AMD VEGA confirmed for 2017 H1

humbug · 20 Sep 2016 at 13:12

24 TFlops half-precision = 12 TFlops single-precision, GTX 1080 is 8.9 TFlops single-precision

Final8y · 20 Sep 2016 at 13:14

KillBoY_UK said:
AMD fans where not saying that with the 5970's when the GTX 480 got released, they where quite happy (and rightly so) to rip nvidia a new on on the heat and power of that thing.

but then the tables turned with the 290/390 its not an issue anymore.

Power and heat with the 290/390 has always been an issue, don't make out that that people said otherwise, the 290 got slammed everywhere the 390 came custom cooled from day one and seeing as most know that the 390 is a tweaked 290 moaning about the power and heat had already run its course with the 290.

The 5970 was a dual GPU card that used less power than the 480.

humbug · 20 Sep 2016 at 13:21

Final8y said:
Power and heat with the 290/390 has always been an issue, don't make out that that people said otherwise, the 290 got slammed everywhere the 390 came custom cooled from day one and seeing as most know that the 390 is a tweaked 290 moaning about the power and heat had already run its course with the 290.

The 5970 was a dual GPU card that use less power than the 480.

It was never an issue for me, i had a 290 for 2 years, it ran cool and quiet.

Like everything else AMD it was vastly over blown by Nvidia fanbois.

Final8y · 20 Sep 2016 at 13:23

humbug said:
It was never an issue for me, i had a 290 for 2 years, it ran cool and quiet.

Like everything else AMD it was vastly over blown by Nvidia fanbois.

The stock cooler on the 290 was a disaster and yes the rest a bit overblown.

Steviepunk · 20 Sep 2016 at 13:27

r7slayer said:
AMD Vega 20 will use Glo Flo 7nm? I highly doubt that as they said they only just decided to skip 10nm and go to 7. Are Global foundries even producing 7nm wafers yet?

Maybe true for Navi but doubtful for vega.

Isn't Vega20 new? I thought the plans were just for Vega10 and Vega11, then Navi10 & 11. sounds like issues with Navi being delayed to 2019, Vega20 may be a die shrink to fill the gap in 2018 between Vega10 and Navi10. Though as you say, 7mm seems ambitious, can't remember if 7mm was scheduled to be available in 2017, 2018 or 2019.

That all does assume that the rumour is true of course

N19h7m4r3 · 20 Sep 2016 at 13:28

r7slayer said:
AMD Vega 20 will use Glo Flo 7nm? I highly doubt that as they said they only just decided to skip 10nm and go to 7. Are Global foundries even producing 7nm wafers yet?

Maybe true for Navi but doubtful for vega.

It's not normal 7nm, probably similar to FDX which is bigger than FinFet but far better.

PCPer had a good discussion about it.

https://youtu.be/2934h0hDDsE?t=1h13m41s

Also Vega 20 is probably a year or more after Vega 10 and 11, which is only coming second half of 2017.

humbug · 20 Sep 2016 at 13:28

Final8y said:
The stock cooler on the 290 was a disaster and yes the rest a bit overblown.

Right, thats all it was, a crap stock cooler, any normal AIB cooler on it and it was perfectly fine.

bru · 20 Sep 2016 at 13:35

Interesting rumour, the first thing to remember is that this is supposedly from the server roadmap, so the numbers are for the server variants not the desktop cards.

what does this tell us, well it would be.

4096 stream processors (effectively two rx470's)
24TF FP16 half precision (hmm not too sure about this seeing as the 470 only does 3.7GF full precision FP16 so about 7.2 half precision FP16)
16GB HBM2 (quite possible on a server card)
512 Gbps memory bandwidth ( seems quite low)
225W TDP

By comparison

NVidia P100 (equivalent HBM2 card)

3584 Cuda cores (not a full chip, so maybe the AMD Vega 10 isn't either seeing as it is only 2x470)
21.2TF FP16 half precision (or 9.3GF full precision FP16 so about 18.6 half precision FP16 for the non mezzanine card)
16GB HBM2
720 Gbps memory bandwidth
300W TDP or 250W for the non mezzanine card

As for a desktop card, performance wise it would be about two 470's. Make of that lot what you will.

Disclaimer: all info taken from various sources around the web and I now have a massive headache, so it may all be complete baloney.

SiDeards73 · 20 Sep 2016 at 13:44

Global Foundries reportedly said recently they was looking to move to 7nm, infact theres been a few articles recently regarding it

http://www.extremetech.com/computin...unces-new-7nm-finfet-process-full-node-shrink

470 is what? slightly slower than a 480? Dual 480 is a bit better than a 1080 where supported and in AMD favoured titles? so you could say that the small vega is roughly around 1080 performance? this is all ifs and buts obviously...

a lot comes down to price, if Small Vega is actually in reality trading blows with 1080 it needs to be cheaper, but i cant see it as i am sure HBM2 will come at a premium.

To be honest its best to just wait til some reviews next year, after the fiasco of the Polaris hype which followed the Fiji hype i am not interested in reading random clickbait sites for speculation, now with regards to AMD i need to see official figures.

