[WCCFTECH]Possible AMD Radeon RX 490 Performance Numbers Show Up in DX12 AOTS Benchmark – On Par Wit

[krumme]

Wow, there's someone who still thinks power consumption has something to do with your electricity bill?

Say that to the hordes of people that talk about power. We got temp and noise for the rest.
But i guess it depends on the situation....

 

[antihelten]

So this is what it has come to? We have now gotten to the point where we are denying the existence of products that are readily available for purchase? The P100 was the first pascal product announced back in April.

http://www.guru3d.com/news-story/nv...ecap-full-gpu-has-3840-shader-processors.html

GP100 does indeed have 60 shader modules, which is obviously twice that of GP102, however the shader modules in GP100 only contain 64 FP32 CUDA cores each, compared to 128 for each module in GP102.

As such GP100 is not twice as fast as GP102 as you previously claimed, quite the contrary it is exactly the same speed. Obviously GP100 also comes equipped with a large number of FP64 CUDA cores, but these are not usable in gaming contexts (at least not at the same time as the FP32 cores, to my understanding).

So currently Nvidia has two options. 1) release a fully enable GP102 card, which would probably be about 10% faster or so than Titan X. 2) design a completely new 600 mm2 gaming focused die (GP102 is only 471 mm2 after all), which (assuming linear scaling) could squeeze in about 38 shader modules (about 30% more than GP102). Given that we haven't had a single leak hinting of the latter (to my knowledge), the most likely scenario would be the first one.

Of course all of this is rather moot unless AMD can produce a challenger for the GP102, which is not what this thread is really about (RX 490 would be a competitor for GP104 apparently, not GP102).

 

[Glo.]

GP100 does indeed have 60 shader modules, which is obviously twice that of GP102, however the shader modules in GP100 only contain 64 FP32 CUDA cores each, compared to 128 for each module in GP102.

As such GP100 is not twice as fast as GP102 as you previously claimed, quite the contrary it is exactly the same speed. Obviously GP100 also comes equipped with a large number of FP64 CUDA cores, but these are not usable in gaming contexts (at least not at the same time as the FP32 cores, to my understanding).

So currently Nvidia has two options. 1) release a fully enable GP102 card, which would probably be about 10% faster or so than Titan X. 2) design a completely new 600 mm2 gaming focused die (GP102 is only 471 mm2 after all), which (assuming linear scaling) could squeeze in about 38 shader modules (about 30% more than GP102). Given that we haven't had a single leak hinting of the latter (to my knowledge), the most likely scenario would be the first one.

Of course all of this is rather moot unless AMD can produce a challenger for the GP102, which is not what this thread is really about (RX 490 would be a competitor for GP104 apparently, not GP102).

GP100 indeed has 60 shader modules, or SM's in other words. They indeed contain only 64 CUDA cores, compared to 128 in Maxwell and Consumer Pascal.

The problem is, that GP100 indeed will be twice as fast. The amount of resources available to each 64 core is doubled compared to Maxwell and consumer Pascal(consumer Pascal indeed is Maxwell architecture on 16 nm TSMC). Nvidia did this before.

Kepler - Maxwell - Pascal. The amount of resources available for each number of cores is exactly the same, but the number of cores that has access to that level of resources is different.

192 cores for Kepler, 128 for Maxwell, 64 for GP100. Nvidia have said that 128 Maxwell cores have 90% of performance of Kepler's 192 cores, because of this very reason. So 64 cores in GP100 will have 90-100% of 128 Maxwell/Consumer Pascal cores.

End of off-top.

 

[Qwertilot]

So currently Nvidia has two options. 1) release a fully enable GP102 card, which would probably be about 10% faster or so than Titan X. 2) design a completely new 600 mm2 gaming focused die (GP102 is only 471 mm2 after all), which (assuming linear scaling) could squeeze in about 38 shader modules (about 30% more than GP102). Given that we haven't had a single leak hinting of the latter (to my knowledge), the most likely scenario would be the first one.

