ThinkSystem AMD MI300X 192GB 750W 8-GPU Board

The AMD Instinct MI300X 192GB 750W Accelerator is a GPU based on next-generation AMD CDNA 3 architecture, delivering leadership efficiency and performance for the most demanding AI and HPC applications. Eight MI300X accelerators are integrated into servers such as the ThinkSystem SR685a V3.

It is designed with 304 high throughput compute units, AI-specific functions including new data-type support, photo and video decoding, plus an unprecedented 192 GB of HBM3 memory on a GPU accelerator. Using state-of-the-art die stacking and chiplet technology in a multi-chip package propels generative AI, machine learning, and inferencing, while extending AMD leadership in HPC acceleration.

Figure 1. ThinkSystem AMD MI300X 192GB 750W 8-GPU Board

The ThinkSystem SR685a V3 includes 8x MI300X GPUs that are fully interconnected using AMD Infinity Fabric which provides 128 GB/s bandwidth between each of the 8 GPUs, for a total of 896 GB/s.

The following table shows the part numbers for the 8-GPU board.

Feature code C1HK contains 8x MI300X GPUs plus the Infinity Fabric high-speed interconnections.

The AMD Instinct MI300X accelerator offers the following features:

• Designed to Accelerate Modern Workloads

  The increasing demands of generative AI, large-language models, machine learning training, and inferencing puts next-level demands on GPU accelerators. The discrete AMD Instinct MI300X GPU delivers leadership performance with efficiency that can help organizations get more computation done within a similar power envelope compared to MI250X accelerators from AMD. For HPC workloads, efficiency is essential, and AMD Instinct GPUs have been deployed in some of the most efficient supercomputers on the Green500 supercomputer list2, these types of systems— and yours—can take now take advantage of a broad range of math precisions to push highperformance computing (HPC) applications to new heights.

• Based on 4th Gen Infinity Architecture

  The AMD Instinct MI300X is one of the first AMD CDNA 3 architecturebased accelerators with high throughput based on improved AMD Matrix Core technology and highly streamlined compute units. AMD Infinity Fabric™ technology delivers excellent I/O efficiency, scaling, and communication within and between industry-standard accelerator module (OAM) device packages. Each discrete MI300X offers a 16-lane PCIe® Gen 5 host interface and seven AMD Infinity Fabric links for full connectivity between eight GPUs in a ring. The discrete MI300X is sold as an AMD Instinct Platform with eight accelerators interconnected on an AMD Universal Base Board (UBB 2.0) with industry-standard HGX host connectors.

• Multi-Chip Architecture

  The MI300X uses state-of-the-art die stacking and chiplet technology in a multi-chip architecture that enables dense compute and highbandwidth memory integration. This helps reduce data-movement overhead while enhancing power efficiency.

  Each OAM module includes:

  • Eight accelerated compute dies (XCDs) with 38 compute units (CUs), 32 KB of L1 cache per CU, 4 MB shared L2 cache shared across CUs, and 256 MB of AMD Infinity Cache™ shared across 8 XCDs. The compute units support a broad range of precisions for both AI/ML and HPC acceleration, native hardware support for sparsity, and enhanced computational throughput.
  • Four supported decoders for HEVC/H.265, AVC/H.264, V1, or AV1, each with an additional 8-core JPEG/MPEG CODEC
  • 192 GB of HBM3 memory shared coherently between CPUs and GPUs with 5.3 TB/s on-package peak throughput
  • SR-IOV for up to 8 partitions
• Coherent Shared Memory and Caches

  Machine-learning and large-language models have become highly data intensive, and they need to split jobs across multiple GPUs. AMD Instinct accelerators facilitate large models with shared memory and caches. The large amount of HBM3 memory is supported with 5.3 TB/s of local bandwidth, and direct connectivity of 128 GB/s bidirectional bandwidth between each GPU, accelerating memory-intensive AI, ML, and HPC models.

• AMD ROCm 6 Open Software Platform for HPC, AI, and ML Workloads

  Whatever your workload, AMD ROCm software opens doors to new levels of freedom and accessibility. Proven to scale in some of the world’s largest supercomputers, ROCm software provides support for leading programing languages and frameworks for HPC and AI. With mature drivers, compilers and optimized libraries supporting AMD Instinct accelerators, ROCm provides an open environment that is ready to deploy when you are.

• Propel Your Generative AI and Machine Learning Applications

  Support for the most popular AI & ML frameworks—PyTorch, TensorFlow, ONYX-RT, Triton and JAX—make it easy to adopt ROCm software for AI deployments on AMD Instinct accelerators. The ROCm software environment also enables a broad range of AI support for leading compilers, libraries and models making it fast and easy to deploy AMD based accelerated servers. The AMD ROCm Developer Hub provides easy access point to the latest ROCm drivers and compilers, ROCm documentation, and getting started training webinars, along with access to deployment guides and GPU software containers for AI, Machine Learning and HPC applications and frameworks.

• Accelerate Your High Performance Computing Workloads

  Some of the most popular HPC programing languages and frameworks are part of the ROCm software platform, including those to help parallelize operations across multiple GPUs and servers, handle memory hierarchies, and solve linear systems. Our GPU Accelerated Applications Catalog includes a vast set of platform-compatible HPC applications, including those in astrophysics, climate & weather, computational chemistry, computational fluid dynamics, earth science, genomics, geophysics, molecular dynamics, and physics. Many of these are available through the AMD Infinity Hub, ready to download and run on servers with AMD Instinct accelerators.

The following table lists the MI300X accelerator specifications.

The following tables list the ThinkSystem servers that are compatible.

Operating system support is based on that of the supported servers. See the SR685a V3 server product guide for details: https://lenovopress.lenovo.com/lp1910-thinksystem-sr685a-v3-server

The MI300X accelerator has the following regulatory approvals:

• Electromagnetic Compliance
  • Australia and New Zealand: CISPR 32: 2015 +COR1: 2016, Class A
  • Canada ICES-003, Issue 7, Class A
  • European Countries: EN 55032: 2015 + A11: 2020 Class B, EN 55024: 2010, EN 55035: 2017
  • Japan VCCI-CISPR32:2016, VCCI 32-1: 2016 Class A
  • Korea KN32, Class A, KN35, RRA Public Notification 2019-32
  • Taiwan CNS 13438: 2016, C6357, Class A
  • USA FCC 47 CFR Part 15, Subpart B, Class A
• Product Safety Compliance
  • UL 62368-1, 2nd Edition, 2014-12
  • CSA-C22.2 No. 62368-1, 2nd Edition, 2014-12
  • EN 62368-1, 2nd Edition, 2014 + A1: 2017
  • IEC 62368-1, 2nd Edition, 2014
  • RoHS Compliance: EU RoHS Directive (EU) 2015/863 Amendment to EU RoHS 2 (Directive 2011/65/EU)
  • REACH Compliance
  • Halogen Free: IEC 61249-2-21:2003 standard

The MI300X accelerator assumes the server’s base warranty and any warranty upgrades.

For more information, refer to these documents:

• ThinkSystem and ThinkAgile GPU Summary:
  https://lenovopress.lenovo.com/lp0768-thinksystem-thinkagile-gpu-summary
• ServerProven compatibility:
  https://serverproven.lenovo.com/
• AMD MI300X product page:
  https://www.amd.com/en/products/accelerators/instinct/mi300/mi300x.html