Worst part is, im stuck with AMD anyhow as i bought a Freesync screen, and Vega will be better than my 290 so it will atleast be an upgrade for me, no way im ever giving Nvidia my cash thats for sure.

TheRealDeal · 20 Sep 2016 at 13:48

bru said:
Interesting rumour, the first thing to remember is that this is supposedly from the server roadmap, so the numbers are for the server variants not the desktop cards.

what does this tell us, well it would be.

4096 stream processors (effectively two rx470's)
24TF FP16 half precision (hmm not too sure about this seeing as the 470 only does 3.7GF full precision FP16 so about 7.2 half precision FP16)
16GB HBM2 (quite possible on a server card)
512 Gbps memory bandwidth ( seems quite low)
225W TDP

By comparison

NVidia P100 (equivalent HBM2 card)

3584 Cuda cores (not a full chip, so maybe the AMD Vega 10 isn't either seeing as it is only 2x470)
21.2TF FP16 half precision (or 9.3GF full precision FP16 so about 18.6 half precision FP16 for the non mezzanine card)
16GB HBM2
720 Gbps memory bandwidth
300W TDP or 250W for the non mezzanine card

As for a desktop card, performance wise it would be about two 470's. Make of that lot what you will.

Disclaimer: all info taken from various sources around the web and I now have a massive headache, so it may all be complete baloney.

You are assuming Vega is Polaris which it appears not to be. If it's the server card then it's probably clocked a good bit lower so possibly higher clocks and tdp on a gaming card with higher bandwidth.

bru · 20 Sep 2016 at 13:57

SiDeards73 said:
470 is what? slightly slower than a 480? Dual 480 is a bit better than a 1080 where supported and in AMD favoured titles? so you could say that the small vega is roughly around 1080 performance? this is all ifs and buts obviously...

a lot comes down to price, if Small Vega is actually in reality trading blows with 1080 it needs to be cheaper, but i cant see it as i am sure HBM2 will come at a premium.

Hmm isn't Vega 10 the bigger of the two Vega chips?

bru · 20 Sep 2016 at 13:59

TheRealDeal said:
You are assuming Vega is Polaris which it appears not to be. If it's the server card then it's probably clocked a good bit lower so possibly higher clocks and tdp on a gaming card with higher bandwidth.

Yup, could be.

Also there a lot of assumptions, guesses and estimations, in this sort of rumour.

SiDeards73 · 20 Sep 2016 at 14:02

bru said:
Hmm isn't Vega 10 the bigger of the two Vega chips?

I thought Vega 10 was the smaller and Vega 20 the bigger?

humbug · 20 Sep 2016 at 14:05

The RX 470 has 5 TFlops single-precision, so 10 for two of them.

Which mean's at 12 TFlops single-precision Vega is clocked 20% higher than the RX 470. 1450Mhz.

TheRealDeal · 20 Sep 2016 at 14:15

bru said:
Yup, could be.

Also there a lot of assumptions, guesses and estimations, in this sort of rumour.

Yea without knowing the Vega architecture it's very hard to guess at performance. Vega could be what Maxwell was to Kepler or it could just be another step up like Gcn 1.1 to 1.2.

bru · 20 Sep 2016 at 14:22

humbug said:
The RX 470 has 5 TFlops single-precision, so 10 for two of them.

Which mean's at 12 TFlops single-precision Vega is clocked 20% higher than the RX 470. 1450Mhz.

Hmm yes your right, looking at the Techpowerup database it lists 4.94 Floating point where this AMD GPU wiki only list 3.793 single precision.
We would have to assume the Techpowerup database is more accurate, seeing as it is more in line with what AMD themselves list.

TheRealDeal · 20 Sep 2016 at 14:25

bru said:
Hmm yes your right, looking at the Techpowerup database it lists 4.94 Floating point where this AMD GPU wiki only list 3.793 single precision.
We would have to assume the Techpowerup database is more accurate, seeing as it is more in line with what AMD themselves list.

Yea the Wiki has it all wrong with Texture/pixel fill rate being way lower than they should be. I think the wiki is taking it's info from the stock clock instead of the boost which is the same with the Rx480.

humbug · 20 Sep 2016 at 14:28

Yeah, the RX 480 is listed with 5.5 TFlops.

RX 470 = 2048 Shaders.
RX 480 = 2304 Shaders. (+12.5%)

5 TFlops + 12.5% = 5.625 TFlops. -62Mhz = about 5.5 TFlops.

orbitalwalsh · 20 Sep 2016 at 14:28

bit on console topic but got me wondering if Xbox Scorpio will use a heavily cut down Vega 10 (ones that didn't cut the mustard) with GDDR5. Would be great business for AMD and i find it hard to stomach that it could be an overclocked RX 480 with the thermals it produces just wouldn't sit right in a console without throttling.
Interesting thought

and good business practice to sell chips that aren't 100% just

Greebo · 20 Sep 2016 at 14:36

Final8y said:
Steam has not surveyed me for 3 years, so my 3x290s have never been registered and at this rate they never will as i would have moved on from them before i'm asked again.

Steam still has me running a 8800GT