VMware Private AI with Intel on Lenovo ThinkAgile VX V3 and ThinkSystem V3

Lenovo ThinkSystem SR650 V3 2U and SR630 V3 1U servers and ThinkAgile VX650 V3 2U and VX630 V3 2U hyperconverged solutions with VMware vSAN powered by 4^th and 5^th gen Intel Xeon Scalable processors are optimized for AI workloads and Accelerated by Intel offerings. Lenovo V3 systems support up to 64 cores per socket with 5^th Gen Intel Xeon processors and up to 60 cores per socket with 4^th Gen processors.

ThinkAgile VX V3 systems are factory-integrated, pre-configured ready-to-go integrated systems built on proven and reliable Lenovo ThinkSystem servers that provide compute power for a variety of workloads and applications and are powered by industry-leading hyperconverged infrastructure software from VMware. It provides quick and convenient path to implement a hyperconverged solution powered by VMware Cloud Foundation (VCF) or VMware vSphere Foundation (VVF) software stacks with "one-stop shop" and a single point of contact provided by Lenovo for purchasing, deploying, and supporting the solution.

Intel provides a comprehensive portfolio of AI development software including data preparation, model development, training, inference, deployment, and scaling. Using optimized AI software and developer tools can significantly improve AI workload performance, and developer productivity, and reduce compute resource usage costs. Intel® oneAPI libraries enable the AI ecosystem with optimized software, libraries, and frameworks. Software optimizations include leveraging accelerators, parallelizing operations, and maximizing core usage.

Intel® Advanced Matrix Extensions (Intel® AMX)

Intel® AMX is a new set of instructions designed to work on matrices and it enables AI fine-tuning and inference workloads to run on the CPU. Its architecture supports bfloat16 (training/inference) and int8 (inference) data types and Intel provides tools and guides to implement and deploy Intel AMX. The Intel AMX architecture is designed with two components,

1. Tiles: These consist of eight two-dimensional registers, each 1 kilobyte in size, that store large chunks of data.
2. Tile Matrix Multiplication (TMUL): TMUL is an accelerator engine attached to the tiles that performs matrix-multiply computations for AI.

Refer more information about Intel AMX here.

With integrated Intel AMX on 4^th and 5^th gen Intel Xeon Scalable processors, many AI inferencing and fine-tuning workloads, including many Generative AI use cases, can run optimally.

Intel AI software and optimization libraries provide scalable performance using Intel CPUs and GPUs. Many of the libraries and framework extensions are designed to leverage the CPU to provide optimal performance for machine learning and inference workloads. Developers looking to leverage these tools can download the AI Tools from AI Tools Selector.

Table 1: Intel AI optimization software and development tools

VMware Cloud Foundation (VCF) is a multi-cloud platform supporting virtual machines and containerization of workloads on common virtualized infrastructure built on top of vSphere, vSAN, and NSX. The suite includes VMware Aria Suite for private/hybrid cloud management and VMware Tanzu for Kubernetes workloads. Refer more details in VMware Cloud Foundation reference design here.

VMware Private AI with Intel solution enables enterprises to develop and deploy classical machine learning models and generative AI applications on the infrastructure powered by Intel AI software and built-in accelerators and managed by VMware Cloud Foundation. VCF provides integrated security capabilities to secure AI, and it is an ideal platform for training and running private LLMs across business functions in an enterprise. The Intel AI software suite and VMware Cloud Foundation are validated on Lenovo ThinkSystem and ThinkAgile servers with 4th and 5th gen Intel Xeon Scalable processors and private LLMs or generative AI models can be deployed at scale along with other AI use cases.

Intel AMX instructions are supported on vSphere 8.0 and above with VMs using virtual HW version 20 and above. The guest OS running Linux should use kernel 5.16 or later and if Tanzu Kubernetes is used, the worker nodes should use Linux kernel 5.16 or later.

Figure 1. VMware Private AI with Intel on Lenovo ThinkAgile VX and ThinkSystem Servers

The Generative AI inference testing with Llama2 7B and 13B model was done on ThinkAgile VX650 V3 server with 4^th Gen Intel Xeon Scalable processors by Intel on May 14, 2024. The test was carried out with different input token sizes 32/256/1024/2048 with varying batch sizes of 1-16 to simulate concurrent requests with static output token size 256. The objective of the testing is to validate different scenario's performance with acceptable latency of less than 100ms latency and to compare the results with/without Intel AMX.

The test and inference serving is targeted on a single node with local storage running ESXi 8.0 U2 and two Ubuntu 22.04.4 guest virtual machines. The model performance can be scaled out by using multiple nodes, but it is not in scope of the current version.

Table 2. Test Hardware Configuration

Table 3. Test Configuration

Figure 2 shows the 2^nd token average latency performance with Intel AMX on 4^th gen Intel Xeon Scalable processors for Llama 7B model and Figure 3 shows the results without Intel AMX. The test with Intel AMX shows up to 42% in 2^nd token latency for the scenario with input/output token size 32/256. The 2^nd token latency for different concurrent requests scenarios (batch sizes 1/2/4/8/16) with input/output token size of 32/256, and 256/256 are within an acceptable threshold of 100 milliseconds and it shows significant throughput increase can be achieved with Intel AMX. The results without Intel AMX show all the scenarios with batch size 8/16 exceeded the 100 milliseconds threshold.

Figure 2. Llama2 7B testing with 4th Gen Intel Xeon CPUs with Intel AMX - 2nd token average latency

Figure 3. Llama2 7B testing with 4th Gen Intel Xeon CPUs without Intel AMX - 2nd token average latency

Figure 4 shows the 2^nd token average latency performance with Intel AMX on 4^th gen Intel Xeon Scalable processors for Llama 13B model and Figure 5 shows the results without Intel AMX. The test with Intel AMX shows up to 18% decrease in 2^nd token latency for the scenario with input/output token size 32/256. The 2^nd token latency for different concurrent user scenarios (batch sizes 1/2/4/8) with input token size of 32/256 are within an acceptable threshold of 100 milliseconds and it shows considerable throughput increase can be achieved with Intel AMX. The results without Intel AMX shows most of the scenarios with batch size 4/8/16 exceeded the 100ms next token latency threshold.

Figure 4. Llama2 13B testing with 4th Gen Intel Xeon CPUs with Intel AMX - 2nd token average latency

Figure 5. Llama2 13B testing with 4th Gen Intel Xeon CPUs without Intel AMX - 2nd token average latency

To deliver the best experience possible, Lenovo and Intel have optimized this solution to leverage Intel capabilities like processor accelerators not available in other systems. Accelerated by Intel means enhanced performance to help you achieve new innovations and insight that can give your company an edge.

To learn more about this Lenovo solution contact your Lenovo Business Partner or visit: https://www.lenovo.com/us/en/servers-storage/solutions/database/

References:

Lenovo ThinkAgile VX650 V3 2U Integrated System and VX650 V3 2U Certified Node

ThinkAgile VX630 V3 1U Integrated System and Certified Node

Intel AI Development Software