Powerful Industry-Standard 8-GPU Solution

Today’s large-scale AI/ML training sets and HPC data need three elements to accelerate workloads: fast acceleration across multiple data types, large memory and bandwidth to handle huge data, and extreme I/O bandwidth. You get all three with the AMD Instinct™ MI300X Platform with 3rd Gen AMD CDNA™ architecture-based GPUs: 42 petaFLOPs of peak theoretical FP8 with sparsity precision performance for generative AI and ML training and 1.3 petaFLOPs peak theoretical FP32 precision for the most challenging HPC codes. Our industry-standard-based universal baseboard (UBB 2.0) platform hosts 8 AMD Instinct™ MI300X accelerators and 1.5 TB of HBM3 memory to help process the most demanding AI models and HPC workloads. With eight x16 PCIe® Gen 5 host I/O connections, you don’t have to worry about data bottlenecks. The bottom line is a platform that’s based on open standards that incorporate proven AMD Instinct™ technology that is expected to drive some of the world’s fastest supercomputers, and an open software platform that is ready to support you.

AI PEAK THEORETICAL PERFORMANCE

TF32	5.2 PFLOPs	10.5 PFLOPs
FP16	10.5 PFLOPs	20.9 PFLOPs
BFLOAT16	10.5 PFLOPs	20.9 PFLOPs
INT8	20.9 PFLOPs	41.8 PFLOPs
FP8	20.9 PFLOPs	41.8 PFLOPs

HPC PEAK THEORETICAL PERFORMANCE

FP64 vector		653.7 TFLOPs
FP32 vector		1307.4 TFLOPs
FP64 matrix		1307.4 TFLOPs
FP32 martrix		1307.4 TFLOPs

DECODERS AND VIRTUALIZATION

Decoders*	32 groups for HEVC/H.265, AVC/H.264,  V1, or AV1
JPEG/MJPEG CODEC	256 cores, 8 cores per group
Virtualization support	SR-IOV, up to 64 partitions

* Video codec acceleration (including at least the HEVC (H.265), H.264, VP9, and AV1 codecs) is subject to and not operable without inclusion/installation of compatible media players. GD-176

SPECIFICATIONS

Form factor	Universal baseboard (UBB) module with 8 Instinct MI300X OAM GPUs
Lithography	5nm FinFET
Active interposer dies (AIDs)	6nm FinFET
GPU compute units	2432
Matrix cores	9728
Stream processors	155,648
Peak engine clock	2100 MHz
Memory capacity	1.5 TB HBM3
Memory bandwidth	5.3 TB/s max. peak theoretical
Memory interface	8192 bits per GPU
AMD Infinity Cache™ (last level)	256 MB per GPU
Memory clock	Up to 5.2 GT/s
Scale-up Infinity Fabric™ Links	7x 128 GB/s per GPU
Ring of 8 aggregate bandwidth	896 GB/s
Scale-out network bandwidth	8 PCIe® Gen 5 x16 (128 GB/s) per GPU
RAS features	Full-chip ECC memory, page retirement, page avoidance
Maximum TBP	750W per GPU

AMD Instinct MI300X Platform

To offer the power of the AMD Instinct MI300X accelerator through industry-standard servers, we have designed a platform to combine the power of eight accelerators on an industry-standard universal baseboard (UBB 2.0). The eight Open Compute Project (OCP) Accelerator Modules (OAMs) are connected with an AMD Infinity Fabric™ mesh that provides direct connectivity between each of the GPUs over 128 GB/s bidirectional links. Each MI300X connects with its peers through seven links, plus one PCIe® Gen 5 x16 connection per OAM device provides upstream server and/or I/O connectivity. Remote DMA I/O transfers can stream data to each GPU where it is needed and where it can be processed in each module’s large 192 GB HBM3 memory.

Based on 4th Gen Infinity Architecture

The AMD Instinct MI300X accelerator is based on the 4th Gen Infinity architecture and the AMD CDNA™ 3 architecture offers high throughput based on generationally improved AMD Matrix Core technology and streamlined compute units. The AMD Instinct MI300X GPU also supports PCIe® Gen 5 with AMD Infinity Fabric™ technology helping to improve I/O performance, efficiency, and scaling within and between each OAM device on the universal baseboard.

More information about MI-Series

© 2023 Advanced Micro Devices, Inc. All rights reserved. AMD, the AMD Arrow logo, AMD Instinct, Infinity Cache, Infinity Fabric, ROCm, and combinations thereof are trademarks of Advanced Micro Devices, Inc. PCIe is a registered trademark of PCI-SIG Corporation. PyTorch, the PyTorch logo and any related marks are trademarks of Facebook, Inc. TensorFlow, the TensorFlow logo, and any related marks are trademarks of Google Inc. Other product names used in this publication are for identification purposes only and may be trademarks of their respective companies. Use of third party marks/logos/products is for informational purposes only and no endorsement of or by AMD is intended or implied GD-83  PID # 232405395-D