AMD VEGA confirmed for 2017 H1

[r7slayer]

O

So let's sum up here. Guru now thinks that's true.

That means this "Vega" card is 6% faster than the stock Fury X according to their original post.
This is when clocked 1200mhz vs 1050mhz.

So that means Vega has regressed in IPC.

We know the Instinct MI25 which is Vega has 12.5TFlops. That means it has a clock speed of 1500Mhz, and it's a passive air cooled accelerator just like Tesla.

We know HBM2 is rated for up to 1000Mhz, that's for 1.6Gbps and 2.0Gbps modules, which are the ones AMD will be using.

NVIDIA are using 1.4Gbps modules, which are clocked at 715Mhz.

Yet "Vega" in that Time spy link shows memory at 700Mhz, which is slower than the slowest available HBM2 modules.

The Vega die is almost the same since as Fiji, while the former is on 14nm compared to 28nm.
So Vega is a much larger chip.


So apparently, gaming Vega, which is going to be Water Cooled as a premium part, is going to have under 10Tflops, and be significantly slower clocked compared to an air cooled server part.
While having somehow lower clocked HBM2 when using the higher bandwidth and higher clocked modules.

All this, and then despite Vega being a larger GPU, and all the architectural improvements means Vega has severely regressed in IPC and overall performance compared to two year old Fiji.

That doesn't add up for me.

Sometimes it's good to throw logic in the mix.
It's so tiresome reading some of the comments that truly believe that Vega is just a 1070/80 competitor and vega is not for gaming. I just don't respond because it's either a troll comment or they really are naive. A furyx die shrink i would expect to be around 1070/80 levels. I mean it's pretty much around 1070 as it is. So vega with all the optimisations and improvements to the architecture, more so than even before since GCN was introduced tells me that performance wise has gotta have better IPC than a Fury lineup.
Also i don't know how Vega is not for gamers when they are releasing VEGA GAMING CARDS and marketing Vega towards gaming.... Ohh super cars are not for the general public or public roads only for racers and racetracks lol.

 

[nashathedog]

We know the Instinct MI25 which is Vega has 12.5TFlops. That means it has a clock speed of 1500Mhz, and it's a passive air cooled accelerator just like Tesla.




That doesn't add up for me.

Nor me but is 1500mhz the confirmed clockspeed of MI25?
Or is it an assumption working with armchair arithmetic?

Have MI25 speeds been announced?

 

[Gregster]

So let's sum up here. Guru now thinks that's true.

That means this "Vega" card is 6% faster than the stock Fury X according to their original post.
This is when clocked 1200mhz vs 1050mhz.

So that means Vega has regressed in IPC.

We know the Instinct MI25 which is Vega has 12.5TFlops. That means it has a clock speed of 1500Mhz, and it's a passive air cooled accelerator just like Tesla.

We know HBM2 is rated for up to 1000Mhz, that's for 1.6Gbps and 2.0Gbps modules, which are the ones AMD will be using.

NVIDIA are using 1.4Gbps modules, which are clocked at 715Mhz.

Yet "Vega" in that Time spy link shows memory at 700Mhz, which is slower than the slowest available HBM2 modules.

The Vega die is almost the same since as Fiji, while the former is on 14nm compared to 28nm.
So Vega is a much larger chip.


So apparently, gaming Vega, which is going to be Water Cooled as a premium part, is going to have under 10Tflops, and be significantly slower clocked compared to an air cooled server part.
While having somehow lower clocked HBM2 when using the higher bandwidth and higher clocked modules.

All this, and then despite Vega being a larger GPU, and all the architectural improvements means Vega has severely regressed in IPC and overall performance compared to two year old Fiji.

That doesn't add up for me.

Good logic but why can't it be an engineering sample that is being tested? Like I said in my response to Alamox, it could well be clocked much lower than what the retail card will be released at, it might even be small Vega, TimeSpy could also be reporting the wrong clocks, as it has done on numerous other cards. My assumption is it is Vega but not a fully operational Vega to speak of and we will see the fully enabled Vega kick some 1080 butt at clocks of 1500Mhz +

 

[opethdisciple]

Hopefully in the next 4 weeks we should get some credible information.

 

[muon]

Why are people so obsessed with dismissing this. 'It can't be true, Vega will have magical improvements'.

For all the marketing, even Pascal was pretty much just Maxwell overclocked. The IPC improvement from Fiji to Polaris 10/20 was minimal if any (r9 390 went toe to toe with rx 480). Vega is still on the same process as Polaris.

People are setting themselves up for massive disappointment.

Who knows, maybe the top end Vega will havr more than 4096 shaders.

 

[akarypid]

Why are people so obsessed with dismissing this. 'It can't be true, Vega will have magical improvements'.

