round v1.1. We delivered an extremely strong performance on BERT training, 22.3 seconds on 1024 Nvidia A100 GPUs and 21.4 seconds on 1368 Nvidia A100 GPUs. This is a 12% improvement TTT (Total Time on Test) over our v1.1 performance in 1024 GPUs (25.06 seconds).
The system used for BERT training consists of 171 nodes, which have two AMD EPYC 7543 processors and eight NVIDIA Tesla A100s as accelerators, which are connected with NVLinks and have their own 80GB memory in HBM.This system is the same as the previous round v1.1, but we just expanded the scalability from 128 nodes to 171 nodes. All hardware and software components we used are available in public, so we changed our system status from research to on-premise.
Based on PyTorch NVidia Release 21.09, we changed the optimizer from LAMB to ADAM and focused on the large batch training with computation and communication overlap. We also tested BERT training with other optimizers, including the standard optimizer LAMB, but we could see ADAM optimizer showed the best performance in our approach.
In addition to AI acceleration in mobile devices, Samsung is actively researching on the scalable and sustainable AI computing. We will work to solve the scaling challenge between computing capability and memory/storage bandwidth through innovation in memory and storage computings such as HBM-PIM, CXL-Memory, and CXL-SSD.
Supermicro Supermicro has its long history of providing a broad portfolio of AI-enabled products for different use cases. In MLPerf Training v2.0, we have submitted results based on four high performance systems to address multiple compute intensive use cases, including medical image segmentation, general object detection, recommendation systems, and natural language processing.
Supermicro’s DNA is to provide the most optimal hardware solution for your workloads and services. For example, we provide four different systems for NVIDIA’s HGX A100 8 GPU platform and HGX A100 4 GPU respectively. Customers can configure the CPU and GPU baseboards based on their needs. Furthermore, we provide upgraded power supply versions to give you choices on using our cost-effective power solutions or genuine N+N redundancy to maximize your TCO. Supermicro also offers liquid cooling for HGX based-systems to help you deploy higher TDP GPU baseboards without thermal throttling. If customers are looking for rack scale design to cluster systems for large machine learning training problems, we can offer rack integration in air cooled solution, RDHx and DLC liquid cooling solution to suit your plug and play need.
Supermicro’s SYS-420GP-TNAR, AS-4124GO-NART, AS-2124GQ-NART and upcoming SYS-220GQ-TNAR with NVIDIA’s HGX A100 GPUs can pass data directly from GPU to GPU, to avoid the pass-through overhead from processors and system memory. By shortening the data path to the accelerator, it shortens the training time for applications such as computer vision and recommendation system.
In addition, Supermicro keeps tuning the current platforms’ performance and aims to extend the benchmark efforts to large portfolios, including Habana, AMD, OpenVino, and Arctic Sounds. MLPerf Training v2.0 is the first time Supermicro submitted Image Classification on SYS-420GH-TNGR with Habana first generation GAUDI accelerators. With Supermicro 400Gbps switches, the MLPerf Training benchmark results show excellent performance and superior scalability to meet the challenges of distributed AI training on multiple nodes.
With multiple configurations of processors, accelerators, system form factors, cooling solutions, and scale out options, Supermicro would like to provide our customers the most comprehensive and convenient solutions to solve the AI problems. We are happy to see all the results we ran on MLPerf using our portfolio of systems, and we will keep optimizing the solutions for customer’s different requirements to help achieve the best TCO.
Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!
