Today’s compute workloads are more complex than ever before and require working with vast amounts of increasingly complex data. System designs need to be more open and to be more flexible to address these applications that have become more parallel in nature, while continuing to improve overall system efficiencies when deploying large systems. AMD’s Radeon Instinct™ family of products have been distinctly designed and optimized to help address these growing needs with open heterogeneous computing solutions bringing about the next era of compute and machine intelligence.

The Radeon Instinct MI6 is a unique, versatile accelerator for Deep Learning edge-training and inference applications delivering 5.7 TFLOPS of peak half or single precision performance in a single-slot GPU card equipped with 16GB of ultra-fast memory at under 150 watts of TDP.4 The Radeon Instinct MI6 accelerator’s distinctive capabilities, combined with AMD’s open ROCm software platform and MIOpen Machine Learning framework libraries, provide customers with a highly versatile, cost-sensitive compute solution for today’s most demanding Deep Learning edge-training and inference applications.

Specification

GPU Architecture:                         AMD “Polaris”

Stream Processors:                      2,304

Performance:

Half-Precision (FP16)                    5.7 TFLOPS

Single-Precision (FP32)                5.7 TFLOPS

Double-Precision (FP64)              358 GFLOPS

GPU Memory:                               16GB gM1

Memory Bandwidth:                    Up to 224 GB/s

Bus Interface:                                PCIe® Gen 3 Compliant Motherboard1

MxGPU Capability:                        Yes

Board Form Factor:                       Full-Height, Single-slot

Length:                                            9.5"

Thermal Solution:                          Passively Cooled

Standard Max Power:                   150W TDP

OS Support:                                    Linux® 64-bit

ROCm Software Platform:            Yes

Programing Environment:

ISO C++, OpenCL™, CUDA (via AMD’s HIP conversion tool) and Python5 (via Anaconda’s NUMBA)