NVIDIA Maxwell Speculation Thread

Those weren't any "rumors" per se. Anything for =/>600mm2 were guessed values based on the GM200 die shots that leaked out. Rumors long before that shot appeared from supposed insider information where in the 560+mm2 range. Die size however doesn't necessarily answer his question. How powerful? My guess is on standard clocks for the full chip =/>35% on top of GTX980 performance on average.

We are expecting at the minimum a 50% increase in resources (ALUs, TMUs, ROPs, B/W) compared to GM204..possibly even more if the die area is >600 mm2 (and there aren't a huge number of FP64 ALUs). Now of course that does not directly correlate to performance..but 50% more resources should result in at least a 40% increase in performance IMHO, especially considering that clocks should be similar, as GM200 will not be power limited.
 
I'm always conservative in my estimates. I said =/>35% for the record and even if you give me a benchmark parcour that gives those +40% I'll just add a few primarily CPU bound games to the mix and you're quickly back to 35%. Besides the bigger the chip the likelier they'll need more conservative frequencies to stay within power targets for the default showcase, irrelevant if custom SKUs blow the living hell out of them in terms of power and easily reach that magical +50 persentage.
 
I'm always conservative in my estimates. I said =/>35% for the record and even if you give me a benchmark parcour that gives those +40% I'll just add a few primarily CPU bound games to the mix and you're quickly back to 35%. Besides the bigger the chip the likelier they'll need more conservative frequencies to stay within power targets for the default showcase, irrelevant if custom SKUs blow the living hell out of them in terms of power and easily reach that magical +50 persentage.

Agreed..you did say 35% and as you say you were being conservative. And you are right that there will be some CPU limited games but those are very few. However, there may also be some games which scale the entire 50% so that will again skew the average more towards the higher figure. Besides, one thing I should have mentioned is that GM204's biggest weakness is in high resolution and/or multi-monitor gaming. Here it's 256 bit bus does show its limitations (R9 290X for example is 80% of GTX 980's performance at 1920X1080 but catches up to 90% at 3840X2160). Anyone buying a GM200 would probably be running it at high resolutions and its extra bandwidth will give it an edge compared to GM204.
 
GTX 980 has a 165W TDP. If we assume a new part that's 50% larger and with similar clocks, we should end up with about 50% more power consumption, which puts it at a bit less than 250W, which is perfectly reasonable for a very high end card. Hence, I don't see any reason it should be clocked and differently than the 980.
 
GTX 980 has a 165W TDP. If we assume a new part that's 50% larger and with similar clocks, we should end up with about 50% more power consumption, which puts it at a bit less than 250W, which is perfectly reasonable for a very high end card. Hence, I don't see any reason it should be clocked and differently than the 980.
980 only has a 165W tdp in nvidia's marketing tho. The card can draw ~200W if you want to get the same performance as the reviewed cards. In most of the reviews I seen, the card draws more power than the 680, which is a 190W TDP card. If they want to make a card that is 50% more of everything, they would be hitting near 300W.
 
...And a larger chip would have higher variation over the chip and therefore likely need higher voltage to run stable at the same clock, thus consuming more power per area. I wouldn't expect GM200 to run at 980 clocks.
 
GTX 980 has a 165W TDP. If we assume a new part that's 50% larger and with similar clocks, we should end up with about 50% more power consumption, which puts it at a bit less than 250W, which is perfectly reasonable for a very high end card. Hence, I don't see any reason it should be clocked and differently than the 980.
GTX 980 reference models TDP is 180W, not 165.
 
WOW!

71450.png


http://www.anandtech.com/show/8962/the-directx-12-performance-preview-amd-nvidia-star-swarm/5
 
I batted an eye at the full system power usage between the Geforce GTX 980 (with it's awesome marketing "165W" TDP) and the Radeon HD 290X (with the horrible perf/watt). The difference amounts to 14 Watt under DirectX12.

It would seem Maxwell v2 does well under DirectX11 because it waits for the API to fetch it things to do. When that limitations is removed it has to work more (albeit it obviously also delivers ~36% higher performance but at a ~48% price increase).
 
Last edited:
I batted an eye at the full system power usage between the Geforce GTX 980 (with it's awesome marketing "165W" TDP) and the Radeon HD 290X (with the horrible perf/watt). The difference amounts to 14 Watt under DirectX12.

It would seem Maxwell v2 does well under DirectX11 because it waits for the API to fetch it things to do. When that limitations is removed it has to work more (albeit it obviously also delivers ~36% higher performance but at a ~48% price increase).

That is Total System Power Consumption mind you and GTX980 has 50% higher performance than 290X, while being the only card that is not nearly bottlenecked with only 2 cores in DX12. So nVIDIA drivers might be pushing more the CPU, generating higher CPU energy consumption in turn. This in fact something you can see clearly on the CPU usage charts. On GTX980 the graphs are on average at around 75% with spikes to 100%, while on 290X after an initial jump to close to that, it quickly reduces to around 50% per core with spikes to 75%. Lets say that GTX980 is using the CPU 25% more than the 290X. That should ammount to some of the difference in power consumption between GTX980 and 290X ;)
 
Last edited:
That is Total System Power Consumption mind you and GTX980 has 50% higher performance than 290X, while being the only card that is not nearly bottlenecked with only 2 cores in DX12. So nVIDIA drivers might be pushing more the CPU, generating higher CPU energy consumption in turn. This in fact something you can see clearly on the CPU usage charts. On GTX980 the graphs are on average at around 75% with spikes to 100%, while on 290X after an initial jump to close to that, it quickly reduces to around 50% per core with spikes to 75%. Lets say that GTX980 is using the CPU 25% more than the 290X. That should ammount to some of the difference in power consumption between GTX980 and 290X ;)

Yes, didn't you read what I wrote? I did write full system power.

