Sharon AI Launches Cloud Media Tech with On-Demand GPU Service
Menu
AMD Instinct MI300X
The AMD Instinct MI300X accelerator, built on next-gen AMD CDNA™ 3 architecture, delivers exceptional efficiency and performance for AI and high-performance computing (HPC) tasks. With 192GB of HBM3 memory and 5.3 TB/s of memory bandwidth, it is designed for demanding AI workloads, generative AI, machine learning training, and large language models, offering significant performance improvements over its predecessors.
Basic Product Information
Product Name
AMD Instinct MI300X
Architecture
AMD CDNA™ 3
Memory
192GB HBM3
Compute Units
304
Release Year
2023
Use Cases
AI/ML training, generative AI, large language models, HPC
Key Advantages
192GB HBM3 Memory
Largest memory capacity in its class, ideal for data-intensive AI and HPC tasks.
5.3TB/s Memory Bandwidth
Unprecedented bandwidth for faster data processing.
304 Compute Units
Enhanced performance for AI and HPC applications.
Up to 13.7X Peak AI/ML Performance
Superior efficiency compared to previous models.
Multi-Chip Architecture
Improves power efficiency and reduces data movement overhead.
Specifications
Performance Specifications
AI Peak Performance (TFLOPS)
FP8: 5,229.8
INT8: 5,229.8
FP16: 2,614.9
TF32: 1,307.4
HPC Peak Performance (TFLOPS)
FP64 Vector
81.7
FP32 Vector
163.4
Memory and Bandwidth
GPU Memory
192GB HBM3
Memory Bandwidth
5.3TB/s
Memory Interface
8192 bits
AMD Infinity Cache™
256MB
Thermal and Power
Max Thermal Design Power (TDP)
750W
Board Specifications
Form Factor
OAM module
Interconnect
7x AMD Infinity Fabric™ Links (128GB/s each) 1 PCIe® Gen 5 x16 (128GB/s)
Supported Technologies
Multi-Chip Architecture
Uses 8 accelerated compute dies with enhanced sparsity and computational throughput.
Decoders
Supports HEVC/H.265, AVC/H.264, VP9, and AV1 codecs.
Virtualization
SR-IOV with up to 8 partitions for resource sharing.
RAS Features
Full-chip ECC memory, page retirement, and page avoidance for reliability.
Server Compatibility
Part of the AMD Instinct Platform with 8 interconnected GPUs on a Universal Base Board (UBB 2.0), compatible with industry-standard HGX host connectors.
Additional Features
01 Infinity Fabric™ Technology
Superior I/O efficiency and scaling between GPUs.
02 Shared Memory and Caches
Coherently shares memory across CPUs and GPUs for better AI and HPC performance.
03 ROCm Software Support
Comprehensive support for AI and HPC frameworks, including PyTorch, TensorFlow, and JAX, with optimized libraries for AMD accelerators.