NVIDIA Maxwell Speculation Thread

This is shaping up to be a very nice chip. The enthusiast that might consider a GTX 980 card might be disappointed in the speed increase, if any over the 780Ti, but on a technical level the chip looks to be pretty impressive.
 
Or the color compression works wonders.

Really hope there is a lot of new features.

Same here


This is shaping up to be a very nice chip. The enthusiast that might consider a GTX 980 card might be disappointed in the speed increase, if any over the 780Ti, but on a technical level the chip looks to be pretty impressive.

Agree, 780ti performances in 165w tdp and 28nm looks amazing
 
GM204 165W TDP: less power, more gaming performance, more ROPs, more GPGPU performance, more GPU Clock, ~5B transistors, 28nm
Tonga 190W TDP: more power, less gaming performance, less ROPs, less GPGPU performance, less GPU Clock, ~5B transistors, 28nm

Good Job, NVIDIA. :)
 
GM204 165W TDP: less power, more gaming performance, more ROPs, more GPGPU performance, more GPU Clock, ~5B transistors, 28nm
Tonga 190W TDP: more power, less gaming performance, less ROPs, less GPGPU performance, less GPU Clock, ~5B transistors, 28nm

Good Job, NVIDIA. :)
Nice job indeed, but let's wait until Tonga shows its true face? It's still shrouded in mystery.

AMD should be thankful that: a) Nvidia will price it pretty high and b) didn't release this part in 2011. ;)
 
64 ROPs? I was arguing on here two, three, four weeks ago there was no way nvidia would go with 64 ROPs on a 256 bit bus. But now, this close to release, it seems legit.

Can't say it's legit, but I said it would make sense on a lot of fronts. Ok, the memory bandwidth is a bit limited for so many ROPs, but it is still more than twice the bandwidth of gtx 750ti (due to using faster memory). And here's the reasons why:
- gk104 had 4x8 pixel rasterizer, and 8x4 pixel shader export (actually 8x128bit IIRC). GM107 and presumably GM204 (assuming the SMM and rasterizers stay the same) will have 4x16 pixel rasterizer, 16x4 pixel shader export (16x128bit actually). Thus, it is easy to see why more ROPs may be useful from that point.
- Kepler and GM107 ROPs are quite simplistic in some way. Half rate fp16 blend, godawful slow fp32 blend for instance. These operations are actually not limited by memory bandwidth at all all on gm107 (unless looking at ddr3 versions) in synthetic tests, despite being very bandwidth hungry. The color compression helps (for fp16 at least, fp32 blend is never bandwidth limited on kepler without any compression in the first place). So, these sort of operations kinda "waste" memory bandwidth in some situations because the ROPs are too slow, which is not good. Nvidia could have improved the ROPs to make these things faster, but another way to achieve this is to just double the amount of ROPs (fp32 blend will go from godawful slow to very slow, but for fp16 blend it should be enough). If you compare that to Tonga, it still has 32 ROPs, but they can do fp16 blend at full rate (I don't know if AMD beefed up the depth/stencil testing too as the chip probably has 4x16 pixel rasterizers too though this is not confirmed - I don't know the shader export rate).
- I assume the internal (L2) bandwidth has also increased.
 
That's funny. When I saw an earlier videocardz post about gamer oriented features my mind immediately went to the B3D thread on ubersampling.

"Dynamic super resolution" sounds interesting but I wonder how it's any different from regular old super sampling.
 
That's funny. When I saw an earlier videocardz post about gamer oriented features my mind immediately went to the B3D thread on ubersampling.

"Dynamic super resolution" sounds interesting but I wonder how it's any different from regular old super sampling.

I doubt it's anything but dynamic super resolution. You set a target frame rate (or just vsync) and resolution is dynamically adjusted to maintain that rate. Would mesh well with g-sync (play Civ 5 at 4k at 40 fps at max available super sampling). The game itself can do that but, like alternative SSAO methods, it's nice if the driver can do it and can force it for games not implementing it.

e: hint hint AMD
 
that has to do with their cascaded display tech, i dont think its applicable to a normal display. maybe it supports it, but i wouldnt count on any consumer hardware using it during the cards lifetime.
 
Nice job indeed, but let's wait until Tonga shows its true face? It's still shrouded in mystery.

AMD should be thankful that: a) Nvidia will price it pretty high and b) didn't release this part in 2011. ;)

Nothing is really a mystery at this point.

AMD has already confirmed that tonga is hiding 4 Cu units which would make the full count of tonga 32cu or 2048. So expect a boost of 12 percent(things never perfectly scale with shaders unless they are small chips) + boost on clocks which isn't much unless they throw out power consumption, which they may be forced to do.

http://techreport.com/review/26997/amd-radeon-r9-285-graphics-card-reviewed/2

"AMD has confirmed to us that Tonga is indeed hiding four more compute units than are active in the R9 285, so the diagram above ought to be accurate in that regard"

I think the budget and R and D constraints are starting to show at AMD. I think they are going to be stuck on this GCN path for a while and it might hurt their competitiveness in the future. Particularly if Nvidia does Pascal late next year.

