Nvidia, Qualcomm Shine in MLPerf Inference; Intel’s Sapphire Rapids Makes an Appearance.

By John Russell

September 8, 2022

The steady maturation of MLCommons/MLPerf as an AI benchmarking tool was apparent in today’s release of MLPerf v2.1 Inference results. Twenty-one organizations submitted 5,300 performance results and 2,400 power measurement. While Nvidia again dominated, showcasing its new H100 GPU, Qualcomm impressed on both performance and power consumption metrics. Intel ran a Sapphire Rapids CPU-based system, albeit in the preview category. AI start-ups Biren, Moffet AI, Neural Magic, Sapeon were among the entrants.

In some ways, making sense of MLPerf results has grown more complicated as the diversity of submitters and system sizes and system configurations have grown. But that’s sort of the goal in the sense that MLPerf’s extensive benchmark suites and divisions (closed, open, etc.) permit more granular comparisons for system evaluators but it takes some effort. Moreover, MLCommons (parent org.) keeps evolving its test suites to stay current and relevant. In this, the sixth set of MLPerf inference results, a new division (Inference over the network) and a new object detection model (RetinaNet) were added.

While training is regarded as a more HPC-related activity, inference is regarded as the enterprise workhorse. Depending on the application scenario, latency may or may not be an issue, accuracy is always an issue, and power consumption, particularly for cloud and edge applications, is critical. The MLPerf inference suite, broadly, has provisions for all of these characteristics. Fortunately, MLCommons had made parsing the results fairly easy. (Link to results)

As has become MLPerf’s practice, vendors are allowed to submit statements describing the systems used in the MLPerf exercise and special features that may assist their AI-related performance. Of note, at least one participant – Neural Magic – is a software company. The companies’ statements, as provided by MLCommons, are provided at the end of the article.

Nvidia Still Rules the Roost

Nvidia is the only company that has run all of the tests in (closed division) in every MLPerf run. It showcased its new H100 GPU in the preview category (available within six months) and its A100 in the closed division. H100 was up to 4.5x faster than the A100-based systems.

David Salvator, director of AI inference, benchmarking, and cloud, at Nvidia, said the big gains were made possible by leveraging Nvidia’s transformer engine which uses FP8 where possible. “Basically [it’s] a technology that makes intelligent use of the FP8 precision. When you go to lower precision, one of the concerns is rounding errors introduce error and inaccuracy in the answers that a neural network is giving you back.” The software-defined transformer engine is able to choose where to use FP8. “The H100 should really be thought of now as our current generation product,” he said.

Not that the wildly successful A100 is going away. It swept all the datacenter closed division metrics and, said Salvator, can even be used at the edge, “The A100 has that ability to be deployed at the edge, particularly our PCIe based version, which can be put into a one rack server that be deploy at the edge if you need a very performant platform at the edge to be doing inference and even some like training work.”

Qualcomm’s AI 100 Continues to Impress

Qualcomm followed up its strong showing in the spring MLPerf Inference (v2.0) with another set. The company’s Cloud AI 100 accelerator is gaining traction and, of course, its Snapdragon processor has long been a mobile industry leader. The AI 100 is mostly used in one of two versions – Pro (400 TOPS) or Standard (300 TOPS) and features miserly power use.

John Kehrli, a senior director of product management, reiterated Qualcomm’s strength in low power devices, “Our story, continues here in terms of power efficiency. In this particular [MLPerf] cycle my perspective is what you’ll really see is showcasing our performance across a number of commercial platforms. You’ll see new submissions from Dell, HP, Lenovo, Thundercomm. You’ll see a new variant from Inventec as well, Foxconn, Gloria is doing a higher submission as well. And see submissions with Standard and Pro.”

Again, it is necessary to compare systems side-by-side to get a clear picture of strengths and weaknesses. The AI 100 was a strong performer at the edge, including the best performance in edge single stream for ResNet-50 and performed well in the multi-stream as well. The company also pulled together some queries-per sec-per-watt analyses which were impressive.

Analyst Karl Freund wrote, “The fact that both Gloria (Foxconn) and Inventec Heimdall, suppliers to cloud service companies, submitted results to MLCommons tells us that Qualcomm may be realizing traction in the Asian cloud market, while Dell, Lenovo, and HPE support indicates global interest in the Qualcomm part for datacenter and “Edge Clouds.”

“The current Cloud AI 100 demonstrates dramatically superior efficiency for image processing compared to both cloud and edge competitors. This makes sense as the heritage of the accelerator is the high-end Snapdragon mobile processor’s Qualcomm AI Engine, which provides AI for mobile handsets where imaging is the primary application. Nonetheless, the Qualcomm platform provides best-in-class performance efficiency for the BERT-99 model, used in natural language processing.”

Good to See Growing Participation

It was interesting to see the Intel Sapphire Rapids-based development system (1-node-2S-SPR-PyTorch-INT8, 2 CPUs) using PyTorch. The preview category is for products that will be available in six months.

Intel’s Jordan Plawner, senior director of AI products, said, “We’re not releasing our core-count and other products specifics. Obviously, there’ll be a range of core counts. This was performed on final silicon, but it is still what I would call with the pre-production software stack. So we expect improvements as we optimize the software between now and launch, which the launch date has not yet been announced.”

Several relative newcomers to AI acceleration submitted entries and in addition to looking at their results it may be useful to look at their provided statements (included below in compilation of statements below).

One hot topic at the formal MLCommons/MLPerf press analyst briefing was so-called sparsity and its use to speed performance. There are, of course, many approaches. It can, however, spoil the apples-to-apples nature of submissions in the closed division.

