https://www.pcper.com/news/Graphics-Cards/NVIDIA-Announces-GP102-based-TITAN-X-3584-CUDA-cores
Almost 30% faster than a single GTX1080.
![titanxpascal2.jpg]( "titanxpascal2.jpg")
![Screen Shot 2016-07-21 at 9.23.01 PM.png]( "Screen Shot 2016-07-21 at 9.23.01 PM.png")
Almost 30% faster than a single GTX1080.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
NVIDIA Announces GP102-based TITAN X with 3,584 CUDA cores
- Thread starter SDAVE
- Start date
- Sort by reaction score
10 TFLOPs, 96 ROPs, 480 GB/s. I have no idea why it would be only around 30% faster than GTX 1080.
Secondly, if there is no error, Nvidia appears to be not able to get full die for GP102 working(6 GPC's equals 3840 CUDA cores, here are only 3584 - cut down part).
Secondly, if there is no error, Nvidia appears to be not able to get full die for GP102 working(6 GPC's equals 3840 CUDA cores, here are only 3584 - cut down part).
May be because the relatively low clock speed. Anyway, I just couldn't understand why they do that, because must keep it at or below 250W TDP? Is it the reference heatsink limit?
On the other hand GTX 1080 is only 21% faster than GTX 1070, I am talking about reference, Founders Edition. It has nothing to do with heatsink. The GPU will use around 240W of power during heavy load. GDDR5X uses less power than GDDR5, to the point that 12 GB GDDR5X will use 27W of power under load compared to 37W of 8 GB GDDR5 used in RX 480.
Secondly, in DX11 games the GPU may really be up to 35% faster because mostly it will be CPU count at this point(Static Scheduling...). IN DX12 scenarios the situation will be much different. Here comes raw compute power from the GPUs.
Fury X 8.6 TFLOPs
GTX 1080 - 9 TFLOPs theoretical.
Titan (P)X - 10.9 TFLOPs. The difference will be up to 30% over Fury X(knowing that currently GTX 1080 is up to 10% faster than Fiji XT).
Thirdly: Why Nvidia is rushing so hard their lineup?
I think it's about the price. There is no real competitor to the 1080 already. However, as you pointed out, the Fury X has more or less same raw power theoretically. Of course, Nvidia has no need to release the new titan to beat only themselves. However, because that's a card that have absolutely no competitor at this moment, that gives them the reason to sell a top of the line card at $1200
Nah, I don't think that is particularly the case. I think it has more to do with available slots in TSMC fabs. Remember that TSMC produces chips for 5 OEMs, and most ovailable space will go soon to... Apple and their A9X, A10 and potentially A10X.
Prices are higher because production costs are high, and the yields are not perfect on big chips, as you can see with GP102. You can get dies, they just need to be disabled.
None of Nvidia hardware was stacked in good quantity. They should have waited for enough available hardware, but they didn't. Whats more, there was quite a lot of previous generation of GPUs in the retail to consider waiting.
The difference: Nvidia sold 50 thousand GPUs arround the world. AMD have had to increase production at GloFo, because 95% of available supply from start was sold. AMD had 3 fabs at GloFo, and from single one they had 177 thousand GPUs at start. Overall AMD sold around 480 thousand GPUs to this day around the world.
Delays of AIB models were made for that very reason: unexpected demand for AMD GPUs, that even AMD themselves were not expecting.
The difference: Nvidia sold 50 thousand GPUs arround the world. AMD have had to increase production at GloFo, because 95% of available supply from start was sold. AMD had 3 fabs at GloFo, and from single one they had 177 thousand GPUs at start. Overall AMD sold around 480 thousand GPUs to this day around the world.
Delays of AIB models were made for that very reason: unexpected demand for AMD GPUs, that even AMD themselves were not expecting.
11 Tflops.
Isn't the 5,1 cMP PSU 1000 Watts? Should be able to get an adapter to work with this (after a firmware update that is)
Thats right, Nvidia seems to be inconsistent in boost vs base TFLOPS. ~10 TFLOPS is at the base clock rate just like ~8 TFLOPS is at the base clock rate on the GTX 1080. Some rough calculations puts this at something like a 480 mm2 chip. GP104 is 320 mm2 and GP100 is 600.
Nvidia has rolled out the Pascal generation in record time. They have delivered 4 Pascal GPUs in the span of 4 months. Apple could adopt these now across the entire lineup. Something like GP102 might be overkill in the current mac pro form factor though. Too big and hot to justify down clocking to get to the 125-150 W thermal/power threshold.
Mobile versions of 1070/1060:
http://www.pcper.com/news/Graphics-Cards/Report-NVIDIA-GeForce-GTX-1070M-and-1060M-Specs-Leaked
Apple uses mobile GPUs in iMac/MacBook Pro for discreet graphics and Workstation level GPU (which usually lags behind mainstream GPUs...they should stop this...and it costs A TON. Apple got a "deal" on the FirePro D700s, $600 or something ....but for $600 you can get a GTX1080 off the shelf).
They could if they wanted to, but they don't.
It took too much effort and resources to release country specific watch bands for the Olympics.
https://www.macrumors.com/2016/07/22/apple-watch-olympic-bands/
It is a good observation by Koyoot that this might be a cut down chip. Its interesting that they chose to do this as that doesn't give them much room to introduce a GTX 1080 Ti which is traditionally a cut version of a Titan (i.e. the 980 Ti is a cutdown version of the Titan X). I am sure they can keep something in their back pocket for whenever AMD releases Vega. Then they can adjust the lineup to compete with it.
Apple has to design a custom logic board for all of its products which is why it might lag behind the add in boards that are already out there. The chips themselves are just down clocked versions of GP104 and GP106.
Apple has to design a custom logic board for all of its products which is why it might lag behind the add in boards that are already out there. The chips themselves are just down clocked versions of GP104 and GP106.
It does give them room to introduce 3200 CC chip as GTX 1080 Ti. And that is most probable core count.
I don't think that GP102 would be stable at that TDP, even considering lower clocks and voltages.
980w, but only 2x6-pin available for GPUs, without mods or an additional PSU.
What firmware update?
When I had a 5,1 Mac Pro, I had a GTX780 with an EFI Mac Firmware, which allowed the usage of 2 pins of power without any mods, and also using PCIe 2x lanes.
I meant Firmware update for the Titan X (ie what MacVidCards does).
Do you know the wattage that comes out of those 2 pins from the 5,1 cMP? If enough, then maybe someone can make a 4+4pin adapter.
Last edited: