The GTX 970 has 4 GB RAM, but if you use more than 3.5 GB, expect a performance drop.
What happened?
With the GM204 GPU, Nvidia decided to launch two products: A full-fledged graphics card, called GTX 980. And a somewhat cut-down/salvage part, the GTX 970.
Nvidia designed the GTX 970 in a way to make sure that the card cannot be fast enough to compete with the GTX 980. They deactive 3 out of 16 shader clusters. This was part of the official specs.
But they also reworked the memory subsystem. The GTX 980 has 8 memory chips of 0.5 GB capacity each. Each memory chip is controlled by its 32 bit GDDR5 controller, and each controller gets its data from/through an L2 cache block of 256 KB.
GPUs store they data interleaved over all RAM chips. This allows to utilize all memory controllers at the same time for memory transactions.
Now we come to the ugly part.
The GTX 970 has only 7 active L2 caches, but still 8 memory chips. The L2 cutdown efficiently reduced the memory bandwidth from 256 to 224 bit (instead of 8x 32 bit, only 7x 32 bit.)
It also renders the last memory chip almost unusable. Only 32 bit bandwidth is available to access it, and if it is accessed, the 3.5 GB block cannot be used at the same time.
Nvidia did not made this information available until pressed by the media who reported the uproar in the forums. The GTX 970 still offers good bang for the buck, but one has to consider that even though it can store 4 GB of RAM, only the 3.5 GB are fast enough for gaming purposes.
The point of the story is not the cut-down of the memory controller, but Nvidia's information policy.
I talked about this some in the build resource thread
and i uploaded a video showing another example (this stuff has been posted before in several games) of the GPU refusing to allocate over 3.5GB of VRAM (3584MB) even when needed, causing load times 3-5x longer than expected, freezing, stuttering (small and big stutters) and hard crashing of the games that need that VRAM (memory usage bouncing up and down at ~3.4 - 3.48GB as stuff is reallocated within the actual usable memory) - which then show no problems at all if you turn down a few settings so that they don't hit ~3.45GB of VRAM used.
For that video, the game was in windowed fullscreen, so ~308MB of VRAM was used by desktop. With the game in fullscreen, the issues were a lot less aparrant, but still there. 3.5GB isn't enough for maxing that game on 1080p - the official recommendation is that 4GB isn't enough, either - but that 14% extra VRAM goes a long way to making sure that you're fine almost all of the time, and you can recover a fair amount more by disabling or turning down a few other settings that let you keep Ultra textures with the true 4GB cards, but not the 3.5GB.
Also, the loss of memory bandwidth from being effectively 224 bit instead of 256 bit does hurt performance. The 980 has that ~14% more memory bandwidth. If i increase my memory clock speed enough to give that much of a bandwidth difference, FPS increases by ~5-7%. That's one of the big reasons why 980, relative to the 970 specifications posted in reviews, performed shockingly well; with overclocked Maxwell being as memory bound as it is, a card with the same amount of memory bandwidth, same ROP count etc shouldn't be able to get a 20% performance increase by adding 23% more SMM.
On paper, the GTX 970 ought to be nearly as fast on this front as the 980—and the Asus Strix card ought to be a smidgen faster. The 3DMark color fill test we use has evidently been limited by memory bandwidth at times in the past, but that shouldn't be an issue since all three cards in question have the exact same memory config.
The 3DMark color fill test we use has evidently been limited by memory bandwidth at times in the past, but that shouldn't be an issue since all three cards in question have the exact same memory config.
has evidently been limited by memory bandwidth
shouldn't be an issue since all three cards in question have the exact same memory config
Ouch.
While the 970 is limited to effectively using 52 ROP's for some things because of the SMM count -
Extra ROPs are still useful to get better efficiency with MSAA and so. But they don’t participate in the peak pixel fillrate.
- Those extra ROP's are disabled in hardware, there are only 56.
The 290 to 290x have the same amount of ROP's, same memory bandwidth etc - a 10% increase in cores there translates only into a ~2-5% increase in performance, at the same clocks in typical games
So what's the optimal course then for someone like me. I was looking to replace my rather old radeon 6950 and was waiting for the 960 release. But the 960 looks rather disappointing and now the 970 doesnt look all too well, too. The 980 is way out of my price range though. Should i just get the 970 anyway, or is there any chance that they will release a new version of the 970 with full 4gb if the backlash gets big enough?
They probably won't release 4GB version because they'd need to keep the last section of l2/rops enabled and that would make all chips with any of the 8 sections unusable to be worthless, instead of sold as a 970. There should be a ~960ti, but it probably won't come until AMD releases next gen stuff so if you're buying now and you want a highish end GPU, just get 970 or 290. If not, a 280 or 960 will be ok.
3.5GB is still a lot of VRAM - 780ti only has 3GB - but it's annoying for those people like me, who would have paid the extra for 980 if the shortcomings of 970 were known relative to it
This is such bullshit, so fucking disappointed in Nvidia. I picked 970 because I wanted a card with 4gb vram and a fast clock speed and I found the 970 to be the best bang for the buck, only it does not fucking have 4gb of effective vram turns out. If amd cards werent so fucking power hungry and not nearly as well supported I would switch next time in spite.
On February 03 2015 12:30 JacobShock wrote: This is such bullshit, so fucking disappointed in Nvidia. I picked 970 because I wanted a card with 4gb vram and a fast clock speed and I found the 970 to be the best bang for the buck, only it does not fucking have 4gb of effective vram turns out. If amd cards werent so fucking power hungry and not nearly as well supported I would switch next time in spite.
You can probably get a refund for your 970, and AMD is offering 10% discounts on some 290's from some retailers - just email roy.taylor@amd.com. Some features aside, that puts 970 at something like 1.4x the GBP/euro cost per FPS.
yeah that's pretty fucked up. was about to get me a GTX970 for the Witcher 3 release, but that's not gonna happen - maybe if there is a juicy discount on the card which knowing nvidia is not very likely.
EPT