Nvidia’s Salvator noted, “Some submitters, I believe, have actually done sparse submissions in the open category as a way to sort of show their abilities to work with sparse data sets. As you probably know, we want to specify a data and you typically have to do a recalibration on the training side [to use or sparsity], right. In today’s MLPerf inference on the closed side, we don’t really have a provision for that yet. It’s something of course, we’re always looking at.

“One of the important things about the closed category is that everybody has to do the same math. There are certain optimizations, shortcuts, whatever you want to call them, that can be taken involving things like sparsity and involving pruning, where you take parts of the network that aren’t really being lit up and basically just trim them away because they’re not being used. Those are optimizations that can be valid. But for the purpose of comparison, and making sure that everyone has done the same work, the closed category does not allow those,” he said.

Stay tuned.

Link to MLPerf Inference Datacenter v2.1 results: https://mlcommons.org/en/inference-datacenter-21/

Link to MLPerf Inference Edge v2.1 results: https://mlcommons.org/en/inference-edge-21/

Link to Nvidia blog: https://blogs.nvidia.com/blog/2022/09/08/hopper-mlperf-inference/

VENDOR STATEMENTS SUPPLIED BY MLCOMMONS

Alibaba

Alibaba Cloud is one of the top-tier cloud service providers in the world, striving to provide the best customer experience around the globe. In MLPerf Inference v2.1, we demonstrated the efficiency of hardware/software co-optimization tools on Nvidia GPUs and Alibaba Yitian 710 CPU.

In the Network Division, Alibaba made the inaugural submission based on a heterogeneous resource pool of accelerators. By leveraging Alibaba’s Sinian vODLA technology which abstracts away the heterogeneity of underlying computing architectures and virtualizes physical compute devices to form a scalable resource pool, we can provide users a single abstract device with arbitrary-sized computation power. Sinian vODLA technology is flexible and optimized to work across many hybrid types of accelerators, going beyond the physical limits of system PCIe slots and pod network links. On the Alibaba EFlops system powered by Sinian vODLA technology, we achieved 107060 QPS for BERT-99 on 4 compute nodes with a total 32 A100 GPUs. It also enabled us to do inference with hybrid types of GPUs.

In the Open Division (datacenter/edge), we used the Sinian full stack optimizer (SinianML) to deeply compress ML models and automatically generate more computation-friendly model architectures at runtime. SinianML also adopts quantization-aware training so that the generated model can benefit greatly from Alibaba Yitian CPU’s efficient ‘int8’ instructions during inference. On the Alibaba Yitian CPU-only system, SinianML gained ~55X speedup against the same model without Sinian optimizations while still satisfying the same accuracy requirements.

In the Closed Division (edge), we leverage AIACC to optimize the ResNet50 within the Single-Stream pattern on Alibaba Cloud. Under equal conditions, the inference latencies were lowered to 0.398ms and 0.377ms on A100 GPU and A10 GPU, respectively.

Alibaba Sinian Platform is a heterogeneous hardware acceleration framework to seamlessly build, train, and deploy ML models without loss of performance portability. AIACC is Alibaba’s ECS ApasraCompute Tool for high-performance optimizations on cloud AI workloads.

ASUSTeK

ASUSTeK is committed to joining ML Commons on AI training and inference benchmarks performance. ASUS ESC8000A-E11 and ESC4000-E10S are specifically designed to deliver intensive AI applications across different inference benchmarks and are tested and validated on the latest MLPerf Inference v2.1.

ESC8000A-E11 is powered by AMD EPYC 7003 family processors and 8 NVIDIA A100 PCIe GPUs, and delivers optimized GPU performance to offload critical workloads from CPU. Design thinking is in the corporate DNA at ASUS, and with the innovative hardware design and proprietary tuning technology for server performance and power efficiency, ESC8000A-E11 can boost more performance on AI related workloads.

ESC4000-E10S is equipped with 3rd Gen Intel Xeon Scalable family processors and 8 NVIDIA A2 GPUs, and also achieves great results on MLPerf Inference v2.1. MLPerf Inference V2.1 is a big leap in inference applications, ASUS is firmly impressed by these improvements which solves more AI issues when powering AI frameworks to another level.

Hardware configuration

ESC8000A-E11 with NVIDIA A100 PCIe GPUs and ESC4000-E10S with NVIDIA A2 PCIe GPUs are tested in all Inference V2.1 benchmarks in server and offline scenarios, benchmarks, including ResNet, RetinaNet, 3D-UNet-99, 3D-UNet-99.9, RNN-T, BERT-99, BERT-99.9, DLRM-99, DLRM -99.9. Users could take these benchmarks as reference when evaluating AI workloads based on ASUS GPU server solutions.

Azure

Azure is pleased to share results from our MLPerf Inference v2.1 submission. For this submission, we benchmarked our NC A100 v4-series, NDm A100 v4-series, and NVads A10_v5-series. They are powered by the latest NVIDIA A100 PCIe GPUs, NVIDIA A100 SXM GPUs and NVIDIA A10 GPUs respectively. These offerings are our flagship virtual machine (VM) types for AI inference and training and enable our customers to address their inferencing needs from 1/6 of a GPU to 8 GPUs.

Some of the highlights from our MLPerf inferencing v2.1 benchmark results are:

–           NC A100 v4-series achieved 54.2K+ samples/s for RNN-T offline scenario

–           NDm A100 v4-series achieved 26+ samples/s for 3D U-Net offline scenario

–           NVads A10 v5-series achieved 24.7K+ queries/s for ResNet50 server scenario

These inference benchmark results demonstrate how Azure is committed to providing our customers with the latest GPU offerings, that are in line with on-premises performance and available on-demand in the cloud, and scales to adapt to all sizes of AI workloads and needs.

Special thanks to our hardware partner NVIDIA for providing the instructions and containers that enabled us to run these benchmarks. We deployed our environment using the aforementioned offerings and Azure’s Ubuntu 18.04-HPC marketplace image.

