AMD VEGA confirmed for 2017 H1

TNA · 12 Dec 2016 at 16:14

AthlonXP1800 said:
http://videocardz.com/64706/amd-vega-doom-4kultra-gaming-performance-demo-possible-specs

Vega Doom performance around GTX 1080 performance level.

Device ID 687F:C1 confirmed as VEGA 8GB HBM2 matched spec from android app GPU-L I posted days ago.

I will be disappointed if that is their best card after this long wait. If that is the case I will definitely be waiting for Volta. Hopefully that is not their best card and a third tier card or something.

LoadsaMoney · 12 Dec 2016 at 16:28

AthlonXP1800 said:
http://videocardz.com/64706/amd-vega-doom-4kultra-gaming-performance-demo-possible-specs

Vega Doom performance around GTX 1080 performance level (when using Vulkan).

Device ID 687F:C1 confirmed as VEGA 8GB HBM2 matched spec from android app GPU-L I posted days ago.

Rroff · 12 Dec 2016 at 16:32

TNA said:
I will be disappointed if that is their best card after this long wait. If that is the case I will definitely be waiting for Volta. Hopefully that is not their best card and a third tier card or something.

Depends what the position in the lineup that card is though - doesn't mean they don't have a bigger card.

Schnippzle · 12 Dec 2016 at 16:35

TNA said:
I will be disappointed if that is their best card after this long wait. If that is the case I will definitely be waiting for Volta. Hopefully that is not their best card and a third tier card or something.

I guess it all depends price, really. If it comes out at ~£300, why not get two?

CAT-THE-FIFTH · 12 Dec 2016 at 16:37

Rroff said:
Depends what the position in the lineup that card is though - doesn't mean they don't have a bigger card.

It depends on what GPU they are using - if its "small" Vega at under 400MM2 with a narrower HBM2 interface its not too bad,and Hynix HBM2 production is really only starting to ramp up recently,and it is what I feared if they stuck with HBM2 for the smaller card,ie, HBM2 production schedules will be the limiting factor I suspect.

If it is a 400MM2 to 500MM2 chip then its somewhat underwhelming.

Davedree · 12 Dec 2016 at 16:38

a lot of effort and expense for just a desktop chip just to compete with a laptop/desktop chip (gp104).
Let's hope we see a 4608/ 5120 shader jobby quite soon.

CAT-THE-FIFTH · 12 Dec 2016 at 16:42

Davedree said:
a lot of effort and expense for just a desktop chip just to compete with a laptop/desktop chip (gp104).
Let's hope we see a 4608/ 5120 shader jobby quite soon.

Well if its the smaller Vega with 4096 shaders its not going to massively bigger I suspect than the GP104. It will be interesting to see what TDP it has - its using only a 2048 bit memory controller,which is half that of the Fiji cards.

Regarding the development of the chip,I have a strong feeling the XBox Scorpio might be based on it. Its a bit like a few other AMD GPUs,you have seen variants of them in consoles,so it is almost like they are using that to spread the dev costs of some lines.

Davedree · 12 Dec 2016 at 16:53

CAT-THE-FIFTH said:
Well if its the smaller Vega with 4096 shaders its not going to massively bigger I suspect than the GP104. It will be interesting to see what TDP it has - its using only a 2048 bit memory controller,which is half that of the Fiji cards.

Regarding the development of the chip,I have a strong feeling the XBox Scorpio might be based on it. Its a bit like a few other AMD GPUs,you have seen variants of them in consoles,so it is almost like they are using that to spread the dev costs of some lines.

Totally agree with you, i'm predicting about 360-380mm2 if using hbm2, most of the die will be for the l2 cache. If you look at the polaris 4 geometry engines then adding 1 more cu per engine would give you 40 cu's, if you scale up by 2x, like tonga to fiji, then a 5120 die isn't going to be that much bigger than Vega 4096. In this sense providing they get the scheduling right Hbm2 will be more beneficial the more the cus are scaled up.

The costs of hbm2 and power efficency/performance are more of a concern for amd in mobilesector than desjtop.

Seanspeed · 12 Dec 2016 at 17:04

humbug said:
Yeah, there are differences between them ^^^^^ the benchmarks for the 290 are of reference cooled ones running 100Mhz lower than reference clocks, about 850Mhz, the 390 with a proper cooler runs at 1000Mhz.

And a 970 can be overclocked much more than a 390. Much less a 290.

Dude, at the very worst, you can say they're very close. And for a card that was higher end when it first came out, compared to a mid range part just a year later, they are obviously both great purchases for anybody who got one on Day 1.

What's with some of you finding it so hard to admit this? It's pretty sad. It's like you just *refuse* to give Nvidia credit where it's due. The 970 was(and is still) a great card. The 290 also being a great card(in open air cooler form) doesn't negate this at all.

Rroff · 12 Dec 2016 at 17:11

Seanspeed said:
And a 970 can be overclocked much more than a 390. Much less a 290.

Dude, at the very worst, you can say they're very close. And for a card that was higher end when it first came out, compared to a mid range part just a year later, they are obviously both great purchases for anybody who got one on Day 1.

