Published: June 21, 2017 | Source: Techpowerup | Author: Mark Campbell

AMD launches their Radeon Instinct MI25

AMD has officially launched their Radeon Instinct MI25 GPU, which offers AMD’s latest Vega GPU hardware with 16GB of HBM2 VRAM.

This GPU will use 64 total NCu compute units, with a total of 4096 Stream processors, which is the maximum that AMD’s Vega GPU is capable of at this time. This core can then offer 12.3TFLOPs of single-precision GPU performance and 24.6TFLOPs of half-precision compute performance.

If users attempt double precision workloads this GPU will only offer 768 GFLOPs of GPU performance, which 16x less than the GPU’s single-precision compute performance. This shows that Vega is not designed for double-precision FP64 workloads, though to say the least most GPUs aren’t. On the Nvidia side, only select Tesla graphics processors offer proper support for FP64 workloads, such as Nvidia’s P100 and V100 GPUs.

This GPU will come with 16GB of ECC HBM2 memory, offering a total of 484GB/s of memory bandwidth, which while less than the R9 Fury X will still offer greater memory performance thanks to updates to AMD’s memory pipeline and the addition of their new Draw-Stream Binning Rasterizer (DSBR).

AMD’s Radeon Instinct MI25 will also come with a full sized, dual slot PCIe form factor with a passive cooling design, though it is worth noting that this GPU is designed to be placed into servers, which will use insanely powerful/loud fans to move plenty of air through these GPUs.

At their EPYC launch event, AMD showcased four of these new GPUs running in a 2U server rack, offering 100 TFLOPS of performance. At this time AMD has not announced the pricing of their new Radeon Instinct MI25 GPUs.

You can join the discussion on AMD’s Radeon Instinct MI25 server graphics processor on the OC3D Forums.