The NC A100 v4-series, NDm A100 v4-series, and NVads A10 v5-series are what we and our Azure customers turn to when large-scale AI and ML inference is required. We are excited to see what new breakthroughs our customers will make using these VMs.

Biren

Founded in 2019, Biren Technology is a provider of high performance accelerated computing products for AI and graphics.

We are pleased to share the results from our first MLPerf Inference submission. Our submission is in the Datacenter Close Division, and the system is powered by our in-house Bili 104 PCIe accelerators (based on BR104 chips) and the BIRENSUPA software development platform.

Some of the result highlights include

  1. BERT 99.9: achieved 22.1K+ samples/s with a 8-accelerator system in the offline scenario
  2. Resnet50: achieved 424.6K+ samples/s with a 8-accelerator system in the offline scenario

The BIRENSUPA software development platform provides a complete development environment for working with Biren devices, including a novel programming model & language, compiler, acceleration libraries, deep learning frameworks, development tools, and SDKs for different areas. Bili 104, together with the BIRENSUPA, provides customers with flexible and powerful solutions to address computation challenges in the AI domain.

The above MLPerf results clearly demonstrate the performance and capabilities that our products can offer. Biren continues to integrate new technologies and optimizations into our products. We are excited to work with customers, and help them to achieve more.

Fujitsu

Fujitsu offers a fantastic blend of systems, solutions, and expertise to guarantee maximum productivity, efficiency, and flexibility delivering confidence and reliability. We have continued to participate in and submit to every inference and training round for the data center division since 2020.

In this round, Fujitsu measured benchmark programs for the edge division with PRIMERGY RX2540 M6 for the first time, not just for the data center division with PRIMERGY GX2570 M6. We also measured power consumption with the same configuration other than GPU power limit with PRIMERGY GX2570 M6. The details of these systems are shown as follows:

  1. PRIMERGY GX2570 M6 is a 4U rack-mount server with Intel (R) Xeon (R) Platinum 8352V CPUx2 and NVIDIA A100 SXM 80GB x8. This server performs far better in our PRIMERGY server portfolio across heavy-duty Deep Learning (AI), Data Science and HPC workloads.
  2. PRIMERGY RX2540 M6 is a 2U rack-mount server with Intel(R) Xeon(R) Platinum 8352Y CPUx2 and NVIDIA A30 x1.

Compared with our past submissions, performance has improved over time, partly because GPU performance has evolved. 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.

Therefore, we will continue to meet the demands of our customers and strive to provide attractive server systems through the activities of MLCommons.

GIGABYTE

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 servers were equipped with various brands of accelerators (NVIDIA and Qulacomm) and processors (AMD, Ampere, and Intel) in configurations to showcase systems that target different markets (x86 and Arm).

For MLPerf Inference v2.1, in closed Data Center, all tasks were run in a GIGABYTE 4U server using the Intel platform supporting NVIDIA SXM GPUs.

Server CPU GPU Testing
G492-ID0 Intel Xeon 8380 NVIDIA SXM-A100 80GB  8-GPU closed, datacenter

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.

H3C

H3C is pleased to share results from our MLPerf Inference v2.1 submission.

 We benchmarked H3C UniServer R5300 G5, R5500 G5 and R4900 G5 which can be applied to AI inference and training:

  • H3C UniServer R5500 G5 supports NVIDIA HGX A100 8-GPU module, and eight NVIDIA A100 GPUs can be fully interconnected with six NVSWITCH at 600GB/s. It adopts modular design and perfectly matches NVIDIA A100 GPU in heat dissipation, power supply and I/O expansibility, giving full play to the strong performance of A100.
  • H3C UniServer R5300 G5 supports NVIDIA HGX A100 4-GPU module, with four A100s interconnected with NVLINK at 600 GB/s. It also supports various PCIe AI accelerators, with up to 8 dual-width or 20 single-width accelerators supported. The CPU and GPU mount ratio can be in various topology configurations, including 1:4 and 1:8.

Some of the highlights from our MLPerf Inference v2.1 benchmark results are:

  1. H3C UniServer R5300 G5 achieved 13.04 samples/s for 3D-UNET offline scenario.
  2. H3C UniServer R5500 G5 achieved 314368 queries/s for ResNet50 server scenario, improved by 11.5% compared with the performance H3C achieved in Inference v2.0.
  3. H3C UniServer R5500 G5 with one A100-SXM-80GB achieved extremely low latency of 1.53ms for BERT SingleStream scenario.

These inference benchmark results demonstrate H3C’s strong technical power. We’re committed to providing our customers with the latest GPU offerings to adapt to all sizes of AI workloads and needs.

HPE

Hewlett Packard Enterprise (HPE) has returned to MLPerf Inference submissions with a strong presence. Aligning to HPE’s edge-to-cloud and heterogeneity goals, our submission has been developed partnering with Nvidia and Qualcomm Technologies.

This year, HPE continued to expand its submissions which highlights the versatility of the HPE Apollo 6500 in supporting a variety of NVIDIA A100 accelerator hardware configurations while maintaining excellent performance. HPE has doubled the number of submissions in this round as compared to 1.1, adding several new closed, datacenter submissions with Apollo-based configurations as well as new submissions in the closed, edge category with Edgeline-based configurations. These Apollo-based configurations include third generation AMD EPYC™ processors with eight A100-SXM-80GB accelerators (with and without MIG) and eight A100-SXM-40GB accelerators, as well as second generation AMD EPYC™ processors with four A100-SXM-40GB accelerators.