I am implying that Maxwell isn't sipping power as some seems to suggest but use just as much as any other 28nm product (and the TDP listed in the BIOS also supports this). I am well aware that it gets better performance per watt but it sure as hell doesn't use a lot less power. They didn't measure power from the PCI Express slots, so your numbers are simply speculations at best for now.

Again, going by the numbers from AnandTech the Geforce GTX 980 is ~36% faster (66.8 vs 42.9 FPS), using ~ 5% less power under DirectX12 but it is also ~48% more expensive.
 
It's "WOW" because DX11 nvidia results are pointless here, with DX11 + deffered contexts + latest WHQL driver I got 52-55 FPS with GTX780 / i7 960 in this benchmark with Extreme preset and there are already much more DX11 + deffered contexts games than both Mantle + DX12


Indeed, GTX 980 scores nearly 3 folds this score under Windows 7/8 and DX11, I don't get why they would disable Deferred Context in the new version of StarSwarm? is it to make DX12 look good?

Again, going by the numbers from AnandTech the Geforce GTX 980 is ~36% faster (66.8 vs 42.9 FPS),
Your math is wrong, GTX 980 is exactly 56% faster than 290X, not 36%.
 
Last edited:
Yes, didn't you read what I wrote? I did write full system power.

I am implying that Maxwell isn't sipping power as some seems to suggest but use just as much as any other 28nm product (and the TDP listed in the BIOS also supports this). I am well aware that it gets better performance per watt but it sure as hell doesn't use a lot less power. They didn't measure power from the PCI Express slots, so your numbers are simply speculations at best for now.

Again, going by the numbers from AnandTech the Geforce GTX 980 is ~36% faster (66.8 vs 42.9 FPS), using ~ 5% less power under DirectX12 but it is also ~48% more expensive.

Funny that my numbers are pure speculation and you seem to believe yours (which are wrong by the way), which did not even take into account the CPU usage numbers that ARE available, are not. On top of that, you cannot seem to be able to calculate percentages, so how do you expect me to give you any credit? You are clearly not here to discuss matters in an educated way, but simply to smear nVIDIA. That is your central argument, nevermind what other sensible things other people might tell you.

And, by the way, if below is not speculation on your part (when saying "it seems" means you are jumping to conclusions), I dunno what speculation is then.

It would seem Maxwell v2 does well under DirectX11 because it waits for the API to fetch it things to do. When that limitations is removed it has to work more (albeit it obviously also delivers ~36% higher performance but at a ~48% price increase).

Plus, it does not even make sense at all. If that would be true, GTX980 would not be much faster than 290X under DX11. Yet IT IS, and even without the defered contexts! With the defered contexts enabled, the difference between DX11 and DX12 is much lower. Again, you are just here making whatever is needed to smear nVIDIA. Oh, and the price reference when this is just a benchmark killed me :D
 
Last edited:
Your math is wrong, GTX 980 is exactly 56% faster than 290X, not 36%.
Yeah, I screwed up on the math. I miscalculated. I went (66,8 - 42,9) / 66,8 = 23,9 / 66,8 = 0,357 * 100 = ~36%, instead of the other way around (~55,7%).

Funny that my numbers are pure speculation and you seem to believe yours (which are wrong by the way), which did not even take into account the CPU usage numbers that ARE available, are not. On top of that, you cannot seem to be able to calculate percentages, so how do you expect me to give you any credit? You are clearly not here to discuss matters in an educated way, but simply to smear nVIDIA. That is your central argument, nevermind what other sensible things other people might tell you.

Oh dear, I am in no way trying to smear NVIDIA, I simply miscalculated the performance difference. I am well aware that the GM204 is better in every metric but also vastly more expensive.
 
Last edited:
It's "WOW" because DX11 nvidia results are pointless here, with DX11 + deffered contexts + latest WHQL driver I got 52-55 FPS with GTX780 / i7 960 in this benchmark with Extreme preset and there are already much more DX11 + deffered contexts games than both Mantle + DX12
Make sure you're running the RTS scenario and not the Follow scenario. I've had this question asked of me a few times now; every time they were running the Follow scenario.

Yes, didn't you read what I wrote? I did write full system power.

I am implying that Maxwell isn't sipping power as some seems to suggest but use just as much as any other 28nm product (and the TDP listed in the BIOS also supports this). I am well aware that it gets better performance per watt but it sure as hell doesn't use a lot less power. They didn't measure power from the PCI Express slots, so your numbers are simply speculations at best for now.

Again, going by the numbers from AnandTech the Geforce GTX 980 is ~36% faster (66.8 vs 42.9 FPS), using ~ 5% less power under DirectX12 but it is also ~48% more expensive.
I would strongly suggest not reading too much into the power numbers. I included them because they were interesting, but mostly because it shows how power usage increased on each card with the change in APIs (mostly from the additional work the system was able to do).

Let's be clear here: these are alpha quality drivers on an alpha quality OS, running a non-deterministic benchmark that is itself alpha quality (Nitrous engine development is ongoing). It is not and should not be used to draw significant conclusions about competitive products at this time.
 
Last edited:
I would strongly suggest not reading too much into the power numbers. I included them because they were interesting, but mostly because it shows how power usage increased on each card with the change in APIs (mostly from the additional work the system was able to do).

Let's be clear here: these are alpha quality drivers on an alpha quality OS, running a non-deterministic benchmark that is itself alpha quality (Nitrous engine development is ongoing). It is not and should not be used to draw significant conclusions about competitive products at this time.

Guess I will just crawl back into my hole.

Thanks for clarifying, Ryan.
 
Back
Top