DegustatoR
Legend
Defects variation.
Not a good look all in all considering that it does affect the performance in comparison to what's shown in reviews.
Nvidia GeForce RTX 50-series Blackwell reviews
- Thread starter Scott_Arm
- Start date
Not a good look all in all considering that it does affect the performance in comparison to what's shown in reviews.
I guess there's minor comfort in the fact that Techpowerup's review sample from Zotac was affected. At least we can nip conspiracy theories about bait and switch in the bud.
The couple dozen people who managed to snag a 5090 might have a class action on their hands
It doesn't have to be a nefarious conspiracy to warrant legal action.They could have simply advertised it as having 168 ROPs, sent 168 ROP cards to reviewers, and either made sure all retail models had 168 ROPs or thrown in a few bonus ROPs on certain cards. Even if the variation was never mentioned, very few users would have complained about getting a faster card than was advertised. At the very least all they had to do was say that ROPs can vary between 168-176. It would be dumb but not this dumb.
Was the ROP count ever mentioned in any official NVIDIA documentation or marketing materials?
Driver only provides the path to read it from GPU. Also it wouldn't affect just random cards.
At least it's in whitepaper, 176 ROPs
arandomguy
Veteran
Would be interesting if some more low level tests could be run on the units to compare. I wonder if the ROP difference also means there's some other differences elsewhere (such as with cache or VRAM access).
Otherwise it's a bit surprising that the ROP difference only could have such an effect on some games.
The current rumor seems say it's not intended and I'd guess they would just replace the affected units especially given there is so few out there. Could also be a factor in the supply issue and the reported ramp up.
What could be interesting with all this is should there be replacements this would be likely rather straight forward for first party purchases. But with scalping and second hand purchases it could get quite interesting. There's a side of me that would actually like it if there were "landmines" like this that discourages scalping. The best case scenario would be at least those who sold on ebay have to take returns (or face buyer protection clawback), wait for the replacement and by the time they get it find the resell market has dried up.
Are we absolutely sure the drop in support is purely due to deprecating things on the software side?
Given that Blackwell is seemingly extracting more performance and functionality out of seemingly the same die/transistor cost, I would wonder if there are changes at the sub SM level that break compatibility with 32 bit CUDA?
Otherwise it's a bit surprising that the ROP difference only could have such an effect on some games.
The couple dozen people who managed to snag a 5090 might have a class action on their hands
The current rumor seems say it's not intended and I'd guess they would just replace the affected units especially given there is so few out there. Could also be a factor in the supply issue and the reported ramp up.
What could be interesting with all this is should there be replacements this would be likely rather straight forward for first party purchases. But with scalping and second hand purchases it could get quite interesting. There's a side of me that would actually like it if there were "landmines" like this that discourages scalping. The best case scenario would be at least those who sold on ebay have to take returns (or face buyer protection clawback), wait for the replacement and by the time they get it find the resell market has dried up.
Are we absolutely sure the drop in support is purely due to deprecating things on the software side?
Given that Blackwell is seemingly extracting more performance and functionality out of seemingly the same die/transistor cost, I would wonder if there are changes at the sub SM level that break compatibility with 32 bit CUDA?
DavidGraham
Veteran
NVIDIA says some 5070Tis are affected as well. The level of QA is this launch is horrendous.
Is that technically possible? It could run 64bit CUDA but not 32bit CUDA?
I googled and it seems the version of NVCC that supports Blackwell does not support the 32bit compiler switch at all.
What's the last version of the CUDA Toolkit to support 32-bit applications?
On godbolt what I observe is that if I compile a trivial code with -m32 compiler switch to nvcc, that on CUDA 11.8 I get a warning message like this: nvcc warning : Compiling in the 32-bit mode is deprecated and may not be supported in future releases when I compile the same code on CUDA...
forums.developer.nvidia.com
Used to say
Now it says:
Yes, and that's part the reason why CPU today are so heavy. One thing Apple can do is to remove 32 bits support from their ARM CPU, but when Intel tries to do that, they faced a lot of resisitance. The X86S project has to be shelved.
