Business Challenge: Do you get the best business value from your data? vSAN clusters are home to increasingly large and valuable data stores. Users can take advantage of AI capabilities that are built into many enterprise applications and analytics tools to help accelerate Solution Overview and Summary.
Solution Overview and Summary
Solution: Intel and VMware offer a new solution that significantly improves performance, underscoring the need for ongoing data center modernization. VMware vSAN 8 includes the new Express Storage Architecture (ESA), an optional alternative architecture that can process and store data with elevated efficiency, scalability, and performance levels. With 4th Generation Intel® Xeon® Scalable processors, VMware vSphere and vSAN 8 software can power the most demanding workloads.
While this solution, with the latest generation of Intel® hardware and VMware software, supports various applications, the benefits for AI workloads like image classification and natural language processing (NLP) are particularly compelling. The built-in AI acceleration from Intel® Advanced Matrix Extensions (Intel® AMX) is designed to provide massive speedup to the low‑precision math operations that underpin AI inference. Mainstream applications already running on vSAN and Intel Xeon processors—such as databases, analytics, business-critical and collaboration applications, and IT automation tools—are being enhanced with AI algorithms and can benefit from Intel AMX. The result is a completely optimized pipeline on a single hardware and software platform that can scale from data center to cloud to edge. Scale AI everywhere using the broad, open software ecosystem and unique Intel tools. Customers can utilize their large and valuable vSAN data store on standard Intel Xeon processor-based servers while gaining the efficiency and performance of a built‑in AI accelerator.
Intel AMX supports INT8 and BF16 data types, augmenting the optimizations from Intel® Advanced Vector Extensions 512 and Intel® Deep Learning Boost to enable fast and efficient AI and deep learning across various industries and use cases.
Results: The testing highlighted in this document demonstrates the performance benefits of vSphere/vSAN 8 on 4th Gen Intel Xeon Scalable processors with Intel AMX, compared to previous generations of hardware, as well as the speedup provided by using BF16 versus FP32. For example, image classification and NLP are about 3x faster gen-over-gen at INT8 precision. Additionally, using BF16 instead of FP32 delivers 4.1x to 5x faster inference with nearly no loss in precision.1 See the next page for a full discussion of testing results.
“Customers can utilize their large and valuable vSAN data store on standard Intel Xeon processor-based servers while gaining the efficiency and performance of a built-in AI accelerator.”
Key Solution ComponentsIntel® Optimization for TensorFlow 2.11 with the Intel® oneAPI Deep Neural Network Library (oneDNN). This library can natively take advantage of Intel® AMX instructions to accelerate AI workloads. Pre-built containers and validated VMs from Model Zoo, a publicly available repository created by Intel and now available on GitHub. In particular, we used:
|
Test Methodology
The testing methodology included tests for NLP use cases (using the BERT-Large model) and object detection and image classification use cases (using the ResNet-50v1.5 model). We benchmarked these two models for AI inference performance at various precisions (FP32, BF16, and INT8). See the “Profiles and Workloads” section for more details on configuring the tests.
FP32 is a standard 32-bit floating-point data type used to train deep learning models and for inferencing. This data type is more computationally demanding than other data types but typically achieves higher accuracies. BF16 is a truncated version of FP32 used for training and inference. It offers similar accuracy but faster computation. INT8 offers higher performance and is the least computationally demanding data type, ideal for constrained environments. It has minimal impact on accuracy.
We experimented with several batch sizes before determining 128 was ideal. Please look at the sidebar for a list of key components of the solution.
Results
Many deep learning workloads are mixed-precision, and 4th Gen Intel Xeon Scalable processors can seamlessly transition between Intel AMX and Intel AVX-512 to use the most efficient instruction set. Intel publicly distributes pre-optimized versions of the most popular deep learning models, which are compatible with Intel® processors at various precisions. Intel also publishes best-known methods for achieving the highest performance on Intel® architecture, including this solution design brief.
The table below summarizes the notable results. Figures 1 and 2 illustrate the gen-over-gen and precision comparisons for image classification and NLP. BF16 precision (not available on older processors) can provide up to 5x faster performance with virtually no loss of accuracy compared to FP32. Running on newer processors, AI is up to 3.2x faster. If time to insight matters to your business, it’s time to upgrade your hardware and software.
Figure 1. The latest generation of Intel® Xeon® Scalable processor with Intel® AMX delivers up to 3x higher INT8 throughput for image classification inference (compared to older hardware). BF16 provides 5x faster inference on the same processors than FB32 with nearly the same accuracy.*
Figure 2. The latest generation of Intel® Xeon® Scalable processor with Intel® AMX delivers up to 3.2x higher INT8 throughput for NLP inference (compared to older hardware) and up to 4.1x higher throughput than FP32.**
Configuration Details
The following tables provide information about the components and settings of the infrastructure used during testing. These tables reflect typical types of configurations used for each generation of processor.
Profiles and Workloads
Residual Network (ResNet) is a popular deep-learning model for image recognition. We selected ResNet50v1.5 for our benchmarking and ran the tests using synthetic data. Bidirectional Encoder Representations from Transformers (BERT‑Large) is a transformer model that is pre-trained on BookCorpus and English Wikipedia data in a self-supervised fashion. BERT has become the ubiquitous baseline for handling various NLP tasks. We used pre-trained BERT-Large for inferencing with the Stanford Question Answering Dataset (SQuAD) to measure performance.
