Nvidia explains the GTX 970's memory 'problems'

[YazX]

Guys, Now Nvidia is trying to cover this up, they hid the main thread of this issue on their forums, if you want to access it, here is the link:

https://forums.geforce.com/default/topic/803518/geforce-900-series/gtx-970-3-5gb-vram-issue/98/

[HalloweenJack]

http://techreport.com/blog/27143/her...an-the-gtx-980

the wierdness was first talked about in October of last year -but where techreport was slightly wrong , the disabled SM`s actually had an effect on ROP count

[herulach]

The problem isn't memory clock speed, it's down to functional units on the processor itself being disabled, limiting throughput.

Often with processors, functional units are arranged in to blocks which can only be disabled with limited granularity, so in order to disable shader cores, it might include other parts. I'm not certain that's fundamental to this issue, but it's not just a case of using different memory ICs.


Caches are controlled by the processor itself and transparent to software. Memory OTOH is software-allocated, and if some software is allocating the bad memory areas it can cause problems, as is happening.

I know that, but I'd be surprised (although I don't know much about 3d programming) if game designers are controlling memory allocation. It seems like something which could be done in the driver, at least after a learning period (similar to the logic used by SSHDs).

Just to be clear I'm not defending not telling anyone about it, but, from a design perspective I can see the logic, 4GB markets better than 3.5 and 4 should perform better than 3.5 even if the last 512 is slower (its still faster than disk buy orders of magnitude).

[HalloweenJack]

the GTX 970m wont have this issue -0 they fused off the entire block ; how did they `make` the GTX980m though? its supposed to be 256bit

[watercooled]

I said software, not specifically games; I'm not certain how memory allocation works with GPUs but I suspect it's abstracted by the driver. Nonetheless, in order to prioritise memory in a way which is effective for mitigating this issue, the driver would need to be aware of the purpose/bandwidth demands of what's stored in a given (virtual) address, which may not always be possible.

What I meant was it's not really comparable to caching which is accounted for. This is something which, in lieu of mitigations in software, really throws a spanner in the works.

More, slower, memory isn't necessarily useful unless it's utilised correctly, similar to how certain VM software treats SMT cores as two real cores and can cause a CPU storm, destroying performance. In these cases, it's often better to ensure only 'real' cores are used and just accept you're not getting the potential performance of SMT. In some cases, more !=better. At the moment it seems simply (and perhaps temporarily) disabling the last 512MB of VRAM would improve everything.

@HalloweenJack: Having read into it a bit more, it seems the SMM and ROP units are fuse-able independently, and the 980M has all ROPs enabled. Similarly, AFAICT everything besides the 970 have even numbers of memory controllers; I'm not sure if this implies they're all full ROPs, but if so they should also be free of the issue as the bottleneck wouldn't be present there. However if they're using multiple part-disabled units then it could be an even more severe problem. Something to watch for in the latter case, is like previous parts we might see multiple die configurations on the market, so some have the issue while others are free of it.

[Moogly]

I may not be as technically savvy as the other commenters here but I can certainly see that Nvidia purposefully spread misinformation about the cards specs. I think that is the issue at hand.

While it does not impact the fact that the card is still a beast, it erodes the trust consumers have in their company.

I don't want to be that guy but I'm hoping AMD seriously brings it this year.

[watercooled]

While we've seen 'official' FPS numbers playing down the issue, they don't really tell us anything about stutter which wouldn't show up (as Nvidia were quick enough to point out about old Xfire drivers OFC): http://www.extremetech.com/extreme/1...mory-problem/2

Seems that while FPS doesn't take a major hit, it can really hurt frame time, as expected TBH. Having said that, what's happened to all of the frame time benchmarks? They were used extensively over the past few years but seem to have dried up with the recent Nvidia releases?

Edit: At the end of the day this could just be a bad driver not segmenting the memory properly as the software sees it; some amount of segmenting does appear to be going on already if the GPU-Z screenshots of ~3.5GB utilisation are anything to go on. However it still stinks as far as marketing goes - while it technically meets the bandwidth/throughput specs, it's not really able to achieve them with actual data. Much like the FLOPS numbers given for the Tegra X1; they forgot to make it clear that they were comparing FP16 numbers against FP32 numbers making the comparison even more useless than cross-architecture FLOPS comparisons already are...

