Upcoming AMD "Polaris" and "Vega" GPU Compute Unit Counts Surface

[btarunr]

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.

View at TechPowerUp Main Site

 

[Fabio Bologna]

Are we gonna see the first big GPU release from AMD, in an awfully long time, without a massive rebadging action?

 

[ZoneDymo]

early 2017....sigh

 

[the54thvoid]

I don't think either camp is releasing flagships that will impress until then.

 

[Fluffmeister]

So not very exciting then?

 

[medi01]

I guess the smaller guy, Polaris 11 was what was demoed vs 960:

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?

early 2017

Only Vega 10 is 2017 (HBM2).
Polaris 10 is expected in June this year.

rebadging

Even 3xx wasn't rebadging (different TDP and performance) let alone Fiji with HBM.

 

[Chaitanya]

Waiting for a good 1440p card that costs less than 300$.

 

[okidna]

I guess the smaller guy, Polaris 11 was what was demoed vs 960:

GTX 950, not 960.

 

[Frick]

So not very exciting then?

If the power effeciency thing iss true, I'd say it's interesting. Here's hoping for good overclocking capabilites!

 

[Stephen.]

GTX 950, not 960.

View attachment 73662

I only have one question about this pic, how are they using i7-4790K with DDR4? last time i checked 4th gen and below Intel processors do not have a DDR4 memory controller in them, though i know for a fact i7-4790K can easily push 2600MHz DDR3

 

[Vayra86]

I only have one question about this pic, how are they using i7-4790K with DDR4? last time i checked 4th gen and below Intel processors do not have a DDR4 memory controller in them, though i know for a fact i7-4790K can easily push 2600MHz DDR3

AMD slides. Need we say more?

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.

 

[medi01]

Fury's tremendous overclocking capabilities. Both of which were seen on AMD slides.

Care to post a link?

 

[Vayra86]

Care to post a link?

Did try to find it but Google won't show me the slide It was part of the Fury presentation anyway, shortly before launch.

 

[yogurt_21]

lots of words in there that added up to "skip" or is that just how I'm interpreting it?

 

[the54thvoid]

Care to post a link?

He doesn't have to, it was spoken at the release of Fiji that Fury X was an overclockers dream. If you go find the presentation and trawl through it you'll find it. It's not a rumour - it was said.

 

[Divide Overflow]

There's an awful lot of "could" statements in this article.

 

[okidna]

I only have one question about this pic, how are they using i7-4790K with DDR4? last time i checked 4th gen and below Intel processors do not have a DDR4 memory controller in them, though i know for a fact i7-4790K can easily push 2600MHz DDR3

That's my question aswell. Probably a typo from their part.

 

[arbiter]

Even 3xx wasn't rebadging (different TDP and performance) let alone Fiji with HBM.

.... um 3xx was a rebadge, just cause tdp and performance is different doesn't uncheck that box. If its same chip used before its a rebadge.

 

[Steevo]

Where is the pudding?

 

[Casecutter]

He doesn't have to, it was spoken at the release of Fiji that Fury X was an overclockers dream. If you go find the presentation and trawl through it you'll find it. It's not a rumour - it was said.

To see that part of the video forward to about 1:01:20

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!

 

[Vayra86]

To see that part of the video forward to about 1:01:20

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!

7-10%, while in game performance rises by 4-5% on average. Mkay

 

[rruff]

That's my question aswell. Probably a typo from their part.

Most likely the whole thing is a "typo"...

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.

 

[HumanSmoke]

AMD slides. Need we say more?

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.

Depending on what their baseline is, 2.5X isn't too fantastical. AMD are obviously using Tonga as their start point since the "28nm GPU" point is late 2014 and the 285/M295X are the only card releases close to that point

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.

 

[Casecutter]

7-10%, while in game performance rises by 4-5% on average. Mkay

I'd have you explore the more recent discussions on various forums, many are indicating between 1120-1150Mhz being the range they can achieve by just following some of the hints as to adjusting voltage and setting power limit. The less understood is, even if you have a 7-10% OC is the "traction" that provides at 4K (its' the only rational res, other than perhaps 144Hz Freesync monitors) doesn't always come thru in every title.
http://www.hardware.fr/articles/937-26/overclocking-gpu-fiji.html
From a like a month ago;
http://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%).
http://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.

 

[FordGT90Concept]

The Fury Nano/GTX 950 demo isn't very relevant going forward. They were both frame locked to 60 fps and Fiji is a monster of a chip running at low clockspeeds getting the performance/watt gains compared to GTX 950 which was running close to the peak of its capability. It's like comparing a Class A truck pulling 30,000 lbs (less than half of its intended load) versus a 1-ton truck pulling 30,000 lbs (almost at maximum). Fuel economy in the Class A truck would fall maybe 20% where the 1-ton truck would fall 50+%.

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.