HPE also showcased a new edge offering which delivered more than 77,500 inferences/second on the ResNet50 computer vision workload and 2,800 inferences/second on the BERT NLP workload with Qualcomm Cloud AI 100.  These results were delivered on HPE Edgeline, optimized for 0-55C operating environments while consuming only about 420-450 Watts. HPE also delivered leading results for high-performance power-efficiency in this category.   HPE Edgeline EL8000 systems with Qualcomm Cloud AI 100 accelerators are available now where high-performance, power efficiency, and server ruggedization is needed.   HPE worked with partners Qualcomm and Krai for the MLPerf v2.1 benchmark submissions powered by Collective Knowledge v2.6.1.

HPE demonstrated in this round a strong compute portfolio for inference by going beyond our big investments in machine learning tools such as Machine Learning Development Environment (MLDE) to train accurate models faster, Swarm learning bringing training to the data sources, and a full platform in Machine learning Development System (MLDS) for scaling AI models. Information on these products can be found at https://www.hpe.com/us/en/solutions/artificial-intelligence.html.

Inspur

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 Inference V2.1, Inspur made submissions on five systems, NF5488A5, NF5688M6, NF5468M6J, NF5468M6 and NE5260M5.

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. 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. NF5468M6J supports up to 24 A100 GPUs and can be widely used in Internet AI public cloud, enterprise-level AI cloud platform, smart security, video codec, etc.

NE5260M5 is an edge server with building blocks optimized for edge AI applications, and 5G applications with capability of operating at temperatures between -5℃~50℃.

In offline scenario of datacenter closed division, the performance of Bert-99, Bert-99.9, 3D-UNET-99 and 3D-UNET-99.9, are improved by 93.8%, 96.7%, 106.3% and 106.3% on NF5468M6J compared with the best performance that Inspur achieved in Inference v2.0. In server scenario of datacenter closed division, the performance of Bert-99 and Bert-99.9 are improved by 100% and 102.4% on NF5468M6J compared with the best performance that Inspur achieved in Inference v2.0. Inspur also submitted Bert and Resnet on NF5468M6 with four BR104 accelerators, and achieved very good performance.

In offline scenario of edge closed division, the performance of 3D-Unet-99, 3D-Unet-99.9, bert-99 and RNNT are improved by 3.97%, 3.97%, 1.9% and 0.35%, respectively on NE5260M5 compared with the best performance achieved in Inference v2.0.

Intel

Intel submitted MLPerf Inference v2.1 results on the 4th Gen Intel Xeon Scalable processor product line (codenamed Sapphire Rapids) as a submission in preview of Intel’s largest, most comprehensive AI launch ever, a platform designed with the flexibility to address the needs from data scientists who want to customize models for large scale data center deployments to the needs of applications engineers who want to quickly integrate models into applications and deploy at the edge. As the only data center CPU vendor to submit MLPerf inference results on a broad set of models, we demonstrate the practicality of running DL inference anywhere on the massive existing install base of Intel Xeon servers alongside other applications.

This preview submission covers the full suite of MLPerf data center benchmarks including image processing, natural language processing (NLP), and recommendation systems. This new platform will deliver significantly more compute and memory capacity/bandwidth than the previous generation, offering a performance boost of 3.9X – 4.7X in Offline and 3.7X – 7.8X in Server, compared to Intel’s previous submissions.

New hardware innovations include Intel® Advanced Matrix Extensions (Intel® AMX) for INT8-based submissions on all workloads, spanning multiple frameworks. This all-new AI accelerator engine sits on every core and delivers 8x operations per clock compared to the previous generation.

To simplify the use of these new accelerator engines and extract the best performance we focused on software. Our customers tend to use the mainstream distributions of the most popular AI frameworks and tools, so Intel’s AI experts have been working for years with the AI community to co-develop this AI Platform integrated with a broad range of open and free-to-use tools, optimized libraries and industry frameworks to deliver a seamless experience deploying AI solutions across generations, getting the best performance out of the box. All software used is available through MLPerf Inference repo, DockerHub, and optimizations are integrated into Intel-optimized PyTorch.

Notices & Disclaimers

Performance varies by use, configuration and other factors. Learn more at www.Intel.com/PerformanceIndex. Performance results are based on testing as of dates shown in configurations and may not reflect all publicly available updates. See backup for configuration details. No product or component can be absolutely secure. Your costs and results may vary. Intel technologies may require enabled hardware, software or service activation.

Krai

We often get asked about the value of submitting benchmarking results to MLPerf. Potential submitters, especially ML hardware startups, are understandably wary of committing precious engineering resources to optimizing industry benchmarks instead of actual customer workloads.

MLPerf is the Olympics of ML optimization and benchmarking. While consulting several leading ML hardware companies as their “Olympic coach”, we have witnessed first-hand the value that our customers extracted from making both actual and “dry-run” MLPerf submissions. As we have discovered, the MLPerf Inference suite is sufficiently diverse that nearly every benchmark presents its own unique challenges, especially when scaling to multiple accelerators and/or hundreds of thousands queries per second. So if nothing else, an intense focus on the MLPerf benchmarks is a health check that serves the broader performance cause, as it often helps resolve similar challenges with customer workloads.

We are proud to have supported Qualcomm’s MLPerf Inference submissions for the fourth time round. In this v2.1 round we have also collaborated with Qualcomm’s partners Dell, HPE and Lenovo, in addition to Alibaba and Gigabyte in the previous v2.0 round. To this end, we have implemented and optimized the Computer Vision and Natural Language Processing benchmarks across a range of Datacenter and Edge platforms powered by the Qualcomm Cloud AI 100 accelerators. As every Olympic coach is proud of their athletes winning Olympic medals, we are proud of the achieved results demonstrating industry-leading performance and energy efficiency.

Finally, we submitted over 4,200 benchmarking results on our own, with over 2,800 results accompanied by power measurements. Our submissions demonstrate accuracy/performance/power trade-offs across a range of Edge platforms, workloads and engines.

