dmcnutt
No crashes in games but get significant drops in fps during shadow of mordar benchmark.

Scarlet-Tech

dmcnutt
Mine new SSC clocks to 1585, I was just concerned because it kept failing on the memory test. I love the card and don't want to RMA. I just thought this might be an issue because of this memory bug everyone is talking about. Which I would not have even known existed because gaming performance is amazing, but it fails this test. I thought maybe something was wrong with the card. Memory won't go above 3.5 in games though.

Nothing is wrong with your card. That test is failing on other cards, and that was what it was designed to do.

Sajin
Um... what? The test works just fine on my cards when tested properly.

The benchmark measures "actually" not the DRAM bandwidth but the bandwidth of the "global memory" s in CUDA. The global memory in CUDA is a virtual memory space, which includes not only the DRAM, the GPU but also DRAM areas of the CPU.

vorko
Everyone still relying on Nai's benchmark as any sort of indicator should read from Nai himself (orig. in German - the link is to google translation to English)
 

The benchmark measures "actually" not the DRAM bandwidth but the bandwidth of the "global memory" s in CUDA. The global memory in CUDA is a virtual memory space, which includes not only the DRAM, the GPU but also DRAM areas of the CPU.

 
Yes, well that is not a perfect translation but it's just as some of us suspected. The bench says nothing about the 0.5MB hardware-wise of bandwidth-wise because at that point it starts to use system resources. So either the 0.5MB module is inaccessible because of faulty hardware or it is a form of a driver bug (in mem controller) which should be fixable... still doesn't change anything about the point that we need to wait for an official word from nVidia

RaghuC99

RainStryke

vorko
Everyone still relying on Nai's benchmark as any sort of indicator should read from Nai himself (orig. in German - the link is to google translation to English)
 

The benchmark measures "actually" not the DRAM bandwidth but the bandwidth of the "global memory" s in CUDA. The global memory in CUDA is a virtual memory space, which includes not only the DRAM, the GPU but also DRAM areas of the CPU.

 
Yes, well that is not a perfect translation but it's just as some of us suspected. The bench says nothing about the 0.5MB hardware-wise of bandwidth-wise because at that point it starts to use system resources. So either the 0.5MB module is inaccessible because of faulty hardware or it is a form of a driver bug (in mem controller) which should be fixable... still doesn't change anything about the point that we need to wait for an official word from nVidia

I've seen people quote his forum posts as some kind of end to the argument because of what they gathered from the translation. If you sit there and read it with the perception of him saying that the benchmark is flawed, you will see what you want to see. If you re-read it with an open mind, you will notice the way google translates his post into English is very disorienting to read. Certain words are off and make it really hard to understand what he's actually saying.

Using his quotes as the end of the argument isn't credible? In that case, why don't you assume what his benchmark says isn't credible? It goes both ways

vorko
Well I don't know about you, but I speak enough German to understand what he's saying. I used google translate only to point the post out to others. Here's the original text:
 
Mein Benchmark sagt nicht aus, was ihr glaubt, dass es aussagt. Es eignet sich *NICHT* um das Vorhandensein eines VRAM-Fehlers nachzuweisen.
 
and here's the translation:
 
My benchmark doesn't say what you think (it does). It is not suitable for VRAM error detection
 
The other post is much longer and quite technical, but pretty much confirms that his tools doesn't use the 2nd VRAM module correctly because it seems to be working with system resources instead of GPU despite directly requesting GPU VRAM addresses

RainStryke
I'm not trying to sway opinion.
 

vorko
Well I don't know about you, but I speak enough German to understand what he's saying. I used google translate only to point the post out to others. Here's the original text:
 
Mein Benchmark sagt nicht aus, was ihr glaubt, dass es aussagt. Es eignet sich *NICHT* um das Vorhandensein eines VRAM-Fehlers nachzuweisen.
 
and here's the translation:
 
My benchmark doesn't say what you think (it does). It is not suitable for VRAM error detection
 
The other post is much longer and quite technical, but pretty much confirms that his tools doesn't use the 2nd VRAM module correctly because it seems to be working with system resources instead of GPU despite directly requesting GPU VRAM addresses

Yeah, I saw that in the more recent postings. But it's clearly not testing motherboard RAM because GTX 970 users see 3.5GB with 8GB of system RAM and GTX Titan users see up to 5.5GB of RAM with even more system RAM on that test. I didn't assume it was error detection. I looked at it like a bandwidth test of the packets it's processing all the way up to the full capacity of the GPU's dedicated RAM. I'm not sure what he intended it to be... But it did get Nvidia to respond and admit that the last .5GB of the VRAM is partitioned and ran differently than the GTX 980's. There is an issue here... We just don't know the true technical reason for it. Some don't have the issue of stuttering, but many do once they reach that VRAM limit. The FPS rate looks fantastic on paper, but it doesn't stop the user from getting stuttering and inconsistent frame rate latency.

billythekid2014
NVIDIA will not do a darn thing  if they were going to they would not of said what they did about the vrm
in there last statement.
So just enjoy your cards and have fun