AMD is reportedly working on a Vega 10 dual GPU that could debut in Q2 2017. The card will be targeted at the professional market where it would serve up an impressive amount of computational power for professional level AI applications based on neural networks and deep learning.
We don’t know much about the Vega 10 dual GPU at this point, but considering AMD’s approach towards the dual GPU configuration, it’s safe to assume that this upcoming graphics chip would double the specs of a fully functional Vega 10 GPU. If that is the case, we are looking at 128 compute units and 8192 stream processors. Featuring 16GB of HBM2 per GPU, the card will boast a total of 32GB of HBM2 coupled with 1TB/s of memory bandwidth.
If previous leaks are to be believed, the Vega 10 GPU offers 24 TFLOPs of 16-bit compute power. Given this is half-precision work, the dual chip will have 12 TFLOPs of single precision compute power. The required clock rates for this amount of floating point performance are estimated to be 1465MHz.
However, a dual GPU card is typically a few hundred megahertz slower than its single chip solution so it’s reasonable to assume the Vega 10 dual GPU will run at around 1100MHz. In that case, you’re looking at an effective single precision performance of 18 TFLOPs – which is indeed massive isn’t it?
To put it into perspective, Nvidia’s P100 GPU aimed at professionals and deep learning systems, offers up to 9.3 TFLOPs of single precision compute. Which means the Vega 10 dual GPU should easily beat the P100, at least on paper, until Nvidia responds with a dual GPU solution of its own.
As mentioned above, this monstrous dual GPU based on the Vega 10 architecture is expected to arrive in the second quarter of 2017 although it is likely that the chip is showcased by the end of the year.