Nvidia Gets SALTY

[Bofferbrauer2]

Pemalite said: 

Bofferbrauer2 said:

Just to add in something I wasn't sure at that point, Vega 10 FP64 can only can reach 1/16th of FP32 while Vega 20 can do 1/2, so the chip must have changed quite a bit under the hood. Double Precision is also something that cost quite a lot of energy, hence why it got cut down  in modern GPUs both by AMD and NVidia. NVidia nowadays mostly uses 1/32th, and all RTX cards, including the Quadro, do so. The original Titan had 1/4, while the Volta-based Titan V has 1/2 like the Vega 20, making the latter the probable target if the FP64 capability of the Radeon VII hasn't been cut down.

Not really. Graphics Core Next is extremely modular remember, you can update part of a chip and leave the rest identical.
Besides... AMD has a Vega GPU with 1/2 FP64 on the market right now... Meaning that the design of Vega 7 isn't new anyway.

In short, Vega 7 is a simply a GPU ported to 7nm with a CU removed to increase yields... Doubling of DRAM and Bandwidth and a big increase in clockrates, it was minimal effort by AMD... And because of that, it's unlikely it will beat a Geforce 1080Ti.

Vega 10, of which Vega 64 and Vega 56 are based on, all have only 1/16 FP64 performance, including Radeon Pro, MI25 and Frontier Edition. Only Vega 20, which is the chip used in the MI50 and MI60 accelerator cards and most probably in the Radeon VII, has that 1/2 FP64 performance.

In short, no, it's not a GPU ported to 7nm with 4 CU removed for yields, but instead a re-purposed professional card that already comes in 7nm. It's minimum effort in so far that that halved the RAM, not doubled it,  and basically kept the speed at the same level as the MI60. And no, 1080Ti doesn't seem to be unbeatable at all by this card.

Last edited by Bofferbrauer2 - on 16 January 2019