What's with some of you finding it so hard to admit this? It's pretty sad. It's like you just *refuse* to give Nvidia credit where it's due. The 970 was(and is still) a great card. The 290 also being a great card(in open air cooler form) doesn't negate this at all.

I played through a good bit of Doom on a 970 - 1440p everything turned up was most of the time comfortably well over 60fps.

Was a bit of a mixed story though - some older games where performance came down a lot to shovelling lots of primitive geometry, etc. my 780 could be quite a bit faster.

They are all old cards now though despite the justifications of some - if it wasn't for the fact nVidia have trickled out performance increments (and AMD were nowhere to be seen) they'd have been utterly obsolete months ago as developers targetted a new generation of hardware.

humbug · 12 Dec 2016 at 17:17

It dips into the high 40's low 50's on Ultra at 1080P a fair bit too tho Roff

^^^

Seanspeed said:
And a 970 can be overclocked much more than a 390. Much less a 290.

Dude, at the very worst, you can say they're very close. And for a card that was higher end when it first came out, compared to a mid range part just a year later, they are obviously both great purchases for anybody who got one on Day 1.

What's with some of you finding it so hard to admit this? It's pretty sad. It's like you just *refuse* to give Nvidia credit where it's due. The 970 was(and is still) a great card. The 290 also being a great card(in open air cooler form) doesn't negate this at all.

Your complaint there is unwarranted.

Plenty of credit is given to the 970, i have said repeatedly they are both great cards and as far as i can see no one is denying that.

easyrider · 12 Dec 2016 at 17:19

Gregster said:

Depends on cost.

TNA · 12 Dec 2016 at 17:38

Rroff said:
Depends what the position in the lineup that card is though - doesn't mean they don't have a bigger card.

I never said it did. I said I hope that is not their best card. My guess would be it would be their third tier card, aka the top Vega 11. I am expecting Vega 10 to be a lot better

Schnippzle said:
I guess it all depends price, really. If it comes out at ~£300, why not get two?

Multi GPU? Not a good idea currently unfortunately.

CAT-THE-FIFTH said:
If it is a 400MM2 to 500MM2 chip then its somewhat underwhelming.

Exactly, but hopefully not

nashathedog · 12 Dec 2016 at 17:38

I like that it's only an 8gb card as that will help keep pricing down,
Performance is around what was expected (by the down to earthers). However,
The fact that they are showing Vulcan only does concern me as I wonder where it'll stand in DX11 titles compared to a 1080.
For me an 8gb Vega with 1080 performance in DX11 would be bang on as it should be priced okay but if it's slower than a 1080 and it's the big Vega it's a shame, Unless it's priced well..

Rroff · 12 Dec 2016 at 17:40

humbug said:
It dips into the high 40's low 50's on Ultra at 1080P a fair bit too tho Roff ^^^

Hmm I only saw one place where it dropped into the 50s - a lot of the time was 70+ FPS though I only played about 1/3rd of the game on the 970.

IIRC that was ultra settings minus one of the motion blur or DOF or whatever settings and using TSSAA I believe.

EDIT: Doom performance is a bit odd mind - at 1080p my 780 was faster for the most part but at 1440p the 780 just fell apart for some reason regardless of quality settings and VRAM utilisation, etc. while the 970 hardly dropped in performance.

tyler_jrb · 12 Dec 2016 at 17:41

easyrider said:
Depends on cost.

Depends if this is the main chip and what they are showing is true.

I.e. Ultra IQ = every setting maxed rather than the main ultra settings only.

Either way looks promising. Just hope it's somewhat cheap.

africanos23 · 12 Dec 2016 at 18:04

Not sure if anyone posted this

Edit: Someone already has

Wonder why they didn't show any temps

tommybhoy · 12 Dec 2016 at 18:08

Gregster said:
If I had to choose between the 2 and no sync screen involved, I would choose the 290X. Grief or not, they are both great cards but the 290X would edge it for me.

It's easy saying that now they are yesterdays cards

but fair play all the same.

I'm still not saying as Iv'e been here too long and it would be too much an

now.

queamin · 12 Dec 2016 at 18:10

tommybhoy said:
It's easy saying that now they are yesterdays cards but fair play all the same.

I'm still not saying as Iv'e been here too long and it would be too much an now.

But we want to know so someone can shout at you.

bru · 12 Dec 2016 at 18:14

Well there is a wealth of info guesses after the NDA on that event lifted, so lets see what we've got.

8GB ram 'confirmed'
512GB/s memory bandwidth 'either 4 stacks at 500MHz or 2 at 1000MHz'
12.5 TFlop/s FP32
< 300w TDP in server style form (fanless)

Can you get 2G HBM2 stacks? I thought it was either 4G or 8G, which would mean 2 stacks at 1000MHz tying in with above.

NCU ? Is this the GCN replacement ? 'next compute unit'

The opposition for reference

Tesla P40 16GB GDDR5x ram (GP102)
345GB/s
12TFlops FP32
250w TDP in server style form (fanless)

So the first info on Vega is certainly looking good, AMD could have a real winner on their hands.

The only downside I can see is the power usage, which is 20% more than the P40, a little bit higher than I expected considering it is using HBM2 as well.

Overall good first look and info, hopefully we will get a bit more tomorrow with the ZEN event.