Vega/Navi Rumors (Updated)

[Armsdealer]

From the pcworld video.

http://www.pcworld.com/article/3203...arks-and-behind-the-scenes-amd-interview.html

Pause at 15:05

Vega FE benchmarks -

Catia-04: 137.68
Creo-01: 88.55
Sw-03: 112.13

Forget quadro p6000 (gp102)

Here are equivalent 8.9 tflop p5000 (gp104) benchmarks

Catia-04: 156.2
Creo-01: 116.24
Sw-03: 156.72

From http://www.geeks3d.com/20170515/test-nvidia-quadro-p5000-vs-geforce-gtx-1080/2/#_42

Ipc improvements anyone?

 

[tajoh111]

From the pcworld video.

http://www.pcworld.com/article/3203...arks-and-behind-the-scenes-amd-interview.html

Pause at 15:05

Vega FE benchmarks -

Catia-04: 137.68
Creo-01: 88.55
Sw-03: 112.13

Forget quadro p6000 (gp102)

Here are equivalent 8.9 tflop p5000 (gp104) benchmarks

Catia-04: 156.2
Creo-01: 116.24
Sw-03: 156.72

From http://www.geeks3d.com/20170515/test-nvidia-quadro-p5000-vs-geforce-gtx-1080/2/#_42

Ipc improvements anyone?

Just to add onto these results.

https://www.spec.org/gwpg/gpc.data/vp12.1/summary.html

P2000 results. = $440usd.

Catia-04: 112.33
Creo-01: 102.48
Sw-03: 139.42

P4000 results price = $900usd

Catia-04: 155.63
Creo-01: 120.93
Sw-03: 170.02

Updated p5000 results. Price = $2000usd

Catia-04: 179.99
Creo-01: 138.52
Sw-03: 196.69.

The comparison vs the p2000 hurts the most. Trades blows for close to a third the cost.

 

[Bacon1]

Updated p5000 results. Price = $2000usd

Catia-04: 179.99
Creo-01: 138.52
Sw-03: 196.69.

You are using this result:

https://www.spec.org/gwpg/gpc.data/...ults_20170207T1002_r1_HIGHEST/resultHTML.html

Not this one:

Catia-4: 153.26
Creo-01: 120.32
Sw-03: 167.57

https://www.spec.org/gwpg/gpc.data/vp12.1/Fujitsu/CELSIUS_R940_E5-2699v4_P5000/resultHTML.html

which clearly shows system configuration makes a huge difference in performance... so using apples - apples is required.

 

[Borf]

I wonder what the specs on the saw are.

 

[Armsdealer]

Nitpicking aside, the big picture from these data is vega seems to be a power hungry gp104 competitor that's costing amd a bundle to manufacture given their decision to go with hbm. I'm sure some select benchmarks will put the rx vega ahead of the gtx 1080, but the burden of proof that it's a genuine gp 102 competitor (or more, as some believe) really rests on the optimists.

To be perfectly frank, heads should roll over there on the hbm decision.

 

[Karnak]

Vega will be a GP102 competitor, only GP104 level of performance is just not gonna happen. Even a Fury X with ~1600MHz would be (almost) on par with a 1080.

 

[Bacon1]

power hungry gp104 competitor

Uhhhh what?

Even a Fury X with ~1600MHz would be (almost) on par with a 1080.

Fury with a ~50% OC would do very well:

Here is one with an OC to 1400:

http://wccftech.com/asus-radeon-r9-...y-unlocked-fury-features-1-tbs-bandwidth-ln2/

 

[Armsdealer]

Vega will be a GP102 competitor, only GP104 level of performance is just not gonna happen. Even a Fury X with ~1600MHz would be (almost) on par with a 1080.

How are we in disagreement? Let's say the top binned part of vega will rest above gp104 and below gp 102, which grants vega some ipc improvement over fiji. When you consider the entire product lineup, with cut dies, that means it's a gp 104 competitor, not a 102 competitor.