That sort of very big die can be (reasonably safely I think) logically presumed to be due with big Volta, due out whenever. (+- 2018ish.). There isn't very much power headroom in GP102 for a gaming GPU - especially not fully enabled - so they've got to up the efficiency a chunk for that to make sense as an idea.

 

[Cookie Monster]

Unless the card is using utterly ridiculous amounts of power I really don't care. A difference of 30,40,50,60W is nothing. It's one fairly dim light bulb yet its somehow made out to be a massive selling point and hugely important metric. Price/perf is the key, nail that and it'll sell.

You are grossly simplifying the importance of power consumption. Clearly consumers look at price/perf but for IHVs that is the not their number 1 priority. For those who want AMD as a company to keep competing or in other words have actual options when out shopping for a video card.. even a product that consumes 10W less is a big win.

Power consumption is tied with so many things on a video card. First is thermals. Thermals relate to the cooling solution. Less power consumption -> Smaller Cooler -> Cheaper BOM. Less heat also means less noisy cooling solutions as you dont need a closed loop AIO solution nor a fan spinning at ridiculous RPMs (while taking up 2~3 slots). More elaborate cooling solutions are more expensive, especially if more moving parts like fans/pumps are involved (more points of failure = more cost). Second is the circuitry themselves. Less power consumption -> less components for VRM/simpler PCB design etc -> Cheaper BOM. From a dual 6 pin PCI-e design to a single 6/8pin also means -> Cheaper BOM (less components, yes even simple passive connectors = less cost!). All of these (there are more) lead to the well being of the company and more flexibility in pricing. Saving even 10c from your BOM when your shipping units in the millions is a big win. Basically power consumption = cost cost cost!! (assuming performance is there). Ive yet to mention about supply issues (which can also lead to a costing issue) because higher power consumption and elaborate cooling systems require better components which are often less available than cheaper more mainstream parts.

For us, it may not mean diddly squat, but for AMD and OEMs in particular, its huge. If VEGA10 is indeed a GP104 competitor, but requires a much more elaborate cooling (like a closed loop AIO cooling solution), on top of using HBM2 which is most likely to be more expensive than any GDDR based memory while being at 250W TDP.. its not going to bode well for AMD. If it performs closer to GP102 then I think it will atleast be alot more attractive option than being a GP104 competitor but thats asking alot (it means they need to release something that is 60~70% faster than FuryX or 120~130% than Polaris 10!! - approx values from TPU performance summary - https://www.techpowerup.com/reviews/NVIDIA/Titan_X_Pascal/24.html) and the performance gap is widening between nVIDIA/AMD.

AMD's GP104 competitor needs to be LEAN (from an engineering perspective) to be competitive financially while the GP102 competitor can be the full monty imo as its a niche/halo part.

 

[frozentundra123456]

The FuryX is already pretty close to 980Ti/1070 performance, I can see AMD coming up with something 1080 level this round.

When I don't know, but AMD graphics is definitely a bigger player than AMD CPUs.

Those of us who have don't have as much disposable cash to blow on gaming PCs had better hope they do. 1080 sales show there is no shortage of people willing to pay $700+ for a mid-range card.

Well, AMD better hope Zen is a better competitor than their dgpus. Despite all the hype for polaris and now in this thread, the reality is that they only have one competitive card, the 470. (Well actually they have two, the 470 and 480, but they are so close in performance they pretty much compete with each other, and only cover the same segment.) Also, if AMD manages to come out with a 1080 competitor, it will be quite interesting to see if the usual suspects bash it as a "mid-range" like they are obsessed with doing for the 1080.

 

[OatisCampbell]

Well, AMD better hope Zen is a better competitor than their dgpus. Despite all the hype for polaris and now in this thread, the reality is that they only have one competitive card, the 470. (Well actually they have two, the 470 and 480, but they are so close in performance they pretty much compete with each other, and only cover the same segment.) Also, if AMD manages to come out with a 1080 competitor, it will be quite interesting to see if the usual suspects bash it as a "mid-range" like they are obsessed with doing for the 1080.

