AMD Announces Three Radeon Instinct Compute Cards Coming Q3 – Includes MI25, MI8 & MI6

While late last year was all the hype surrounding Zen and Polaris for AMD, this year, it’s EPYC, Threadripper, Vega, and now Instinct. The role of the GPU has changed somewhat over the years, moving beyond simple graphics accelerators. GPUs now do general purpose compute, ray tracing, video rendering, finance, and now AI and deep learning, which is where Instinct comes in.

While NVIDIA has its Tesla series compute cards, AMD now has its Radeon Instinct series. These cards are headless (no display output), designed to fit in servers and workstations, used for complex math and huge datasets; although they can still be used for graphics. The main focus though, is all the latest industry buzzwords surrounding HPC (High Performance Computing), Artificial Intelligence, deep learning, and machine intelligence. It’s the enterprise market in general that AMD is after, and it’s committed to a broad range of open standards to get there, including HSA, ROCm, and MIOpen (all of which tie-in and extend AMD’s GPUOpen initiative).

Radeon Instinct’s initial launch will leave some scratching their heads, as the three cards (MI25, MI8, and MI6) are not from a single generation, but spread across the last three Radeon architectures; Vega, Polaris, and Fiji. Not only that, but the architecture has no bearing on performance, as the middle performer, the M18, is a compute variant of the Fury cards, i.e. it’s a Fiji-based GPU. This is because of the inheritance from the original AMD FirePro S-series compute cards being rebranded, although Instinct supports new versions of AMD’s ROCm software suite. The table below will give you a better overview of each of the three cards and how they line up.

	AMD Radeon Instinct Compute Cards
	MI25	MI8	MI6
Cores	4096	4096	2304
Memory	16GB HBM2 (ECC)	4GB HBM1	16GB GDDR5
Memory Bandwidth	484 GB/s	512 GB/s	224 GB/s
Double-precision	768 GFLOPS	512 GFLOPS	358 GFLOPS
Single-precision	12.3 TFLOPS	8.2 TFLOPS	5.7 TFLOPS
Half-precision	24.6 TFLOPS	8.2 TFLOPS	5.7 TFLOPS
TDP	300W	175W	150W
Architecture	Vega	Fiji	Polaris
Size	Dual Slot / 10.5″	Dual Slot / 6″	Single Slot / 9.5″

Just as an aside, it only came to my attention while writing this post that the MI in each name is not M1 (M One), but stands for Machine Intelligence; I mean, no one ever confuses 1 and I, do they? Mini-rant aside, the three cards are definitely an oddball mix of generations, but each cater to a different audience, size, and task, with careful consideration to server density.

Radeon Instinct supports the latest ROCm 1.6 software platform, which works as its own development environment and supporting language for AMD’s Heterogeneous System Architecture (HSA) which allows for simultaneous programming of both GPU and CPU within the same application, as well as cross-compiling of code (including some CUDA support), and a broad range of libraries.

The MI25 also makes liberal mention of AMD’s new server platform, EPYC (formally Naples), that makes use of the new Zen architecture powering Ryzen, and soon, Threadripper. From what we can gather, the quoted “configuration advantages with EPYC server processors” most likely comes down to server density, as EPYC allows for 128 PCIe lanes, allowing for multiple MI25 cards to be accessible from a single server.

ROCm 1.6 will be made available June 29th, ahead of Radeon Instinct’s official launch. The compute cards themselves will begin shipping to system integrators and HPC builders some time Q3 this year.