Oak Ridge National Laboratory’s exascale Frontier system may be stealing some of the spotlight, but the lab’s 148.6 Linpack petaflops Summit system is still churning out powerful science. Recently, for instance, the Read more…
In this monthly feature, we’ll keep you up-to-date on the latest career developments for individuals in the high-performance computing community. Whether it’s a promotion, new company hire, or even an accolade, we’ Read more…
HPCwire takes you inside the Frontier datacenter at DOE's Oak Ridge National Laboratory (ORNL) in Oak Ridge, Tenn., for an interview with Frontier Project Director Justin Whitt. The first supercomputer to surpass 1 exafl Read more…
You may be surprised how ready Python is for heterogeneous programming, and how easy it is to use today. Our first three articles about heterogeneous programming focused primarily on C++ as we ponder “how to enable programming in the face of an explosion of hardware diversity that is coming?” For a refresher on what motivates this question... Read more…
MLCommons’ latest MLPerf Training results (v2.0) issued today are broadly similar to v1.1 released last December. Nvidia still dominates, but less so (no grand sweep of wins). Relative newcomers to the exercise – AI Read more…
Hamburg-based Indivumed specializes in using the highest quality biospecimen and comprehensive clinical data to advance research and development in precision oncology. Its IndivuType discovery solution uses AWS to store data and support analysis to decipher the complexity of cancer. Read more…
Consumers use many accounts for financial transactions, ordering products, and social media—a customer’s identity can be stolen using any of these accounts. Identity fraud can happen when setting up or using financial accounts, but it can also occur with communications such as audio, images, and chats. Read more…
In February 2020, the United States’ National Oceanic and Atmospheric Administration (NOAA) announced that it would be procuring two HPE Cray systems, allowing the organization to triple its operational supercomputing capacity for weather and climate applications. Now, those efforts have come to fruition: NOAA has inaugurated the two systems, which are... Read more…
HPCwire takes you inside the Frontier datacenter at DOE's Oak Ridge National Laboratory (ORNL) in Oak Ridge, Tenn., for an interview with Frontier Project Direc Read more…
You may be surprised how ready Python is for heterogeneous programming, and how easy it is to use today. Our first three articles about heterogeneous programming focused primarily on C++ as we ponder “how to enable programming in the face of an explosion of hardware diversity that is coming?” For a refresher on what motivates this question... Read more…
MLCommons’ latest MLPerf Training results (v2.0) issued today are broadly similar to v1.1 released last December. Nvidia still dominates, but less so (no gran Read more…
In February 2020, the United States’ National Oceanic and Atmospheric Administration (NOAA) announced that it would be procuring two HPE Cray systems, allowing the organization to triple its operational supercomputing capacity for weather and climate applications. Now, those efforts have come to fruition: NOAA has inaugurated the two systems, which are... Read more…
With the Linpack exaflops milestone achieved by the Frontier supercomputer at Oak Ridge National Laboratory, the United States is turning its attention to the next crop of exascale machines, some 5-10x more performant than Frontier. At least one such system is being planned for the 2025-2030 timeline, and the DOE is soliciting input from the vendor community... Read more…
HPE's early stab at ARM servers close to a decade ago didn't pan out, but the company is hoping the second time is a charm. The company introduced the ProLiant RL300 Gen11 server, which has Ampere's ARM server processor. The one-socket server is designed for cloud-based applications, with the ability to scale up applications in a power efficient... Read more…
In this regular feature, HPCwire highlights newly published research in the high-performance computing community and related domains. From parallel programmin Read more…
You may recall that efforts proposed in 2020 to remake the National Science Foundation (Endless Frontier Act) have since expanded and morphed into two gigantic bills, the America COMPETES Act in the U.S. House of Representatives and the U.S. Innovation and Competition Act in the U.S. Senate. So far, efforts to reconcile the two pieces of legislation have snagged and recent reports... Read more…
Getting a glimpse into Nvidia’s R&D has become a regular feature of the spring GTC conference with Bill Dally, chief scientist and senior vice president of research, providing an overview of Nvidia’s R&D organization and a few details on current priorities. This year, Dally focused mostly on AI tools that Nvidia is both developing and using in-house to improve... Read more…