Can't speak for the usual suspects, but the 980 and 1080 were the next gen mid range cards.

If big Vega is 1080 FE level performance it will be mid range for this generation as well. Can't just pretend the Titan X Pascal doesn't exist, and the 1080Ti will be launched soon as well. Facts are facts.

In the era of console ports, that level of performance is good enough for most.

 

[frozentundra123456]

Can't speak for the usual suspects, but the 980 and 1080 were the next gen mid range cards.

If big Vega is 1080 FE level performance it will be mid range for this generation as well. Can't just pretend the Titan X Pascal doesn't exist, and the 1080Ti will be launched soon as well. Facts are facts.

In the era of console ports, that level of performance is good enough for most.

Yea, lets look at the facts. The fact is, that except for the fringe Titan X, the 1080 is the highest performing, most expensive card available. Things can obviously change as new cards are released, but right now, and ever since it came out, the 1080 *is* a high end card. And even if you consider the Titan X, the 1080 is still a high end card. (There can be more than one high end card you know). And if you consider the 1080 "mid range", what is the 1070, and even more so, the 480, 1060, 470, etc and all the new cards of even lower performance? They must be "low end" then, right?

 

[Ranulf]

They're high end in price, certainly not in performance.

 

[frozentundra123456]

What is faster?

 

[AtenRa]

People easily forget or chose not to talk about Fury Nano 175W TDP vs GTX 980Ti 250W TDP in 2016 games.

This card completely destroys GTX 980Ti in perf/watt at 1440p and especially in 4K.
Its amazing what this little card (28nm) can manage at 175W TDP and half the card size vs GTX980Ti.

 

[Shmee]

I wonder how much vega would cost, if it is about on par with 1070 or better, I would likely get 2 for CF

 

[Pariah]

People easily forget or chose not to talk about Fury Nano 175W TDP vs GTX 980Ti 250W TDP in 2016 games.

This card completely destroys GTX 980Ti in perf/watt at 1440p and especially in 4K.
Its amazing what this little card (28nm) can manage at 175W TDP and half the card size vs GTX980Ti.

Nvidia doesn't produce golden sample limited run desktop products. It's an apples to oranges comparison. You seem to have easily forgotten that Nvidia took their golden samples and made a GTX 980 that was crammed into a 17" laptop with a TDP of 125-150w (OEM configurable) without the benefit of HBM. Which product do you think was more profitable?

 

[AtenRa]

Nvidia doesn't produce golden sample limited run desktop products. It's an apples to oranges comparison. You seem to have easily forgotten that Nvidia took their golden samples and made a GTX 980 that was crammed into a 17" laptop with a TDP of 125-150w (OEM configurable) without the benefit of HBM. Which product do you think was more profitable?

Since Fury Nano has way higher perf/watt than Desktop GTX 980, a Mobile Fury Nano will still have even higher Perf/Watt than mobile GTX980 due to HBM memory.

 

[Pariah]

Since Fury Nano has way higher perf/watt than Desktop GTX 980, a Mobile Fury Nano will still have even higher Perf/Watt than mobile GTX980 due to HBM memory.

Edit: misunderstood what you wrote the first time.

Even in the cherry picked benchmarks you chose to put the Nano in the best light above, the nano is only leading the 980 by 30.7% on average. While the TDP of the Nano is 17%-40% higher than the 980 gtx notebook depending on configuration. If you chose a more representative selection of games, instead of the half dozen the Nano doesn't suck at now, the 980 notebook would be significantly more efficient.

 

[AtenRa]

With a conservative 15w drop to 110W TDP for the 980 notebook the Nano's TDP would be 60% higher. Even in the cherry picked benchmarks you chose to put the Nano in the best light above, it's no where near 60% faster on average. If you chose a more representative selection of games, instead of the half dozen the Nano doesn't suck at now, the 980 notebook is probably more efficient as it is currently configured.