For all the marketing, even Pascal was pretty much just Maxwell overclocked. The IPC improvement from Fiji to Polaris 10/20 was minimal if any.

...and the same for Polaris, only with the addition of the primitive discard accelerator.

Vega is not that. It has a better geometry engine, a longer pipeline for higher clocks, it seems to be the biggest departure from the traditional GCN architecture we've had for many years.

I don't know if it's a true result of a down-clocked engineering sample with cache disabled, or just a plain fake result, but what I do believe is N19h7m4r3 wrote: saying this is full fat Vega at its best simply does not add up.

 

[N19h7m4r3]

Good logic but why can't it be an engineering sample that is being tested? Like I said in my response to Alamox, it could well be clocked much lower than what the retail card will be released at, it might even be small Vega, TimeSpy could also be reporting the wrong clocks, as it has done on numerous other cards. My assumption is it is Vega but not a fully operational Vega to speak of and we will see the fully enabled Vega kick some 1080 butt at clocks of 1500Mhz +

Fury X at 1450Mhz is close to 2Ghz GTX 1080 in Firestrike.

Even if it is an engineering sample it means that Vega has worse IPC than Fiji looking at performance to clocks alone.

Only way it makes remotely sense is if it's cut down small Vega, as the TFLOPS of that card is 9.8.

Nor me but is 1500mhz the confirmed clockspeed of MI25?
Or is it an assumption working with armchair arithmetic?

Have MI25 speeds been announced?

MI25 has 25TFLOPS of FP16, and runs that at 2:1 of FP32. That means it has 12.5TFLOPS of FP32.
http://www.anandtech.com/show/10905/amd-announces-radeon-instinct-deep-learning-2017/2

Big Vega has 4096 Stream processors; which at 12.5TFLOPs = 1525Mhz core clock.

TFLOPS = Shaders * 2 * Clock Frequency

Why are people so obsessed with dismissing this. 'It can't be true, Vega will have magical improvements'.

For all the marketing, even Pascal was pretty much just Maxwell overclocked. The IPC improvement from Fiji to Polaris 10/20 was minimal if any.

Rather simple really, a Fury X at 1050Mhz is 6% slower than the "Vega" at 1200Mhz. That implies reduced IPC, while being a substantially larger GPU.

Fiji was 596mm2 on 28nm, Vega is over ~530mm2 on 14nm. Yet somehow performs worse clock for clock while being a much bigger design?

If that's true, Vega is the Bulldozer of GPUs. High clocks, bigger chip, but worse IPC.

 

[muon]

In which case the top end Vega must have more than 4096 shaders to be so big. So what makes this result fake?

As for IPC, I sort of get where you are coming from. But no one is saying this is the final performance.

 

[nashathedog]

If that's true, Vega is the Bulldozer of GPUs. High clocks, bigger chip, but worse IPC.

Let's hope not. After waiting so long that'd destroy my belief in RTG. Good drivers alone aren't enough.

 

[akarypid]

Rather simple really, a Fury X at 1050Mhz is 6% slower than the "Vega" at 1200Mhz. That implies reduced IPC, while being a substantially larger GPU.

Fiji was 596mm2 on 28nm, Vega is over ~530mm2 on 14nm. Yet somehow performs worse clock for clock while being a much bigger design?

If that's true, Vega is the Bulldozer of GPUs. High clocks, bigger chip, but worse IPC.

That's a bit harsh. The IPC in itself doesn't matter. What matters is IPC*CLOCK. It's ok to sacrifice some IPC if you can get much higher CLOCK.

Let's say a node-shrink of Fiji could get to 1400MHz on 14nm. If Vega is 6% slower in IPC but can get to 1600MHz which is 14% higher clock, then it's totally worth it and not what I would call a Bulldozer.

 

[N19h7m4r3]

In which case the top end Vega must have more than 4096 shaders to be so big. So what makes this result fake?

As for IPC, I sort of get where you are coming from. But no one is saying this is the final performance.

Unless AMD for the first time has designed two drastically different nodes, similar to NVIDIA for Tesla vs Quadro/GeForce P100 vs GP102, I don't see that happening.

P100 has FP64, and FP16
GP102 only has FP32

The 4096 Shader die found in the AMD Instinct MI25 is the largest Vega we know of. As AMD used the same highend design for both Gaming and Enterprise.

7970 vs FirePro W9000
290X vs FirePro W9100
Fury vs Instinct Mi18
RX 480 vs Instinct MI6

All have FP 64 (although at 1/16 since Tahiti), 32, and 16

AMD hasn't had the resources for the R&D required to do what NVIDIA has been doing, with two massive GPUs, one aimed at the top end of Enterprise, and another for Gaming.

Is it possible AMD has a core out there with more Stream Processors? It's certainly possible, but given their severe constraints the last few years it's not very likely.