Overall, we prepared 81% of all performance results and 97% of all power results in this round, thanks to our powerful workflow automation. We also provided a preview of our next generation workflow technology called KRAI-X.

Lenovo

Lenovo delivers Smarter Technology for All, to enrich the lives of people, advance research and usher in a new era of digital transformation for organizations of all sizes. We aim to deliver a diverse portfolio of compute platforms specifically designed for AI workloads ranging from the data center to the edge and deliver AI solutions that open new possibilities for  enterprises and tech companies of all sizes. Choosing the best infrastructure for your AI applications should not create a barrier to launching or scaling your AI Initiatives.  Our goal through MLPerf inference v2.1 is to bring clarity to infrastructure decisions so our customers can focus on the success of their AI deployment overall.

Our submissions covered 13 different benchmarks across 6 submissions with 3 variations of platforms and acceleration choices. Our Lenovo ThinkSystem SR670 V2 produced more than three-hundred thousand ResNet50 Inferences a second in a 3U data center footprint leveraging eight NVIDIA A100 PCIe GPUs while the ThinkSystem SR670 V2 HGX solution demonstrated high performance inferencing in a hybrid cooled server design with our latest Lenovo Neptune™ liquid-cooling technology.

MLPerf Inference v2.1 was also Lenovo’s debut of the Qualcomm AI Cloud 100 accelerator featured in the Lenovo ThinkEdge SE350 platform, delivering more than 23,000 ResNet50 inferences a second in ruggedized edge server with a footprint less than 3 liters.

We are very proud of the results and advancements that MLPerf has contributed to AI performance improvements, standards and resources for the data science community.

Moffett

The inventor of the dual sparsity algorithm, Moffett AI has the world’s leading sparse computing techniques with more than 30 patents worldwide. The company creates a new generation of AI computing platform with hardware and software co-design to achieve order-of-magnitude acceleration of computing performance, reducing latency and low TCO. The results of S4, S10, and S30 in MLPerf v2.1 have demonstrated the potential of sparse computing in inference performance and energy efficiency, which leads to a lower total cost of ownership (TCO).

For MLPerf Inference v2.1, Moffett has submitted results of three high-sparsity accelerators (S4, S10, and S30) running ResNet-50 and Bert-large 99.9%, respectively.

Submission for: Datacenter, Offline scenario, Open division

S30 S4 S10
resnet50 95784 FPS 31679 FPS 65593 FPS
Bert-99.9 3837 SPS 1219 SPS 2536 SPS

Besides the performance results, energy efficiency is another significant highlight of these three sparse computing accelerators. For example, the peak power consumption of S30 is merely 250W, and 75W for Antoum S4.

The Antoum architecture through hardware and software co-design and Moffett’s original sparsity algorithm are the reasons to achieve great performance with high energy efficiency.

The accelerators for AI inference applications in data centers are equipped with Moffett’s 1st generation Antoum processor – the first commercial AI processor with 32x sparsity in the world. Besides the sparse processing units (SPU) for native sparse convolution and matrix computing in Antoum, the processor also integrates a Vector Processing Unit (VPU), which enables flexible programmability to keep up with the fast evolution of AI models.

Also, the on-chip Video Codec, which supports 192-way 1080p video decoding at 30 FPS, and the JPEG decoder, which supports 1080p image decoding up to 6960 FPS, provide an end-to-end capability for video and image inference workloads.

Nettrix

Nettrix Information Industry (Beijing) Co., Ltd. (Hereinafter referred to as Nettrix) is a server manufacturer integrating R&D, production, deployment, and O&M, as well as an IT system solution provider. It aims to provide customers industry-wide with various types of servers and IT infrastructure products such as common rack based on artificial intelligence, multiple nodes, edge computing and JDM life cycle customization. Nettrix has developed server products for industries including Internet, telecommunications, finance, medical care and education.

Nettrix X640 G40 is an all-round GPU server with both training and reasoning functions. It supports up to 8 training GPUs, and provides comprehensive performance support for high-density GPU computing. The product supports a variety of different GPU topologies, and optimizes GPU interconnection for different applications and models. It is an efficient and all-round computing platform. At the same time, it has been adapted to the mainstream GPUs on the market, and is perfectly compatible with a variety of GPU types. Meets the flexible needs of customers.

Nettrix X660 G45 LP is a liquid cooled GPU server. It is a high-performance computing platform specially developed for deep learning training and is designed for huge computing clusters. As a liquid cooled server as well, X660 G45 LP is designed with cold plate liquid cooling for the main heat sources, including CPU, GPU, and memory, which can achieve component-level precise cooling so as to release the full performance of CPU and GPU and reduce energy consumption. The liquid cooling technology has played a significant role in this test, effectively reducing the GPU temperature, ensuring the optimal working frequency, and giving full play to the GPU computing power.

Nettrix has been focusing on server business for 15 years. Each liquid cooled server has experienced different sorts of strict tests from R&D, manufacturing to delivery inspection, including impact resistance test, vibration resistance test, signal integrity test, and electromagnetic compatibility test, which can ensure its stable and reliable operations. In the AI ​​era, Nettrix is ​​committed to creating rich AI server products for customers, continuously improving computing power and energy efficiency returns, and helping customers’ intelligent transformation.

Neural Magic

Neural Magic introduces performant model sparsity to MLPerf. It brings an open-source compression framework that unifies state-of-the-art algorithms such as pruning and quantization (SparseML), and a sparsity-aware inference engine for CPUs (DeepSparse). The power of sparse execution enables neural networks to be run accurately, at orders of magnitude faster speeds, using commodity CPU hardware. This marks an important milestone in efficient machine learning powered by better algorithms and software.

