NVIDIA Maxwell Speculation Thread

Not really. Announced at GTC. Only useful for some very specific compute use cases. Priced for no volume at all.

And if the gossip rumors are true, it was planned to be released earlier and then delayed to to issues. I can't be bothered to look up exactly how long it was delayed due to this, but shit happens in engineering.

But feel free to disagree, and express outrage if you prefer. It's of no importance.

I think it was originally meant to appeal to consumers, but plans changed when the 295 X2 came out.

(Not that I understand what the problem is with paper launches either.)
 
I think it was originally meant to appeal to consumers, but plans changed when the 295 X2 came out.

(Not that I understand what the problem is with paper launches either.)

Yeah remember this GTC, Nvidia was present Gameworks, G-sync if im right.. teasing a future product was not a bad idea. AMD had a lot of attention, Mantle or low API overhead was in all discussion and nearly all presentation ( even if not speaking directly about Mantle ).

Then the 295x2 is released.. I dont know if the problem was to try got higher clock of what was initially decided, but well the Z have been delayed for nearly 2 months.

Anyway, im not sure paperlaunch today are really a problem.. when it is a question of week.. seriously, whatever is the brand you have allways a delay between country. here, outside some limited pieces, at a prohibited price, we need in general wait 2 weeks
in addition of other EU country for get the gpu's in shop at standard prices ).. So paper launch or not, we will not get the cards in shop the same days.

On the other hand, in the case of a paper launch with a product who dont exist, is not finalized, and where you provide tons of benchmarks etc ... that could be more another problems.

But in the case of the presentation of a new architectures, showing the cards but saying they will be available from 2 weeks to 3 weeks from now, i dont think it is a real problems, reviews will take place at this time.
 
What I remember most about this year's GTC is that Volta disappeared off Nv's roadmap and Pascal was announced to be coming after Maxwell.
Yeah, that's another thing I don't understand: the fascination with product names of which we don't know anything in the first place.

Take Volta: announced as having HBM DRAM. That's really all we knew about it. Now it's Pascal that will have HBM and it's occupying exactly the same space as Volta in that slide.

Does it matter at all? Are we any wiser? Other than that the generation after Pascal may or may not be called Volta? I'm sure there is a reason why it's called Pascal instead of Volta, but from the outside it's nothing more than a renaming of the same thing.
 
Just how much more performance can nvidia get out of 28nm without making unrealistically huge chips? It seems to me that they can only adjust the archtecture so much at 28nm before hitting a wall. If I were nvidia, I'd consider holding Maxwell back until 16nm was available.
 
Just how much more performance can nvidia get out of 28nm without making unrealistically huge chips? It seems to me that they can only adjust the archtecture so much at 28nm before hitting a wall. If I were nvidia, I'd consider holding Maxwell back until 16nm was available.
There's no reason for bigger Maxwells not to have similar perf/w and perf/mm2 improvements that were seen for GTX750Ti. That should be sufficient to lift them quite a bit beyond current performance levels. 16nm will give them yet another lift, but that's still at least a year away, and probably more. There's no good reason to wait for that.
 
13 SMX salvage part?

Looks like the full config for the 256-bit chip would be 16 SMX. :???:

p.s.: the # of TMUs doesn't match the SMX's.
 
15 SMM may be a reasonable possibility since they could be going for 3x GM107 in terms of SMM count.

p.s.: the # of TMUs doesn't match the SMX's.
138 = 2·3·23 doesn't work well for basically any number of SMMs anywhere near 15, so I don't think that number is correct at all (as opposed to the TMU count being correct and the shader count being incorrect).
 
15 SMM may be a reasonable possibility since they could be going for 3x GM107 in terms of SMM count.

138 = 2·3·23 doesn't work well for basically any number of SMMs anywhere near 15, so I don't think that number is correct at all (as opposed to the TMU count being correct and the shader count being incorrect).

Or it could be that GPUz isn't reading the core count correctly.

GK104 was 4x the cores of GK107. GF/104GF114 was 4x the cores of GF107. I think when it's all said and done, a fully enabled GM204 chip will have 2,560 maxwell cores.
 
15 SMM may be a reasonable possibility since they could be going for 3x GM107 in terms of SMM count.

138 = 2·3·23 doesn't work well for basically any number of SMMs anywhere near 15, so I don't think that number is correct at all (as opposed to the TMU count being correct and the shader count being incorrect).

138 = 13(SMM) * 16(TMU/SMM) * 2/3(maybe some kepler/maxwell issue)
IMO GPU-z got 13 CU and 128 ALU per CU from OpenCL, and then used 192ALU : 16TMU ratio to get TMU count :)
 
Or it could be that GPUz isn't reading the core count correctly.

GK104 was 4x the cores of GK107. GF/104GF114 was 4x the cores of GF107. I think when it's all said and done, a fully enabled GM204 chip will have 2,560 maxwell cores.

With Kepler they didn't have a theoretical upper boundary of how many transistors they can squeeze into 28nm to reach the 551mm2 of the GK110 either. On the other hand GM200 is bound by that and there has to be a reasonable performance difference for GM200 based SKUs (salvage parts included) above any GM204 based SKUs.

I don't think the rumors that the GM204 based GPUs will have a lower MSRP is a coincidence either and no I don't expect the biggest 204 based SKU to exceed a GK110b by ~20% in performance as you're suggesting. It sounds more like give or take the same performance ballpark with vendor specific higher clocked variants to reach that much higher than a 780Ti.
 
The future-nvidia-gpu speculative parts in the paper don't seem to contain any insider info, just conjecture based on the GTC slideware...
 
With Kepler they didn't have a theoretical upper boundary of how many transistors they can squeeze into 28nm to reach the 551mm2 of the GK110 either. On the other hand GM200 is bound by that and there has to be a reasonable performance difference for GM200 based SKUs (salvage parts included) above any GM204 based SKUs.

I get what you are saying, and I don't disagree at all. But given Maxwell's efficiency, and assuming that efficiency is maintained as performance scales up with bigger chips, there is an awful lot of headroom left to work with even within the die size / transistor density constraints of 28nm. Case in point, GM107 is obviously headroom capped purposefully by Nvidia. They could have easily slapped it with 7ghz vram and shipped out 1275mhz boost clocked chips, attaining or surpassing GTX 660 performance. (And I believe they may still do that in a future 800 series rebadge of GM107).
 
Back
Top