If you're going to argue ipc improvements are going to put it on par with 1080ti, well, I hope you're right but so far we're seeing hard evidence suggest otherwise with full die vega being ~20% - 25% worse off than full die 104 in the benchmarks above. Do you honestly believe in gaming they're going to be 20% better? Possible but it's irrational to think that.

 

[tviceman]

Vega will be a GP102 competitor, only GP104 level of performance is just not gonna happen. Even a Fury X with ~1600MHz would be (almost) on par with a 1080.

At 500 watts.... AMD's problem isn't IPC. People have these fantasies of AMD cards clocking as high as Nvidia and WOW imagine the performance!!! ...Reality needs to set in; the only thing that matters is end performance, not mhz. AMD and Nvidia are taking different routes to get there.

AMD's problem is perf/w. We are already seeing AMD hasn't fixed that problem with Vega and it's ominus TDP levels 20% and 50% higher than Titan Xp. And don't bother commenting on TDP vs. real world readings; both companies are accurately pegging TDP levels these days, with Nvidia being somewhat more conservative than AMD.

 

[Elfear]

How are we in disagreement? Let's say the top binned part of vega will rest above gp104 and below gp 102, which grants vega some ipc improvement over fiji. When you consider the entire product lineup, with cut dies, that means it's a gp 104 competitor, not a 102 competitor.

If you're going to argue ipc improvements are going to put it on par with 1080ti, well, I hope you're right but so far we're seeing hard evidence suggest otherwise with full die vega being ~20% - 25% worse off than full die 104 in the benchmarks above. Do you honestly believe in gaming they're going to be 20% better? Possible but it's irrational to think that.

In order to give credence to what you say, we'd have to assume that Vega had an IPC regression compared to Polaris or at the very least it's on par. Here are the latest results from TPU:

If we base Vega's performance off of Polaris, there is a 26% increase in clockspeed (1266Mhz vs 1600Mhz) and a 78% increase in stream processors (2304 vs 4096) resulting in a score of 370% on the above chart.

If we base it on Fiji performance, there is a 52% increase in clock speed which would put Vega at 318% on the chart above.

Obviously, performance isn't a 1:1 increase with clock speed or with stream processors but you're essentially saying Vega will have the same IPC as Fiji if it falls in between the 1080 and 1080Ti. That's a regression in comparison to Polaris...

 

[Bacon1]

At 500 watts....

Vega has a TDP of 300w not 500w...

And don't bother commenting on TDP vs. real world readings; both companies are accurately pegging TDP levels these days, with Nvidia being somewhat more conservative than AMD.

250w 1080 Ti uses 280w in actual gaming usages after a small factory OC.

We already know Vega's clocks are @1600 range due to the TFLop and core count.

TFlops = Cores * Clock * 2

13.1TFlop = 4096 * x * 2

13.1TFlop = 8192 * x

x = 13.1TFlop / 8192

x = 1600MHz

So at worse case it should perform like a Fiji OC'd @ 1600Mhz. With all of the arch improvements I doubt it would do equal to that, let alone worse than Fiji @ 1600Mhz.

 

[JDG1980]

Release date for Vega FE is supposed to be tomorrow. Do we get real reviews from sites like TPU and TechReport? It's clear that AMD doesn't want any game benchmarks for FE, because they've done everything they can to avoid leaking any real gaming numbers (which means they must suck).

 

[3DVagabond]

Release date for Vega FE is supposed to be tomorrow. Do we get real reviews from sites like TPU and TechReport? It's clear that AMD doesn't want any game benchmarks for FE, because they've done everything they can to avoid leaking any real gaming numbers (which means they must suck).

They've said the "gaming cards" will be better. We'll have to see but that means that gaming benchmarks might not be particularly useful.

Typically these cards don't go to sites like TPU.

 

[railven]

Vega has a TDP of 300w not 500w...



250w 1080 Ti uses 280w in actual gaming usages after a small factory OC.

We already know Vega's clocks are @1600 range due to the TFLop and core count.