When applied to BERT-Large SQuAD v1.1 question answering task, Neural Magic’s benchmarks:

  • Maintain >99% of its original F1 score
  • Decrease model size by orders of magnitude from 1.3 GB to ~10 MB
  • Improve throughput performance from ~10 samples/second to up to 1,000 samples/second when executed in the sparsity-aware DeepSparse Engine.

The benchmark was evaluated using a server with two 40-core Intel Xeon Platinum 8380 CPUs. More details on each of the models and methods can be found on GitHub under Neural Magic’s BERT-Large DeepSparse MLPerf Submission.

Detailed model methods:

NVIDIA

In MLPerf 2.1 Inference, we are excited to make our first H100 submission, demonstrating  up to 4.5X higher performance than our A100 GPU. NVIDIA H100, based on the groundbreaking NVIDIA Hopper Architecture, supercharges the NVIDIA AI platform for advanced models, providing customers with new levels of performance and capabilities for the most demanding AI and HPC workloads. Hopper features our Transformer Engine, which applies per-layer intelligence to the use of FP8 precision, delivering optimal performance for both AI training and inference workloads while preserving model accuracy.

The NVIDIA AI platform delivers great performance on a broad range of models, accelerates the end-to-end AI workflow from data prep to training to deployed inference, and is available from every major cloud and server maker. We make these resources available to the developer community via NGC, our container repository.

We are excited to see our 12 NVIDIA partners submit great inference results on A100-based systems across all tests, both for on-prem as well as cloud platforms. A100 continues to deliver excellent inference performance across the full suite range of MLPerf tests, spanning image, speech, reinforcement learning, natural language and recommender systems.

We also wish to commend the ongoing work MLCommons is doing to bring benchmarking best practices to AI and HPC, enabling peer-reviewed apples-to-apples comparisons of AI and HPC platforms to better understand and compare product performance.

OctoML

Speaker: Grigori Fursin, Vice President of MLOps and the author of the Collective Knowledge framework

OctoML is a Series-C startup developing a machine learning deployment platform to automatically optimize, benchmark and deploy models in production from any deep learning framework to any target hardware based on user requirements and constraints.

Since 2021, OctoML is supporting and extending the Collective Knowledge framework (CK) to make it easier for companies to use, customize and extend the MLPerf benchmarking infrastructure, lower the barrier of entry for new MLPerf submitters and reduce their associated costs.

After donating the CK framework to the MLCommons, OctoML continued developing this portable workflow automation technology as a community effort within the open education workgroup to modularize MLPerf and make it easier to plug in real-world tasks, models, data sets, software and hardware from the cloud to the edge.

We are very glad that more than 80% of all performance results and more than 95% of all power results were automated by the MLCommons CK v2.6.1 in this round thanks to submissions from Qualcomm, Krai, Dell, HPE and Lenovo!

Furthermore, the feedback from the users has helped our workgroup to develop the next generation of the CK technology called Collective Mind available to everyone under Apache 2.0 license at https://github.com/mlcommons/ck/tree/master/cm – we used CM in this round to test the new framework and demonstrate how to simplify and automate submissions.

We invite you to join our open education workgroup to continue developing this portable workflow automation framework as a community effort to help everyone:

  • learn how to use, customize and extend the MLPerf benchmarking infrastructure;
  • modularize MLPerf and make it easier to plug in real-world tasks, models, data sets, software and hardware from the cloud to the edge;
  • lower the barrier of entry for new MLPerf submitters and reduce their associated costs.

Qualcomm Technologies, Inc.

Speaker: John Kehrli, Senior Director of Product Management at Qualcomm Technologies, Inc.

Qualcomm’s® MLPerf v2.1 inference results continue to demonstrate our performance-to-power efficiency leadership, from edge to cloud with incremental upside in performance for NLP and computer-vision networks. We expanded Qualcomm® Cloud AI 100 benchmark submission scope with new partners, Dell, HPE and Lenovo, and new edge and datacenter platforms, including HPE Proliant e920d, Dell PowerEdge 7515 and Lenovo ThinkSystem SE350. For the first time PCIe HHHL-Standard (350 TOPs) accelerator-based submissions are made in addition to PCIe HHHL-Pro (400 TOPs) accelerators. We introduced three new edge device platforms based on Snapdragon® with Qualcomm Cloud AI 100 accelerators, including Foxconn Gloria High-end with 200 TOPs, and Thundercomm TurboX EB6 and Inventec Heimdall each with a 70 TOPs accelerator – covering more than 200 MLPerf v2.1 results based on Qualcomm Cloud AI 100 accelerators.

We continue to innovate and optimize AI solutions across submissions. The 2U server platform with 18x Qualcomm Cloud AI 100 accelerators achieved ResNet-50 offline peak performance of 428K+ inference per second (418K in v2.0).  ResNet-50 power efficiency for Foxconn Gloria High-end peaked at 311 inferences/second/Watt. BERT performance across platforms has increased up to 15% and BERT power efficiency on Foxconn Gloria Entry improved by 15%.

We expanded network coverage with the RetinaNet object detector in the Open division achieving 187 inference/second for PCIe-HHHL-Pro accelerator. Among Qualcomm Cloud AI 100 accelerator based RetinaNet submissions, the highest power efficiency for datacenter submissions was measured at 2.65 inferences/second/Watt, whereas Gloria High-end achieved 4.16 inferences/second/Watt.

All submissions are prepared in collaboration with Krai and are powered by Collective Knowledge v2.6.1.

Qualcomm and Snapdragon are trademarks or registered trademarks of Qualcomm Incorporated. Qualcomm Cloud AI and Snapdragon are products of Qualcomm Technologies, Inc. and/or its subsidiaries.

SAPEON

SAPEON Inc. is a corporation founded by SK Telecom, SK Hynix, and SK Square to make inroads in the global AI semiconductor market.