To prepare the benchmark, we used a single VM with 56 vCPUs on the 3rd Gen Intel Xeon Scalable processor system and 64 vCPUs on the 4th Gen Intel Xeon Scalable processor system. Intel® Hyper-Threading Technology was enabled. The OS was Ubuntu 22.04, with Docker installed. We used a native (OS control) CPUFreq governor and set the BIOS CPU settings to Performance. Tests were conducted in containers downloaded from the relevant repository. Refer to the ResNet50 documentation and the BERT-Large documentation for all necessary steps. The following tables provide details about the various test configurations.
Conclusion
Running your AI workloads using vSphere/vSAN 8 and 4th Gen Intel Xeon Scalable processors can improve both time to insight and time to market for AI solutions (compared to previous-generation Intel processors). Intel and VMware worked together to ensure that the built-in acceleration of AI computation provided by Intel AMX works out-of-the-box, without any vSAN configuration changes. Plus, Intel makes many pre-trained AI models and container images publicly available, compatible with 4th Gen Intel Xeon Scalable processors and Intel AMX, with support for various precision levels. The solution highlighted in this brief consists of a pre-validated hardware and software configuration that is benchmarked and tested so you can confidently run your enterprise AI workloads on vSAN.
Further Information
- 4th Gen Intel® Xeon® Scalable processors
- Intel® Ethernet 800 Series
- VMware vSphere 8
- VMware vSAN with ESA
- ResNet50 model inference execution guidance on Model Zoo
- BERT-Large model inference execution guidance on Model Zoo
Authors:
Ewelina Kamyszek - Cloud Solutions Engineer, DCAI/CESG Intel Group
Patryk Wolsza - Cloud Solutions Architect, vExpert DCAI/CESG Intel Group
Learn more about the Intel and VMware Partnership and Data Center solutions.
Contact your Intel representative to learn more about this solution.
*3rd Gen Intel® Xeon® Scalable platform configuration: Test by Intel as of March 2023. 4-node cluster. Each node: 2x Intel® Xeon® Gold 6348 processor, 1x Intel® Server Board M50CYP2UR, total memory 512 GB (16x 32 GB DDR4 3200 MHz), Intel® Hyper-Threading Technology = enabled, Intel® Turbo Boost Technology = enabled, NUMA enabled noSNC, Intel® Volume Management Device (Intel® VMD) = enabled, BIOS: SE5C620.86B.01.01.0006.2207150335 (microcode:0xd000375), storage (boot) = 2x 80 GB Intel® Optane™ SSD P1600X, storage (ESA flat tier): 9x 3.84 TB Solidigm D7-P5510 SSD, network devices: 1x Intel® Ethernet E810-CQDA2, FW 4.0, at 100 GbE with RDMA using RoCEv2, network speed: 100 GbE, OS/Software: vSphere/vSAN 8.0, 20513097, Ubuntu Server 22.04 VM (vHW=20, vmxnet3), vSAN ESA – optimal default policy RAID-5, Kernel 5.15, Intel® Optimization for TensorFlow 2.11.0, ResNet50v1.5, batch size=128, VM=56 vCPU+64 GB RAM, multi-instance scenario (4 threads per instance), BERT-Large, SQuAD 1.1, batch size=128, VM=56 vCPU+64 GB RAM, multi-instance scenario (28 threads per instance).
**4th Gen Intel Xeon Scalable platform configuration: Test by Intel as of March 2023.4-node cluster. Each node: 2x Intel Xeon Gold 6448Y processor QS pre-production, 1x Intel Server Board M50FCP2SBSTD, total memory 512 GB (16x 32 GB DDR5 4800 MHz), Intel Hyper-Threading Technology = enabled, Intel Turbo Boost Technology = enabled, NUMA enabled noSNC, Intel VMD = enabled, BIOS: SE5C741.86B.01.01.0002.2212220608 (microcode:0x2b000161), storage (boot) = 2x 240 GB Solidigm D3-S4520 SSD, storage (ESA flat tier): 9x 3.84 TB Solidigm D7-P5510 SSD, network devices: 1x Intel Ethernet E810-CQDA2, FW 4.0, at 100 GbE with RDMA using RoCEv2, network speed: 100 GbE, OS/Software: vSphere/vSAN 8.0, 20513097, Ubuntu Server 22.04 VM (vHW=20, vmxnet3), vSAN ESA – optimal default policy RAID-5, Kernel 5.15, Intel Optimization for TensorFlow 2.11.0, ResNet50v1.5, batch size=128, VM=64 vCPU+64 GB RAM, multi-instance scenario (4 threads per instance), BERT-Large, SQuAD 1.1, batch size=128, VM=64 vCPU+64 GB RAM, multi-instance scenario (32 threads per instance).
Performance varies by use, configuration, and other factors. Learn more on the Performance Index site.
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.
Intel and the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries.
You must be a registered user to add a comment. If you've already registered, sign in. Otherwise, register and sign in.