Intel has shared more details on a new interconnect that is the foundation of the company’s long-term plan for x86, Arm and RISC-V architectures to co-exist in a single chip package. The semiconductor company is taking a modular approach to chip design with the option for customers to cram computing blocks such as CPUs, GPUs and AI accelerators inside a single chip package. Read more…
In April 2018, the U.S. Department of Energy announced plans to procure a trio of exascale supercomputers at a total cost of up to $1.8 billion dollars. Over the ensuing four years, many announcements were made, many deadlines were missed, and a pandemic threw the world into disarray. Now, at long last, HPE and Oak Ridge National Laboratory (ORNL) have announced that the first of those... Read more…
AMD/Xilinx has released an improved version of its VCK5000 AI inferencing card along with a series of competitive benchmarks aimed directly at Nvidia’s GPU line. AMD says the new VCK5000 has 3x better performance than earlier versions and delivers 2x TCO over Nvidia T4. AMD also showed favorable benchmarks against several Nvidia GPUs, claiming its VCK5000 achieved... Read more…
The 59th installment of the Top500 list, issued today from ISC 2022 in Hamburg, Germany, officially marks a new era in supercomputing with the debut of the first-ever exascale system on the list. Frontier, deployed at the Department of Energy’s Oak Ridge National Laboratory, achieved 1.102 exaflops in its fastest High Performance Linpack run, which was completed... Read more…
The first-ever appearance of a previously undetectable quantum excitation known as the axial Higgs mode – exciting in its own right – also holds promise for developing and manipulating higher temperature quantum materials... Read more…
The battle for datacenter dominance keeps getting hotter. Today, Nvidia kicked off its spring GTC event with new silicon, new software and a new supercomputer. Speaking from a virtual environment in the Nvidia Omniverse 3D collaboration and simulation platform, CEO Jensen Huang introduced the new Hopper GPU architecture and the H100 GPU... Read more…
PsiQuantum, founded in 2016 by four researchers with roots at Bristol University, Stanford University, and York University, is one of a few quantum computing startups that’s kept a moderately low PR profile. (That’s if you disregard the roughly $700 million in funding it has attracted.) The main reason is PsiQuantum has eschewed the clamorous public chase for... Read more…
AMD is getting personal with chips as it sets sail to make products more to the liking of its customers. The chipmaker detailed a modular chip future in which customers can mix and match non-AMD processors in a custom chip package. "We are focused on making it easier to implement chips with more flexibility," said Mark Papermaster, chief technology officer at AMD during the analyst day meeting late last week. Read more…
Intel reiterated it is well on its way to merging its roadmap of high-performance CPUs and GPUs as it shifts over to newer manufacturing processes and packaging technologies in the coming years. The company is merging the CPU and GPU lineups into a chip (codenamed Falcon Shores) which Intel has dubbed an XPU. Falcon Shores... Read more…
Additional details of the architecture of the exascale El Capitan supercomputer were disclosed today by Lawrence Livermore National Laboratory’s (LLNL) Terri Read more…
Just a couple of weeks ago, the Indian government promised that it had five HPC systems in the final stages of installation and would launch nine new supercomputers this year. Now, it appears to be making good on that promise: the country’s National Supercomputing Mission (NSM) has announced the deployment of “PARAM Ganga” petascale supercomputer at Indian Institute of Technology (IIT)... Read more…
The long-troubled, hotly anticipated MareNostrum 5 supercomputer finally has a vendor: Atos, which will be supplying a system that includes both Nvidia and Inte Read more…
MLCommons today released its latest MLPerf inferencing results, with another strong showing by Nvidia accelerators inside a diverse array of systems. Roughly fo Read more…
Supercomputing has been indispensable throughout the Covid-19 pandemic, from modeling the virus and its spread to designing vaccines and therapeutics. But, desp Read more…
Close to a decade ago, AMD was in turmoil. The company was playing second fiddle to Intel in PCs and datacenters, and its road to profitability hinged mostly on Read more…
© 2022 HPCwire. All Rights Reserved. A Tabor Communications Publication
HPCwire is a registered trademark of Tabor Communications, Inc. Use of this site is governed by our Terms of Use and Privacy Policy.
Reproduction in whole or in part in any form or medium without express written permission of Tabor Communications, Inc. is prohibited.