The AI accelerator project, SAPEON, started in 2017 at SK Telecom with the goal of enhancing the internal AI infrastructure for SK Group. It has since been evolved to the second-generation product, the X220. The accelerators have been used for various applications like NUGU, Korea’s first AI speaker, and our intelligent video security system SK Shieldus, the second largest physical security company in Korea.

 After being spun off from SK Telecom, SAPEON Inc. disclosed the X220’s performance results for the first time, which have been validated by MLPerf in order to introduce its product to the market. SAPEON X220, the second-generation product debuted in 2020, accelerates deep learning workloads in data centers.

 SAPEON is optimized for real-time interactive AI inferences for a large number of users at one time, and the results are presented in MLPerf “Inference: Datacenter” benchmarks to measure the performance of those types of services.

 The highlights from our MLPerf 2. 1 benchmark results are:

– SAPEON X220-Compact achieved 6,145 queries/s and 6,769 samples/s for ResNet 50-v1.5 in server and offline scenario respectively.

– SAPEON X220-Enterprise achieved 12,036 queries/s and 13,101 samples/s for ResNet 50-v1.5 in server and offline scenario respectively.

The cost and energy-efficiency of the X220 will be superior in the next-generation, X330, released in the second half of 2023. The X330 is a superior performing AI Accelerator for large-scale AI inference systems and provides precise operations for more accurate services. It can also be used for training infrastructure.

 SAPEON is targeting the global AI markets with vertical solutions which integrates AI semiconductor chips and software such as AI algorithms and services with a proven track record in SK Group businesses such as media, security, smart factories, etc.

Supermicro

Supermicro has its long history of providing a broad portfolio of products for different AI use cases. In MLPerf Inference v2.1, we have submitted three systems with nine configurations in the datacenter and edge inference category. These are to address the performance for multiple use cases, including medical image segmentation, general object detection, recommendation systems, and natural language processing in centralized datacenter and distributed edges.

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 8GPU platform and HGX A100 4GPU 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.

For customers looking for PCIe platforms, Supermicro provides even more choices. In MLPerf v2.1, we submitted results for IoT SuperServer SYS-E403-12P-FN2T, a compact high performer with box PC form factor. With multiple selections of GPUs, the system is a perfect fit for edge AI applications such as predictive analysis, and network operation monitoring and management. The system is currently shipping worldwide.

Supermicro’s SYS-120GQ-TNRT is a high density 1U server that accommodates 4 double width GPUs. It is designed for customers looking for compact but high AI performance. It can achieve 50% of the performance while only using 25% of the rack unit of our high end GPU servers.

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.

Subscribe to HPCwire's Weekly Update!

Be the most informed person in the room! Stay ahead of the tech trends with industy updates delivered to you every week!

Supercomputing Helps Explain the Milky Way’s Shape

September 30, 2022

If you look at the Milky Way from “above,” it almost looks like a cat’s eye: a circle of spiral arms with an oval “iris” in the middle. That iris — a starry bar that connects the spiral arms — has two stran Read more…

Top Supercomputers to Shake Up Earthquake Modeling

September 29, 2022

Two DOE-funded projects — and a bunch of top supercomputers — are converging to improve our understanding of earthquakes and enable the construction of more earthquake-resilient buildings and infrastructure. The firs Read more…

How Intel Plans to Rebuild Its Manufacturing Supply Chain

September 29, 2022

Intel's engineering roots saw a revival at this week's Innovation, with attendees recalling the show’s resemblance to Intel Developer Forum, the company's annual developer gala last held in 2016. The chipmaker cut t Read more…

Intel Labs Launches Neuromorphic ‘Kapoho Point’ Board

September 28, 2022

Over the past five years, Intel has been iterating on its neuromorphic chips and systems, aiming to create devices (and software for those devices) that closely mimic the behavior of the human brain through the use of co Read more…

DOE Announces $42M ‘COOLERCHIPS’ Datacenter Cooling Program

September 28, 2022

With massive machines like Frontier guzzling tens of megawatts of power to operate, datacenters’ energy use is of increasing concern for supercomputer operations – and particularly for the U.S. Department of Energy ( Read more…

AWS Solution Channel

Shutterstock 1818499862

Rearchitecting AWS Batch managed services to leverage AWS Fargate

AWS service teams continuously improve the underlying infrastructure and operations of managed services, and AWS Batch is no exception. The AWS Batch team recently moved most of their job scheduler fleet to a serverless infrastructure model leveraging AWS Fargate. Read more…

Microsoft/NVIDIA Solution Channel

Shutterstock 1166887495

Improving Insurance Fraud Detection using AI Running on Cloud-based GPU-Accelerated Systems

Insurance is a highly regulated industry that is evolving as the industry faces changing customer expectations, massive amounts of data, and increased regulations. A major issue facing the industry is tracking insurance fraud. Read more…

Do You Believe in Science? Take the HPC Covid Safety Pledge

September 28, 2022

ISC 2022 was back in person, and the celebration was on. Frontier had been named the first exascale supercomputer on the Top500 list, and workshops, poster sessions, paper presentations, receptions, and booth meetings we Read more…

How Intel Plans to Rebuild Its Manufacturing Supply Chain

September 29, 2022

Intel's engineering roots saw a revival at this week's Innovation, with attendees recalling the show’s resemblance to Intel Developer Forum, the company's ann Read more…

Intel Labs Launches Neuromorphic ‘Kapoho Point’ Board

September 28, 2022

Over the past five years, Intel has been iterating on its neuromorphic chips and systems, aiming to create devices (and software for those devices) that closely Read more…

HPE to Build 100+ Petaflops Shaheen III Supercomputer

September 27, 2022

The King Abdullah University of Science and Technology (KAUST) in Saudi Arabia has announced that HPE has won the bid to build the Shaheen III supercomputer. Sh Read more…

