Guys over at Coolaler managed to (allegedly) get their hands on the upcoming Maxwell GM204 graphics card.
GeForce GTX 870 has 1664 CUDA cores?
First of all, we have (supposedly) the number of CUDAs this card might have in its final form. According to the leaker GeForce GTX 870 would sport 1664 CUDA cores (13 SMMs), 138 TMUs and 32 ROPs. We have to say, that numbers look poorly compared to full GK110 chip, but Maxwell may actually be that powerful.
Our job is not to confirm that these leaked specs are legit, our job is to confirm they are not.
To start with, device ID 13C2 has not been used before. The last entry is 13BB, which is yet unannounced Quadro based on GM107 chip. If NVIDIA was to launch a new GPU, that would be the ID to use.
138 TMUs and 32 ROPs?
We are going to make a side-by-side comparison to GM107. The layout could’ve changed, but the basic principles should not.
1664 CUDA cores mean there are 13 SMMs, since each SMM can hold 128 CUDAs. Each SMM has 8 TMUs, so the number of TMUs should actually be 104, not 138, 138 does not even divide by 8. This could either mean that someone was manipulating with the picture, or GPU-Z does not fully support second generation Maxwell yet.
In GM107 there were 16 ROPs per one GPC (which has 5 SMMs). The full GM204 could have 15 SMMs, thus 48 ROPs. Something is telling me that NVIDIA would not remove 16 ROPs from full chip (880).
Does NVIDIA D17U-20 actually exists?
The short answer: yes. If you look closely at the pictures, you will notice that the leaker didn’t uncheck the option to upload the scores to 3Dmark database. We actually did try to find them. Unfortunately our search did not bring any similar results.
However, we did find D17U-20.
As you can see it does confirm that N17U does exist, in fact, it even confirmed it has 4GB frame buffer. Core clock is not always read correctly by 3DMark (due to frequent profile changes), but memory clock is indeed 7GHz. Here is the link.
Bear in mind, it is rather easy to trick 3DMark to read anything you want.
GeForce GTX 870 specifications
Let’s move on to leaked specs. Along with 1664 CUDAs, we have 4GB GDDR5 memory and 256b interface. That’s exactly what we anticipated. NVIDIA would still use 7 GHz GDDR5 memory modules, but due to very short bus, bandwidth will not exceed 225 GB/s.
GeForce GTX 870 (allegedly) has a base clock of 1051 MHz and boost clock of 1178 MHz. We were told that GM204 cards indeed have higher clocks, so these numbers could be true.
|GeForce GTX 870 Specifications|
|GeForce GTX 780||GeForce GTX 870|
|Base Clock||863||1051 MHz|
|Boost Clock||902||1178 MHz|
|Memory Clock||1502||1753 MHz|
|Memory||3072 MB||4096 MB|
|Bandwidth||288 GB/s||244.4 GB/s|
|Launch Date||May 2013||September 2014|
GeForce GTX 870 scores P11919 points in 3DMark11
In performance preset the GTX 870 would be slower than GTX 780, however in Extreme profile (where 870 scores 4625 points), it is actually faster than 780 (that’s almost R9 290X score).
If these numbers are anywhere accurate, then we have a nice replacement for GTX 780 users. Due to very limited bus interface, Radeon 290X users will likely stay with their cards.
Last but not least, here is our finding. This is the score with affordable CPU (i5-4670), were we lose 138 points to i7-4820K. I think it is safe to assume that GTX 870 will sit between R9 290 and GTX 780.
As far synthetic scores can mean anything to gamers, GeForce GTX 870 could be 15-20% faster than GTX 770. I just don’t see how 256b can survive in UltraHD revolution.
Both GeForce GTX 870 and GTX 880 are expected in September.
TechPowerUP GPU-Z utility has now preliminary support for GM204 (source):
- Added preliminary support for NVIDIA GM204
- Added preliminary support for AMD Tonga
- Added support for AMD Radeon R9 M275X, FirePro W5100, W9100
- Added support NVIDIA GeForce GTX 780 6 GB, GTX 860M, GT 830M, GTX 780M, GT 740, GT 730, GT 720, Quadro NVS 510, FX 380M, GRID K520, Tesla K40c
- Added release date for R9 290
- More robust PhysX detection
- Fixed fan speed monitoring on some recent AMD cards
- Fix for sensor graph over/underflow
- Performance improvements to sensor graph drawing
- Fix for French translation