Why do you actually believe a mobile Nano will not be faster at 120W vs a 120W Mobile GTX 980 ???

Desktop

175W TDP Nano = Faster than 180W TDP GTX 980

Mobile

120W TDP Nano = Faster than 120W TDP GTX 980

Also, 7 of the latest 2016 games is not cherry pick.

 

[Pariah]

I edited my previous post. I don't get what you are saying with your response. It looks like you are inventing a mobile nano instead of comparing actual products.

 

[AtenRa]

I edited my previous post. I don't get what you are saying with your response. It looks like you are inventing a mobile nano instead of comparing actual products.

Obviously im comparing a hypothetical 120W Mobile Nano against the Mobile GTX 980m.

If the Desktop Nano is faster with higher perf/watt than the GTX 980. In a smaller/higher thermal and lower power environment like the Mobile, Nano will have an even higher advantage due to HBM. So at the same Watt the mobile Nano will still be faster than the Mobile GTX 980. Its not that hard to understand.

GTX 980m has to lower both GPU clocks AND Memory bandwidth to be able to operate at those lower thermals. Due to HBM, Nano will have to lower the GPU clocks less making it even faster than the GTX980m.

Easy to understand actually.

 

[Pariah]

GTX 980m has to lower both GPU clocks AND Memory bandwidth to be able to operate at those lower thermals. Due to HBM, Nano will have to lower the GPU clocks less making it even faster than the GTX980m.

Easy to understand actually.

The last part here is not necessary. All it does it open yourself up to ridicule when your post indicates you don't know what you are talking about. The GTX980m and GTX 980 notebook are 2 unrelated products. Nvidia did a horrid job of naming them, but it is what it is. The GTX 980 notebook's memory runs at the exact same specs as the desktop version while the GPU base clock drops 62Mhz from 1126 to 1064. A 5.5% drop in clock rate with identical memory is not going to hurt performance very substantially.

 

[antihelten]

GP100 indeed has 60 shader modules, or SM's in other words. They indeed contain only 64 CUDA cores, compared to 128 in Maxwell and Consumer Pascal.

The problem is, that GP100 indeed will be twice as fast. The amount of resources available to each 64 core is doubled compared to Maxwell and consumer Pascal(consumer Pascal indeed is Maxwell architecture on 16 nm TSMC). Nvidia did this before.

Kepler - Maxwell - Pascal. The amount of resources available for each number of cores is exactly the same, but the number of cores that has access to that level of resources is different.

192 cores for Kepler, 128 for Maxwell, 64 for GP100. Nvidia have said that 128 Maxwell cores have 90% of performance of Kepler's 192 cores, because of this very reason. So 64 cores in GP100 will have 90-100% of 128 Maxwell/Consumer Pascal cores.

End of off-top.

The only resources that are doubled per CUDA core, relative to GP102, are the registry sizes, and that alone won't allow 64 CUDA cores in GP100 to get anywhere near 128 CUDA cores in GP102, when it comes to gaming (which is what we're talking about here). Bandwidth per core also goes up, but not double (from 480 GB/s to 732 GB/s or a 52.5% increase), so does onboard shared memory (from 96KB for 128 cores to 64KB for 64 cores, or an increase of 33%).

So no, the 64 cores in GP100 will not have 90-100% the performance of 128 cores in GP102, at best it will have 50-60% the performance, mainly due to the increase in memory bandwidth (in reality it will probably have less than 50% the performance, since GP100 likely can't clock as high as GP102).

That sort of very big die can be (reasonably safely I think) logically presumed to be due with big Volta, due out whenever. (+- 2018ish.). There isn't very much power headroom in GP102 for a gaming GPU - especially not fully enabled - so they've got to up the efficiency a chunk for that to make sense as an idea.

I absolutely agree with you that Volta will probably be the next performance jump we see from Nvidia, and not a 600 mm2 gaming Pascal GPU .

With that being said though I do think that there is theoretically a bit more headroom than one might initially think. Obviously with GP102 at 471 mm2 and a TDP of 250W, there is very little room, however this theoretical 600 mm2 GPU would probably need a 512 bit memory interface, at which point Nvidia might as well switch over to HBM2. This would both leave a bit more room on the die for shaders and whatnot (40-44 SM or 33-47% more than GP102), plus it would lower power usage (512bit GDDR5 at 8GHz switched over to 4096 bit HBM1 at 1 GHz allegedly saves 55 W, GDDR5X to HBM2 would likely be in the same ballpark), thus hopefully keeping the whole thing within 300 W.

But again this is all theorycrafting, and I don't really see this happening.

 

[Glo.]

Nvidia doesn't produce golden sample limited run desktop products. It's an apples to oranges comparison. You seem to have easily forgotten that Nvidia took their golden samples and made a GTX 980 that was crammed into a 17" laptop with a TDP of 125-150w (OEM configurable) without the benefit of HBM. Which product do you think was more profitable?

R9 Nano NEVER was Golden Sample, best binned chip. It was exactly the same ASIC as Fury X. It was just power gated to 175W of power average. You could've done exactly the same thing with Fury X, and get exactly the same results.

 

[Keysplayr]

R9 Nano NEVER was Golden Sample, best binned chip. It was exactly the same ASIC as Fury X. It was just power gated to 175W of power average. You could've done exactly the same thing with Fury X, and get exactly the same results.

umm. You cant know that. It implies you know the ASIC of every Fury and Nano produced. kicker is, no two chips have the exact same ASIC. There are a whole lotta decimal places.

 

[Glo.]

umm. You cant know that. It implies you know the ASIC of every Fury and Nano produced. kicker is, no two chips have the exact same ASIC. There are a whole lotta decimal places.

It has been said either by Raja Koduri or Mark Papermaster in 2015, when Nano was actually released.

 

[OatisCampbell]

Yea, lets look at the facts. The fact is, that except for the fringe Titan X, the 1080 is the highest performing, most expensive card available. Things can obviously change as new cards are released, but right now, and ever since it came out, the 1080 *is* a high end card. And even if you consider the Titan X, the 1080 is still a high end card. (There can be more than one high end card you know). And if you consider the 1080 "mid range", what is the 1070, and even more so, the 480, 1060, 470, etc and all the new cards of even lower performance? They must be "low end" then, right?

Perspective and semantics.

I would consider the 680/980/1080 mid range because they all had multiple products that perform much higher in their chip families. (no, the 1080Ti hasn't launched yet, but we all know it will)

I consider the 660/960/1060 "low end" for gaming and cards below that I don't consider. I'm not wealthy but I don't see much point in the X50 cards for desktop gaming. Might as well save money and buy a Xbox or PS4.

 

[frozentundra123456]

Obviously im comparing a hypothetical 120W Mobile Nano against the Mobile GTX 980m.

If the Desktop Nano is faster with higher perf/watt than the GTX 980. In a smaller/higher thermal and lower power environment like the Mobile, Nano will have an even higher advantage due to HBM. So at the same Watt the mobile Nano will still be faster than the Mobile GTX 980. Its not that hard to understand.

GTX 980m has to lower both GPU clocks AND Memory bandwidth to be able to operate at those lower thermals. Due to HBM, Nano will have to lower the GPU clocks less making it even faster than the GTX980m.

Easy to understand actually.

You are comparing a theoretical mobile nano which never existed, and certainly never will, with a previous generation card that has been replaced. OK, that tells us a lot. The relevant comparison is between the most efficient current laptop chips. In fact, nVidia's architecture on 16nm is so efficient that they can essentially put a desktop chip into a laptop. So the comparison should be between a current mobile 1060/1070/1080 and whatever current AMD *laptop* chip offers comparable performance (if there is one).