Nvidia outspends AMD R and D and for a company doing desktop CPU's and GPU, I think this shouldn't be the case.
 
Nice job indeed, but let's wait until Tonga shows its true face? It's still shrouded in mystery.

Full fat Tonga is going to perform around 7970GE / 280x speeds, maybe 5% faster on average unless there is a magically hidden 128-bits worth of memory controllers and 32 ROPs that has been disabled and not found by anyone examining the chip.
 
Full fat Tonga is going to perform around 7970GE / 280x speeds, maybe 5% faster on average unless there is a magically hidden 128-bits worth of memory controllers and 32 ROPs that has been disabled and not found by anyone examining the chip.

Well, outside some addition on the Hawaii cores ( mostly on features side like for the color compression ) and maybe some little changes here and there.

Tonga seems to have the same core design as Hawaii with half memory bus and Vram, ROPs count, let alone the TMU. I can imagine the shaders processors are exactly similar of what you can find on Hawaii.

On contrario maxwell is a new architecture: shader processors (CUDA Cores ) have got a lot of change it seems for allow better performance, and a lot of thing are redesigned. The layout of the SMM is really different of what you can find on kepler SMX.

Basically Maxwell is an evolution of Kepler who can be compared to the Fermi to kepler evolution. a single 128 CUDA core SMM can deliver 90% of the performance of a 192 CUDA core SMX at a much smaller size. ( A number who seems right in place with the 2048cores of the GTX980 given lately )

I dont see Tonga be an evolution of the Hawaii architecture, most likely a little Hawaii with some "features" added who was not ready when hawaii have been released. ( with the addition of the use of HPM process it seems ). For be honest i dont really see what Tonga can hide..

They are not even adressed at the same performance ratio, basically Tonga is the real successor of the 7870 in the lineup and take place under the R9 290.. I still dont understand why AMD have not release it at the same time of the R9 290x launch.

On TDP thats another question, but i will wait to see what the next AMD GPU arch will bring about it. ( in reality, i will better like a big gap in performance + new technology added vs a decrease on TDP, but thats another concern. )..

( understand me, on a techological point of view it is interessant to compare both, but not much that to compare the GK104-110 vs GM204 )
 
Last edited by a moderator:
Nothing is really a mystery at this point.
Sure there is a mystery: 700 million transistors + transistors from 128bit worth of memory controllers had to go somewhere, and the changes/additions weren't really that big to warrant that huge amount of transistors
 
Well, unfortunately for AMD, I can't see that it matters a whole lot where the transistors went. Even a fully enabled Tonga would likely trade blows or slightly edge out the 980 (or worst case slightly lose), meanwhile getting absolutely crushed on performance/watt, heat, and noise.
 
Well, unfortunately for AMD, I can't see that it matters a whole lot where the transistors went. Even a fully enabled Tonga would likely trade blows or slightly edge out the 980 (or worst case slightly lose), meanwhile getting absolutely crushed on performance/watt, heat, and noise.

Tonga would obviously lose to 980 no matter what secret sauce is inside, even Hawaii is having hard time with it by the looks of it.
 
Tonga would obviously lose to 980 no matter what secret sauce is inside, even Hawaii is having hard time with it by the looks of it.

Exactly, the 980 seems perform equal, maybe even a bit better of the 780TI.. i dont really see how a gpu who should perform between the R9 290 ( non x ) and 280-285 could come close of it. ( look like it is now confirmed that the 285 have only 1 SM disabled, so no magic secret on SM / SP count )

With the high boost speed reported for the GTX980 ( ~ 1200mhz vs 900-928mhz on the 780-780TI ) with the increase on performance / cores on Maxwell ( i should say performance by SM, it will be more accurate )... its not even really a surprise.
 
Last edited by a moderator:
With the high boost speed reported for the GTX980 ( ~ 1200mhz vs 900-928mhz on the 780-780TI ) with the increase on performance / cores on Maxwell ( i should say performance by SM, it will be more accurate )... its not even really a surprise.

I haven't seen 4k speeds quoted yet, and there is a bandwidth deficit. Might not help Tonga, but it likely gives the edge to the 780ti under some circumstances.

I'm interested that it has both hdmi 2.0 and three DPs (and love that at least two articles came out and all but said 'no, not 1.3, Dave'). I admit I had not expected that. Heck, I'm on record griping to ASUS about no hdmi 2.0 on the pb279q, pointing out the relative probability of multiple DPs vs hdmi2.0, and now feel fortunate that I didn't give voice to my internal doubts as to the marriage status of their progenitors ;/. If the pricing on the 970 is really $300, that's going to be pretty tempting. I've got a 5750 in need of upgrading, and while this'll be the most I've spent on a card, if 960 really hits at $250, $300 for the 970 sounds like a steal.
 
Back
Top