Intel’s New Programmable Chips Next Year to Replace Aging Products

September 27, 2022

Intel shared its latest roadmap of programmable chips, and doesn't want to dig itself into a hole by following AMD's strategy in the area.  "We're thankfully not matching their strategy," said Shannon Poulin, corporate vice president for the datacenter and AI group at Intel, in response to a question posed by HPCwire during a press briefing. The updated roadmap pieces together Intel's strategy for FPGAs... Read more…

Intel Ships Sapphire Rapids – to Its Cloud

September 27, 2022

Intel has had trouble getting its chips in the hands of customers on time, but is providing the next best thing – to try out those chips in the cloud. Delayed chips such as Sapphire Rapids server processors and Habana Gaudi 2 AI chip will be available on a platform called the Intel Developer Cloud, which was announced at the Intel Innovation event being held in San Jose, California. Read more…

More Details on ‘Half-Exaflop’ Horizon System, LCCF Emerge

September 26, 2022

Since 2017, plans for the Leadership-Class Computing Facility (LCCF) have been underway. Slated for full operation somewhere around 2026, the LCCF’s scope ext Read more…

Nvidia Shuts Out RISC-V Software Support for GPUs 

September 23, 2022

Nvidia is not interested in bringing software support to its GPUs for the RISC-V architecture despite being an early adopter of the open-source technology in its GPU controllers. Nvidia has no plans to add RISC-V support for CUDA, which is the proprietary GPU software platform, a company representative... Read more…

Nvidia Introduces New Ada Lovelace GPU Architecture, OVX Systems, Omniverse Cloud

September 20, 2022

In his GTC keynote today, Nvidia CEO Jensen Huang launched another new Nvidia GPU architecture: Ada Lovelace, named for the legendary mathematician regarded as Read more…

Nvidia Shuts Out RISC-V Software Support for GPUs 

September 23, 2022

Nvidia is not interested in bringing software support to its GPUs for the RISC-V architecture despite being an early adopter of the open-source technology in its GPU controllers. Nvidia has no plans to add RISC-V support for CUDA, which is the proprietary GPU software platform, a company representative... Read more…

AWS Takes the Short and Long View of Quantum Computing

August 30, 2022

It is perhaps not surprising that the big cloud providers – a poor term really – have jumped into quantum computing. Amazon, Microsoft Azure, Google, and th Read more…

US Senate Passes CHIPS Act Temperature Check, but Challenges Linger

July 19, 2022

The U.S. Senate on Tuesday passed a major hurdle that will open up close to $52 billion in grants for the semiconductor industry to boost manufacturing, supply chain and research and development. U.S. senators voted 64-34 in favor of advancing the CHIPS Act, which sets the stage for the final consideration... Read more…

Chinese Startup Biren Details BR100 GPU

August 22, 2022

Amid the high-performance GPU turf tussle between AMD and Nvidia (and soon, Intel), a new, China-based player is emerging: Biren Technology, founded in 2019 and headquartered in Shanghai. At Hot Chips 34, Biren co-founder and president Lingjie Xu and Biren CTO Mike Hong took the (virtual) stage to detail the company’s inaugural product: the Biren BR100 general-purpose GPU (GPGPU). “It is my honor to present... Read more…

Newly-Observed Higgs Mode Holds Promise in Quantum Computing

June 8, 2022

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…

AMD’s MI300 APUs to Power Exascale El Capitan Supercomputer

June 21, 2022

Additional details of the architecture of the exascale El Capitan supercomputer were disclosed today by Lawrence Livermore National Laboratory’s (LLNL) Terri Read more…

Tesla Bulks Up Its GPU-Powered AI Super – Is Dojo Next?

August 16, 2022

Tesla has revealed that its biggest in-house AI supercomputer – which we wrote about last year – now has a total of 7,360 A100 GPUs, a nearly 28 percent uplift from its previous total of 5,760 GPUs. That’s enough GPU oomph for a top seven spot on the Top500, although the tech company best known for its electric vehicles has not publicly benchmarked the system. If it had, it would... Read more…

Exclusive Inside Look at First US Exascale Supercomputer

July 1, 2022

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…

Leading Solution Providers

Contributors

AMD Opens Up Chip Design to the Outside for Custom Future

June 15, 2022

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…

Nvidia, Intel to Power Atos-Built MareNostrum 5 Supercomputer

June 16, 2022

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…

UCIe Consortium Incorporates, Nvidia and Alibaba Round Out Board

August 2, 2022

The Universal Chiplet Interconnect Express (UCIe) consortium is moving ahead with its effort to standardize a universal interconnect at the package level. The c Read more…

Using Exascale Supercomputers to Make Clean Fusion Energy Possible

September 2, 2022

Fusion, the nuclear reaction that powers the Sun and the stars, has incredible potential as a source of safe, carbon-free and essentially limitless energy. But Read more…

Is Time Running Out for Compromise on America COMPETES/USICA Act?

June 22, 2022

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…

Nvidia, Qualcomm Shine in MLPerf Inference; Intel’s Sapphire Rapids Makes an Appearance.

September 8, 2022

The steady maturation of MLCommons/MLPerf as an AI benchmarking tool was apparent in today’s release of MLPerf v2.1 Inference results. Twenty-one organization Read more…

India Launches Petascale ‘PARAM Ganga’ Supercomputer

March 8, 2022

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…

Not Just Cash for Chips – The New Chips and Science Act Boosts NSF, DOE, NIST

August 3, 2022

After two-plus years of contentious debate, several different names, and final passage by the House (243-187) and Senate (64-33) last week, the Chips and Science Act will soon become law. Besides the $54.2 billion provided to boost US-based chip manufacturing, the act reshapes US science policy in meaningful ways. NSF’s proposed budget... Read more…

  • arrow
  • Click Here for More Headlines
  • arrow
HPCwire