REGISTER | LOGIN
Breaking News
News & Analysis

Intel’s Nervana Attacks GPUs

AI strategy incomplete without Movidius
11/17/2016 11:00 PM EST
5 comments
NO RATINGS
1 saves
Page 1 / 3 Next >
More Related Links
View Comments: Newest First | Oldest First | Threaded View
Gondalf
User Rank
Author
Re: ...
Gondalf   11/20/2016 4:10:22 PM
NO RATINGS
I don't think the process node is so much relevant here. Actual 28nm, expecially from TSMC, is right now a good choice for high power devices. 16nm HP is not as good like low power 16nm SOC, the trend is pretty evident. Upcoming finer HP processes are not a mess like in the golden era of silicon scaling, they help a bit but they can not solve the problems, like in old days, against new solutions. But this is only IMO

 

witeken
User Rank
Freelancer
Re: ...
witeken   11/18/2016 1:13:34 PM
NO RATINGS
That's not really the point, I think. Intel will have specialized chip vs. Nvidia general purpose. Of course the dedicated chip in theory has advantage. Only hurdle is process node, it seems they plan to move it to Intel foundry by 2019/2020.

realjjj
User Rank
CEO
Re: ...
realjjj   11/18/2016 11:14:06 AM
NO RATINGS
Would be a bit foolish to make a guess with little info.

Based on history, they would likely keep evolving the GPU to better fit this task, the question is how much and how fast. They should have different silicon than in consumer but likely not hugely different just yet - the differences would increase as the TAM increases.

For Tegra Xavier aimed at cars, pretty sure that they will have a computer vision accelerator, not "just" the CPU and GPU.

For them there is the risk of getting stuck on the idea of using the GPU but it might be that, for now, they are just navigating the present by trying to stay ahead while keeping costs in check.  So maybe there is room to be optimistic that, once the algorithms are more mature and the TAM large enough, they'll have something that doesn't resemble a GPU at all.

AMD is moving forward too.We know that the PS4 Pro supports packed 16bit formats so it is safe to assume that Vega will include the feature and their software efforts (Radeon Open Compute Platform) are starting to take shape. AMD should have a "yuge" CES with Zen and Vega and maybe ,if all goes well, they'll start to be able to afford to invest a bit more going forward.

rick merritt
User Rank
Author
Re: ...
rick merritt   11/18/2016 10:19:16 AM
NO RATINGS
Good point. They already have a 16nm GPU with a cache coherent link to an IBM Power host which is a step ahead of Nervana's 28nm accelerator on a PCI Express bus.

What would you guess Nvidia will have in mid-2017 when Nervana ships?

realjjj
User Rank
CEO
...
realjjj   11/18/2016 2:16:56 AM
NO RATINGS
Nvidia isn't standing still though.  A couple of months ago Nvidia teased their next Tegra, code named Xavier and the perf claim was 20 deep learning tera-ops ( measures 8-bit integer ops). The fact that they aren't using FP16 and FP32 when talking perf is rather interesting. Don't have more info but it smells like they are about to take a huge step forward next year.

Like Us on Facebook
EE Times on Twitter
EE Times Twitter Feed