TFlops = Cores * Clock * 2

13.1TFlop = 4096 * x * 2

13.1TFlop = 8192 * x

x = 13.1TFlop / 8192

x = 1600MHz

So at worse case it should perform like a Fiji OC'd @ 1600Mhz. With all of the arch improvements I doubt it would do equal to that, let alone worse than Fiji @ 1600Mhz.

Pretty sure he's talking about the metaphorical 1600mhz Fury X, not Vega, regarding the power consumption.

 

[tviceman]

Vega has a TDP of 300w not 500w...

Yes I know what the TDP of Vega is. The 500w reference was to all the comments about Fiji @ 1600mhz being able to match a GTX 1080. If you would have read what I quoted you would have realized that.

250w 1080 Ti uses 280w in actual gaming usages after a small factory OC.

Almost all custom AIB cards consume considerably more power than reference models. A reference 1080 TI consumes 230w, and after that it depends entirely on the vendor.
OC'd at 250w https://www.techpowerup.com/reviews/EVGA/GTX_1080_Ti_SC2/29.html
OC'd at 250w https://www.techpowerup.com/reviews/Gigabyte/GTX_1080_Ti_Xtreme_Gaming/28.html

Quit lying to yourself, trying to fudge numbers, including AIB custom models which almost always dips considerably in perf/w.

We already know Vega's clocks are @1600 range due to the TFLop and core count.

TFlops = Cores * Clock * 2

13.1TFlop = 4096 * x * 2

13.1TFlop = 8192 * x

x = 13.1TFlop / 8192

x = 1600MHz

So at worse case it should perform like a Fiji OC'd @ 1600Mhz. With all of the arch improvements I doubt it would do equal to that, let alone worse than Fiji @ 1600Mhz.


If we base Vega's performance off of Polaris, there is a 26% increase in clockspeed (1266Mhz vs 1600Mhz) and a 78% increase in stream processors (2304 vs 4096) resulting in a score of 370% on the above chart.

If we base it on Fiji performance, there is a 52% increase in clock speed which would put Vega at 318% on the chart above.

Obviously, performance isn't a 1:1 increase with clock speed or with stream processors but you're essentially saying Vega will have the same IPC as Fiji if it falls in between the 1080 and 1080Ti. That's a regression in comparison to Polaris...

380x TFLOPS: 4
Fury X TFLOPS: 8.6

Fury X was 95% faster at 4k and 89% faster at 1440p despite having 2.15x more tflops.

If that kind of drop off in real world performance carries over between Polaris and Vega, and it's very likely it will since this phenomena is inherent to chips that scale up in size, then when factoring RX 580's reference speed 6.17 TFLOPS, full Vega will be about 92% faster than a reference RX 580 at 4k and 86% faster at 1440p, which puts it right in between a 1080 and 1080 TI while having a 300w TDP.

 

[Bacon1]

Release date for Vega FE is supposed to be tomorrow. Do we get real reviews from sites like TPU and TechReport? It's clear that AMD doesn't want any game benchmarks for FE, because they've done everything they can to avoid leaking any real gaming numbers (which means they must suck).

They've said a few times that the gaming related "RX Vega" will be faster and cheaper so gamers should wait for those next month.

Pretty sure he's talking about the metaphorical 1600mhz Fury X, not Vega, regarding the power consumption.

So it has no bearing on the discussion then ok. Vega has about the same power for 4x the memory and >50% higher clocks.

So using the 1080 Ti SC2 @ 251w: 370 - 251 = 119w for the rest of the system.

374.8 - 119 = 255.8w -> 305w for the Titan XP.

Vega has a TDP of 300w for the air and we'll see actual power consumption tomorrow.

 

[Bacon1]

380x TFLOPS: 4
Fury X TFLOPS: 8.6

Fury X was 95% faster at 4k and 89% faster at 1440p despite having 2.15x more tflops.

Because TFLops != gaming performance at all. Its literally just the clocks * core count and ignores arch and the rest of the hardware required for gaming needs.

