Monday, April 11th 2016
Upcoming AMD "Polaris" and "Vega" GPU Compute Unit Counts Surface
AMD's upcoming GPUs based on the "Polaris" and "Vega" architectures appear to be taking advantage of performance/Watt gains to keep stream processor counts low, and chips small, according to a VideoCardz analysis of curious-looking CompuBench entries. Assuming that a Graphics CoreNext (GCN) compute unit (CU) of the "Polaris" architecture, like the three versions of GCN before it, consists of 64 stream processors, AMD's Polaris 11 silicon, codenamed "Baffin," could feature over 1,024 stream processors, across 16 CUs; Polaris 10, codenamed "Ellesmere," could feature over 2,304 stream processors spread across over 36 CUs; and Vega 10 featuring 4,096 stream processors, spread across 64 CUs.
The "Baffin" silicon succeeds current generation "Curacao" silicon, driving mid-range graphics cards. It is expected to feature a 128-bit wide GDDR5 memory interface, holding 4 GB of memory. The "Ellesmere" silicon succeeds current-generation "Tonga" silicon, driving performance-segment SKUs. It could feature up to 8 GB of GDDR5(X) memory. These two chips could see the light of the day by mid-2016. The third chip out of AMD's stable, Vega 10, could succeed "Fiji," overcoming its biggest marketing shortcoming - 4 GB memory. Taking advantage of HBM2 interface, it could feature 16 GB of memory. It could launch some time in early-2017. AMD is claiming a massive 2.5X performance-per-Watt increase for "Polaris" over the current GCN 1.2 architecture, which drives the "Tonga" and "Fiji" chips, and so these stream processor counts could look deceptively insufficient.
Source:
VideoCardz
The "Baffin" silicon succeeds current generation "Curacao" silicon, driving mid-range graphics cards. It is expected to feature a 128-bit wide GDDR5 memory interface, holding 4 GB of memory. The "Ellesmere" silicon succeeds current-generation "Tonga" silicon, driving performance-segment SKUs. It could feature up to 8 GB of GDDR5(X) memory. These two chips could see the light of the day by mid-2016. The third chip out of AMD's stable, Vega 10, could succeed "Fiji," overcoming its biggest marketing shortcoming - 4 GB memory. Taking advantage of HBM2 interface, it could feature 16 GB of memory. It could launch some time in early-2017. AMD is claiming a massive 2.5X performance-per-Watt increase for "Polaris" over the current GCN 1.2 architecture, which drives the "Tonga" and "Fiji" chips, and so these stream processor counts could look deceptively insufficient.
24 Comments on Upcoming AMD "Polaris" and "Vega" GPU Compute Unit Counts Surface
Polaris 10, the faster guy, should be demoed by the end of May.
What I am puzzled about is that there isn't that much of a perf gap between 390x and Fury, to squeeze one more item in, so. dafuq would the rest of the line be doing, if there is 490 & 490x that is much faster than 390&390x?Only Vega 10 is 2017 (HBM2).
Polaris 10 is expected in June this year.Even 3xx wasn't rebadging (different TDP and performance) let alone Fiji with HBM.
I'll go out on a limb and say the 2.5x perf/watt increase is about as realistic as a Fury's tremendous overclocking capabilities. Both of which were seen on AMD slides.
:toast:
He was supposedly got "caught up in the moment" and said something that you quickly you see Raja Koduri trying to get him off that subject. That said you'd think your Chief Technology Officer would've known the actual product and not go shooting the company in the foot. A day or so later AMD release information, saying more or less toning it back to a more conservative "upward of 7-10%", but by that point the damage was done.
He'll never live that down... Being mostly know most as a Chairman of JEDEC's Subcommittee I don't know what he brings to that GPU discussion other than about HBM and now it shows. While get a haircut, and sit up straight... geez!
140W doesn't make sense for GTX 950 (a 90W TDP card), unless they are talking about full system power... and that doesn't make sense either (too low). Not that they couldn't find an inefficient 950 and *make* it consume 140W, which is what I'm assuming they did. In that case we already have GTX 950s that run on the slot only (<75W), so Polaris probably won't even beat current Maxwell Nvidia in FPS/W.
Bearing that in mind, and the fact that the 285 was far from a performance-per-watt leader (and the binned Fiji Nano already close to doubled the efficiency), 2.5X almost seems conservative.
www.hardware.fr/articles/937-26/overclocking-gpu-fiji.html
From a like a month ago;
www.play3r.net/reviews/graphics/xfx-amd-r9-fury-x-4gb-graphics-card-review/14/
The regular Sapphire Nitro Fury went from the 1050Mhz to 1139Mhz (8.5%).
www.hardwareasylum.com/reviews/video/sapphire_r9fury-nitro/page13.aspx
One thing that many who first started the OC do find is that it matters on radiator hose position, and the official installation guidelines should be followed regargeding the radiator hoses (should come out from the bottom).
To finalize, Joe Macri start in about cooler and the amount of heat it could dissipate, but end-up making it more about the entire card... doh! I say AMD should've done more to back-away from his comments. The internet took a vague statement(s), and it took on a life quickly. If AMD knew the first release units had some amount of being locked-out and/or production shortcomings they should've addressed that more. They where derelict in not seeing the drum beating before the reviewers' came off their NDA, and should've nip it in the butt.
The performance per watt change, at both camps, is going to come predominantly from switching to 14/16nm fabs. I don't expect much change from the architecture because those improvements were made on 28nm.