Vega for consumers is also due within the next two months. So unless that "leak" is a Vega sample from over 1-2 years ago, I don't see what they could do to suddenly improve IPC dramatically.

 

[N19h7m4r3]

That's a bit harsh. The IPC in itself doesn't matter. What matters is IPC*CLOCK. It's ok to sacrifice some IPC if you can get much higher CLOCK.

Let's say a node-shrink of Fiji could get to 1400MHz on 14nm. If Vega is 6% slower in IPC but can get to 1600MHz which is 14% higher clock, then it's totally worth it and not what I would call a Bulldozer.

IPC does matter though, especially when it's been reported that Vega will have improved IPC. It's why I'm highly skeptical of that "leak".
http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

Finally, along with outlining their new packed math formats, AMD is also confirming, at a high level, that the Vega NCU is optimized for both higher clockspeeds and a higher IPC. It goes without saying that both of these are very important to overall GPU performance, and it’s an area where, very broadly speaking, AMD hasn’t compared to NVIDIA too favorably.

 

[akarypid]

That's a bit harsh. The IPC in itself doesn't matter. What matters is IPC*CLOCK. It's ok to sacrifice some IPC if you can get much higher CLOCK.

Let's say a node-shrink of Fiji could get to 1400MHz on 14nm. If Vega is 6% slower in IPC but can get to 1600MHz which is 14% higher clock, then it's totally worth it and not what I would call a Bulldozer.

IPC does matter though, especially when it's been reported that Vega will have improved IPC. It's why I'm highly skeptical of that "leak".
http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

I didn't say IPC doesn't matter. All I said is that it's just 1 part of the equation and CLOCK is equally important (we don't have IPC squared times clock, they both contribute equally). So knowing how much the IPC changed means nothing if we don't know what effect it had on CLOCK.

To put it in the extreme, imagine if this Vega with 6% less IPC turns out to have a boost clock of 1900Mhz: nobody would be complaining about a 6% drop in IPC then, would they?

I'm sure AMD have done their math and decided it's worth it.

 

[N19h7m4r3]

I didn't say IPC doesn't matter. All I said is that it's just 1 part of the equation and CLOCK is equally important (we don't have IPC squared times clock, they both contribute equally). So knowing how much the IPC changed means nothing if we don't know what effect it had on CLOCK.

To put it in the extreme, imagine if this Vega with 6% less IPC turns out to have a boost clock of 1900Mhz: nobody would be complaining about a 6% drop in IPC then, would they?

I'm sure AMD have done their math and decided it's worth it.

Which conflicts with that I just linked you. Vega is reported to have Improved IPC, not less.

http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

So that "leak" still doesn't add up at all.

 

[eddyr]

Why are people so obsessed with dismissing this. 'It can't be true, Vega will have magical improvements'.

For all the marketing, even Pascal was pretty much just Maxwell overclocked. The IPC improvement from Fiji to Polaris 10/20 was minimal if any (r9 390 went toe to toe with rx 480). Vega is still on the same process as Polaris.

People are setting themselves up for massive disappointment.

Who knows, maybe the top end Vega will havr more than 4096 shaders.

Its a case of discarding useless information and not drawing any conclusions, rather than trying to find a way to give it significance, and it is certainly not insisting that the opposite of it must be true.

 

[muon]

Its a case of discarding useless information and not drawing any conclusions, rather than trying to find a way to give it significance, and it is certainly not insisting that the opposite of it must be true.

I don't think it is useless unless someone can show this informatiom is fake.

Which nobody can. Only that it doesn't line up with people's prior expectations of a massive IPC improvement. In fact I would bet against one given all the information from the last few months. I can be hopeful there will be one, but that is all it is.

 

[hyperseven]

I don't think it is useless unless someone can show this informatiom is fake.

Which nobody can. Only that it doesn't line up with people's prior expectations of a massive IPC improvement. In fact I would bet against one given all the information from the last few months. I can be hopeful there will be one, but that is all it is.

^
so much this.

 

[N19h7m4r3]

I don't think it is useless unless someone can show this informatiom is fake.

Which nobody can. Only that it doesn't line up with people's prior expectations of a massive IPC improvement. .

Can you show me that the information is accurate?

 

[akarypid]

Which conflicts with that I just linked you. Vega is reported to have Improved IPC, not less.

http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser/2

So that "leak" still doesn't add up at all.

So then the leak is likely totally fake. I'd much rather take AMD's word from this slide over that leak...

EDIT: Or it's a true Vega but with the L2 cache disabled or something else artificially hindering performance.

 

[muon]

Can you show me that the information is accurate?

Hardware ID identifies the card as an RX Vega and it has been run on 3D Mark which does basic checks on fake results.

TPU have identified the same specs separately.

Over to you.