Trying to match TFLops to gaming performance for different architectures doesn't make any sense at all.

 

[Armsdealer]

In order to give credence to what you say, we'd have to assume that Vega had an IPC regression compared to Polaris or at the very least it's on par. Here are the latest results from TPU:

Your calculation is highly theoretical and undeniably carries some water, but you yourself solved the quandary: Scaling is highly imperfect. In the graph above just glance at the relative performance of the 1080 vs the 1070. 1080 is 8.9 tflops vs 6.5 for the 1070 or a 37% difference. In tpu's actual tests they found a 25% difference. When you account for a similar imperfection in scaling in your polaris -> vega or Fiji -> vega calculation, you'll end up with roughly the same answer as me - that full die vega is not really going to give 1080ti a run for its money regardless of which reference point you use. This is FURTHER suggested by the ACTUAL results we have from amd themselves. Either way I hope you're right and it's a bombshell product. Unfortunately I'm not seeing it.

 

[railven]

So it has no bearing on the discussion then ok.

Should probably ask the most recent poster that brought it up. A OC'd to the moon Fury X seems to some how keep popping up (oddly enough same with Polaris).

 

[Bacon1]

Should probably ask the most recent poster that brought it up. A OC'd to the moon Fury X seems to some how keep popping up (oddly enough same with Polaris).

Because performance would be similar, but obviously the power usage is not since we already know the current power is nowhere near 500w.

 

[exquisitechar]

Should probably ask the most recent poster that brought it up. A OC'd to the moon Fury X seems to some how keep popping up (oddly enough same with Polaris).

Vega clocks like a Fury X OC'd to the moon (and better), so it's reasonable to bring it up in a prediction for Vega's baseline performance. Obviously, the power usage will be nowhere near the same since it's fabricated on 14nm and designed for higher clocks.

 

[Ambrosia1313]

Speaking of Vega/Navi I have a question regarding memory and bus width.
Looking at these 2 articles I have a question about the math itself as to how to fit more memory without widening the bus.

http://www.guru3d.com/news-story/sk-hynix-hbm2-and-gddr6-specs-and-availibity.html
http://www.guru3d.com/news-story/sk-hynix-showcases-first-gddr6-double-the-bandwidth.html

How would one go about adding 16/20/24 or even 32Gb of VRAM while keeping the bus at 384-bit?
The best I came up with and looking at GPU's in general was they all come back to 32 bit per chip (e.g 8*1Gb*32-bit = bus width).
But for the love of god I can't figure out how it actually works, so have a look at GDDR6, it's density and IIRC it's max bandwidth is 64GB/s up from 56Gb/s GDDR5x and 28GB/s GDDR5.

 

[beginner99]

So at worse case it should perform like a Fiji OC'd @ 1600Mhz. With all of the arch improvements I doubt it would do equal to that, let alone worse than Fiji @ 1600Mhz.

I agree with you but we had the same logic before Bulldozer released and IPC did regress. You never know with AMD.

 

[railven]

Because performance would be similar, but obviously the power usage is not since we already know the current power is nowhere near 500w.

Vega clocks like a Fury X OC'd to the moon (and better), so it's reasonable to bring it up in a prediction for Vega's baseline performance. Obviously, the power usage will be nowhere near the same since it's fabricated on 14nm and designed for higher clocks.

And then when someone tries to make a estimation using Fury X "you can't, different uarchs, IPC gains" list goes on.

Oh well. Back to waiting for more Vega info, tomorrow can't some soon enough.

 

[exquisitechar]

And then when someone tries to make a estimation using Fury X "you can't, different uarchs, IPC gains" list goes on.

Oh well. Back to waiting for more Vega info, tomorrow can't some soon enough.

Well, you can't say that's all it can do, but it serves as a baseline. And it certainly proves that Vega could easily be superior to GP104, since many seem to believe it'll be no better than the GTX 1080 (although I believe it's already surpassed overclocked GTX 1080s in demos AMD has shown). Anything more than that is difficult to say with what we know now.