[CAT-THE-FIFTH]

http://www.anandtech.com/show/8935/g...ory-allocation

[HalloweenJack]

'If this was a "miscommunication" between marketing and engineering how does the fact that GPU-z reports 64 ROPS for this card, did marketing write the BIOS for the cards '

[tob5id1]

The point everyone seems to be forgetting is that the cards price/performance/power consumption are still great, nothing there has changed.

You still didnt get what you paid for.

[KeyboardDemon]

Here's my 2 cents worth.

I'll talk about Performance first, that's the thing that matters most to me when I am looking at buying a GPU and I can see that with or without this 'new' knowledge, if I were looking at getting a GTX970 now, I would be looking at benchmarks and making a quality assessment based on what other cards at or around the same price or less could achieve the same results or better, or if spending a little more would gain me far better results. Ultimately, I would be looking at the benchmarks I know, trust and understand, so those benchmarks that can be run in games, Unigine and 3d Mark are usually my choices to trust because I can repeat those tests using my current hardware and make assumptions based on the results I get compared to the results that the reviewers get.

You still didnt get what you paid for.

So in the case of the GTX970, based on results then I would argue that you did get what you paid for.

However, I would also argue that if nVidia were issuing press releases that they either knew were misleading and then keeping quiet about it or if they later found out that the press releases were wrong and did nothing to reveal this themselves then that just shouts of dishonesty and bad practice. I think things like this are damaging to nVidia's image as company, I think that if they knew or found out about this then they should have come clean rather than waiting for someone else to reveal this or hope that no one discovers it.

If I were a potential customer looking to buy a GPU and read about this now my immediate feeling would be that nVidia have lied about the specs of the GTX970, if they did it with the GTX970 then what else have they lied about? If I can't trust the information that was released about one card why should I trust the information about their other products?

[crossy]

To a lesser extent, it's also hard to swallow that no-one foresaw this at any stage of the processor design; the block fusing as I said tends to be fairly fixed (and is part of the design), hence it would have been known that fusing cores to make the 970 would take part of the memory subsystem with it. Even if it's possible to mitigate it to extent in software/firmware, why did it make it out without that in place?

Oh great, I was feeling all happy with life and then I read this and thought "yep, that's it, I'm now officially dumb enough to apply for the position of pointy-haired boss ... or something in the marketing department". Grr! (Thanks for the info though - got a Powerpoint presentation to go along with it? LOL)

Self-flagellation aside, it's strange that it's a 3.5GB : 0.5GB split - the last time I saw that was on the Dell D620 I bought from eBay. Despite having 4GB installed, I could only ever use 3.5GB, wonder if it's the same kind of addressing deal here?

It would have been better, I think, for NVidia to admit this to start with - since it's only the attempt at concealment that I object to. If I had a 970 then I think I'd probably still be okay with it - the performance is still going to be pretty good for the price. TBH the only folks who are really going to hate this are all those smart-backsides who were saying to buy a 970 as a cut-price 980. So if you absolutely really need high-tex 4K then it's going to have to be a 980. Meanwhile, I'll continue to "slum it" avec mon 7970.

[shaithis]

Does it matter?

Surely if it did, reviewers would have picked up on it and benchmarks wouldn't have been as good as they were.

Seems the only people with egg on their faces are the reviewers who feel slighted that their reviews were inaccurate with regards to technical data....and then didn't pickup on the drop-off from games utilising > 3.5GB VRAM.

[[DW]Cougho]

Does it matter?

It doesn't matter from a performance point, I agree.

However a blatant attempt to mislead customers matters a lot in my book. They supplied incorrect specifications to review sites, lets give them the benefit of the doubt and say this was a mistake - however someone in nVidia would have noticed this, and they didn't come clean until they were found out.

This news shouldn't turn people against the 970, but it should make people wary of nVidia.

[CAT-THE-FIFTH]

but it should make people wary of nVidia.

Cough,bumpgate?

[[DW]Cougho]

Cough,bumpgate?

Sorry of course that should read 'even more wary', you should be wary of every company by default (can you tell I'm a cynical person?).