The AMD Vega GPU Architecture Teaser: Higher IPC, Tiling, & More, Coming in H1’2017
by Ryan Smith on January 5, 2017 9:00 AM ESTFirst Thoughts: A Peek At What’s To Come
Wrapping things up, while today’s reveal from AMD is only a teaser of what they have been working on over the last few years with Vega, it’s none the less an important one. Based on what we know so far, Vega stands to be the biggest change to AMD’s GPU architecture since GCN 1.0 was released 5 years ago, and the changes to the ALUs, the ROPs, the memory structure, and other aspects of Vega reinforce this notion. To be sure, Vega is not a wholly new architecture – it is clearly a further refinement of GCN – but then this is exactly why GCN was designed to be able to evolve through refinements over a very long period of time.
What we have for now then is a quick look at what’s to come from AMD. There are still many things we don’t know, not the least of which is the actual GPU configurations. But for a teaser it’s enough to show that AMD has been hard at work. It sets the stage for the hardware and marketing ramp-up to come over the next few months.
But for now, let’s close with an image. As I mentioned before, the first Vega has taped out, and Radeon Technology Group’s frontman and Chief Architect, Raja Koduri, has one. The chip was just a few weeks old as of December, and while trying to discern die size may be a little too error-prone, we can see one important detail: 2 HBM2 packages.
Raja and AMD will not tell us what chip we’re looking at – like Polaris, two Vega chips have been confirmed – but either way we are looking at one of them in all its alpha silicon glory. Bearing in mind HBM2’s much greater bandwidth per pin, we could very well be looking at a design for a Fiji-like 512GB/sec of memory bandwidth in the chip Raja holds. And for AMD, that is one more teaser for today to keep attention focused right where they want it: on Vega ahead of its H1’17 launch.
Facebook
Twitter
RSS
155 Comments
View All Comments
FireSnake - Thursday, January 5, 2017 - link
ve.ga is not accessible .... no Ryzen servers yet, haha :)Michael Bay - Thursday, January 5, 2017 - link
Desktop graphics game is lost. AMD should just close all non-gpu divisions and outright reorient towards producing specific ASIC solutions for miners, this newfangled machine bullshit and such.DATS WHERE DEM MONES BE BRAH
rpmrg - Thursday, January 5, 2017 - link
Here's another armchair analyst.Michael Bay - Thursday, January 5, 2017 - link
At least try to hide your envy next time, honey.lobz - Thursday, January 5, 2017 - link
yeah rpmrg dude, don't you ever be envious of his armchair...ddriver - Thursday, January 5, 2017 - link
People are confusing "armchair analyst" with "idiot" :)While it is disappoint for me, as I am predominantly interested in compute, vega has the makings of a good gaming GPU, disappointing for me as this comes from sacrificing high-precision performance for the sake of efficiency and optimizing for low-prevision performance, which is the precision which game graphics utilize.
extide - Thursday, January 5, 2017 - link
Well, the only bit about DP FP on the slides mentioned that it was 'configurable' -- so we don't know what we will see 1/2, 1/3, 1/4... ? Rate vs SP DP? AMD has always been more gratuitous with not disabling that higher end compute on consumer parts -- but we wont know what's actually in the silicon until later. I wouldn't be surprised to see at least one 1/2 rate DP FP part ship at some point, whether it be sold as a consumer card or only as a pro card, only time will tell.eachus - Saturday, January 14, 2017 - link
AMD has stated that Vega 20 (due in 2H18) will have 1/2 DPFP, while Vega 10 will be 1/16th. The configurable means that different chips can have different double precision rates. There is at least one additional Vega chip in the works (Vega 11) plus Navi. However, I think that 1/2 is as fast as we will see. ;-)Michael Bay - Friday, January 6, 2017 - link
But not in your case. ^_^Alexvrb - Friday, January 6, 2017 - link
Why not, his armchair is the brains of the operation! Seriously, he's totes right though... they should just throw away GPU revenue. Take chips that could otherwise be sold in several product categories, and only sell them as "newfangled machine bullshit" chips. Plus some low-margin ephemeral ASICs for miners. See, it's brilliant. Cuts down volume which saves on shipping costs. Plus it means less wafers so you don't get those pesky economies of scale benefits. Nvidia would obviously go this route too but they just don't have the balls to stop selling graphics cards. Excelsior!!