AMD announces Radeon Instinct accelerator with Vega GPU

AMD has announced a series of accelerators for performing gpgpu calculations. One of the cards is equipped with a Vega GPU. According to the specifications, that MI25 accelerator can make more calculations per second than the Tesla P100 with Pascal GPU from Nvidia.

Although AMD puts its new Vega GPU in the MI25 accelerator, the manufacturer provides few details about the specifications of the GPU. According to AMD, the card is good for a processing power of 25Tflops at 16fp half-precision computing. That’s more than Nvidia’s Tesla P100 accelerator, which scores 21.2Tflops at fp16. At what clock speed the GPU runs and how many computing cores the Vega chip has, has not been disclosed.

The plug-in card uses high bandwidth memory, probably hbm2. How much memory there is and what bandwidth it has has not been disclosed. According to AMD, the consumption of the card is less than 300 watts and the card is passively cooled. In practice, however, the card ends up in server racks that are equipped with the necessary fans to dissipate large amounts of heat.

AMD touts the Radeon Instinct MI25 as a ‘training accelerator’ that can be used for deep learning applications. Companies such as Facebook and Google use many of these accelerators. Along with the cards, AMD is introducing the MIOpen platform, an open source deep learning library.

The MI25 is one of three accelerators in the new AMD Instinct series. Only the top model has a Vega GPU on board. The MI6 and MI8 are based on Polaris and Fiji respectively. The former has an fp16 processing power of 5.7Tflops, which is equivalent to the RX 480, and has 16GB of RAM. The MI8 is equipped with 4GB HBM memory, is good for 8.2Tflops and uses the same Fiji GPU as the R9 Fury Nano.

AMD shows that a number of server manufacturers are releasing racks with the MI25 accelerators. For example, Inventec comes with a 39U rack that contains 120 Radeon Instinct MI25 cards, good for 3 petaflops of fp16 computing power. Systems with Radeon Instinct cards should be available in the first half of 2017. AMD has not disclosed pricing.

AMD will later also use its Vega GPU in video cards for gamers. Those video cards are also expected in early 2017. The computing power of a GPU variant for video cards can still differ from the shown accelerator. Nvidia also initially presented its Pascal GPU as an accelerator, with the introduction of the Tesla P100 with hbm2 memory in April. GeForce variants followed later with a modified version of that gpu and gddr5x memory in the GTX 10-series.