NVIDIA Maxwell Speculation Thread

[A1xLLcqAgt0qc2RyMz0y]

Engineering sample Maxwell PCB Power

How about because this being an engineering sample pcb the power supplied could be either one 8-pin or two 6-pin.

 

[Wynix]

That looks like it could be roughly the size of AMD's Tahiti chip, maybe slightly bigger?

 

[xDxD]

Can we expect ~ 15% more performance (overall) between gm204 and gk110 (or ~50% more between gm204 and gk104) ? I think it is realistic, don't you?

 

[boxleitnerb]

Depends on which GK104 SKU you take as comparison. The 780 Ti is almost 50% faster than the GTX 680, but only about 45% faster than the GTX 770.
I think 60-70% faster than the GTX 680 is to be expected.

 

[Deleted member 13524]

Depends on which GK104 SKU you take as comparison. The 780 Ti is almost 50% faster than the GTX 680, but only about 45% faster than the GTX 770.
I think 60-70% faster than the GTX 680 is to be expected.

780Ti is a GK110, not GK104.

 

[Picao84]

780Ti is a GK110, not GK104.

On the other hand (which is what he meant) GTX 770 was GK104.

 

[boxleitnerb]

780Ti is a GK110, not GK104.

So? xDxD mentioned GK110 in his post.

 

[Deleted member 13524]

Ok sorry, translation mistake on my part.

 

[Esrever]

Looks too small to replace the 780ti. Maxwell doesn't increase performance/mm^2 that much by looking at the 750ti. I wonder if its real and if it really will be named the 880.

 

[boxleitnerb]

Looks too small to replace the 780ti. Maxwell doesn't increase performance/mm^2 that much by looking at the 750ti. I wonder if its real and if it really will be named the 880.

75 percent faster than GK107 and 25 percent larger...40 percent better perf/mm2

 

[Esrever]

75 percent faster than GK107 and 25 percent larger...40 percent better perf/mm2

Oh I was looking at the 650ti numbers when I was comparing it to kepler, I guess this will be slightly faster than the gk110 if those are the gains.

 

[McHuj]

I'm hoping that something like the 870 can be a bug upgrade (+75-100%) over my 670 and will come with 4GB of VRAM.

 

[Psycho]

75 percent faster than GK107 and 25 percent larger...40 percent better perf/mm2

GK107 was the very early and first 28nm chip, so probably quite conservative - GK110 wasn't widely/comercially available until a year later. Plus those small chips probably spend a considerable percentage of space on stuff that just has to be there, unrelated to performance. So GK107 is very favorable to compare against.

GK106 is 45% larger (for 35% more transistors), and 660 is ~25% faster than 750ti. So in that comparison GM is around 15% more area efficient than GK.

 

[tviceman]

GK107 was the very early and first 28nm chip, so probably quite conservative - GK110 wasn't widely/comercially available until a year later. Plus those small chips probably spend a considerable percentage of space on stuff that just has to be there, unrelated to performance. So GK107 is very favorable to compare against.

GK106 is 45% larger (for 35% more transistors), and 660 is ~25% faster than 750ti. So in that comparison GM is around 15% more area efficient than GK.

Yet gm107 didn't replace gk106 in nvidia's product stack, nor will it be positioned similarly in maxwell's family of chips, making that comparison less valid and useful.

 

[homerdog]

So this card should be extremely fast then. Hard to imagine a chip that much smaller than GK110 and that much faster on the same node. Reminds me of RV670 vs RV770.

 

[Psycho]

Yet gm107 didn't replace gk106 in nvidia's product stack, nor will it be positioned similarly in maxwell's family of chips, making that comparison less valid and useful.

What does this suddently have to do with the product stack!?
We were investigating maxwell's area efficiency gains in an attempt to predict the performance of a chip 25%(*) smaller than GK110. And I was merely pointing out that it may be a good idea to use more than a single (that may not be the best) sample point for that.

(*) The above gif anim makes gm204 76% of gk110 or ~421mm2 (pretty close to hawaii), while gk104 for verification is the expected 53%/293mm2. (actually i would expect the pixel measurements to show too little difference between the chips - as we are measuring the outer packaging)

 

[Homeles]

What does this suddently have to do with the product stack!?
We were investigating maxwell's area efficiency gains in an attempt to predict the performance of a chip 25%(*) smaller than GK110. And I was merely pointing out that it may be a good idea to use more than a single (that may not be the best) sample point for that.

(*) The above gif anim makes gm204 76% of gk110 or ~421mm2 (pretty close to hawaii), while gk104 for verification is the expected 53%/293mm2. (actually i would expect the pixel measurements to show too little difference between the chips - as we are measuring the outer packaging)

There is a baseline level of circuitry that these GPUs use for things like PCI-E and display output that have nothing to do with the GPU's performance. A larger GPU will have higher areal efficiency than a smaller one from the same family, unless design decisions are made that would significantly impact areal efficiency (e.g. the large, fast bus on Tahiti, or the additional FP64 hardware on GK110).

 

[tviceman]

What does this suddently have to do with the product stack!?

It's not a direct transitional comparison between GM107 and GK106, unless you're doing it just for the sake of discussion. GM107 replaced GK107, while retaining the same bus and ROP configuration. It's the closest, most valid comparison to make when extrapolating how GM204 might perform compared to GK104.

We were investigating maxwell's area efficiency gains in an attempt to predict the performance of a chip 25%(*) smaller than GK110. And I was merely pointing out that it may be a good idea to use more than a single (that may not be the best) sample point for that.

Using GK110 when comparing die sizes and extrapolating performance is an even less valid comparison. GK110 has a significant amount of compute-dedicated transistors that don't exist on GK104, 106, 107, and GM107. All of those chips (except GK107) beat out GK110 on perf/mm^2 by different amounts (i.e. gtx 780 TI is 45% faster than gtx 770 but has 89% larger die, gtx780 ti is 122% faster than gtx660 but has 157% larger die).

As boxleitnerb pointed out, GM107 is 25% larger than GK107 but outpaces it by 70-75% with the same power consumption. Kepler scaled linearly among all it's chips in perf/watt (GK104 and GK107 had roughly the same perf/watt), and I don't see why that metric won't continue with Maxwell. If this new chip has the same operating TDP of GK104, then it should roughly 70-75% faster than gtx770 (about 20% faster than gtx780 TI) assuming it's a balanced chip with respect to cache, ROP's, bandwidth, and cuda cores. If Nvidia engineered it to run at a higher TDP, then it should increase the performance gap

 

[xDxD]

videocardz.com/51009/nvidia-preparing-four-maxwell-gm204-skus

I like this part, however: Bear in mind that Charlie’s website is called semi accurate for a reason. It could all be wrong from the beginning.

 

[McHuj]

I'm not sure I believe that they'll skip 20 nm completely. For consumer level products, I think it's believable and at this point likely.

But for the high end, low volume where cost isn't a factor, I bet there will be a 20 nm GPU that will show up as a Tesla product and then maybe filter to a consumer GPU as a GTX Titan II or something similar with an outrageous price tag.

However, I don't think I buy the time table for the 16nm refresh. I think the earliest would be Q4 2015 and at that point they'll probably call it the 9xx series.