hello, question are all silly: what does Nvidia mean by compute capability? , because the Titan RTX, GTX 10xO and RTX 20x0 have the same score, but are of different technologies and the benchmark is not explained
compute capability refers to a specific design member of an architectural generation
GTX 10x0 (compute capability 6.1) does not have the same compute capability as Titan RTX and RTX 20x0 (compute capability 7.5).
6.1 = pascal generation
7.5 = turing generation
If you study some tables in the programming guide, you can see how compute capability is used to determine certain technical capabilities and throughputs of the chip:
thank you if i understand correctly and you correct me if i am wrong: the 16xx and 20xx series having the same architecture, therefore do the same thing
At this point in time, this is a true statement. GeForce 16xx and GeForce 20xx SKUs all use GPUs with the Turing architecture.
That is a very ambiguous statement. All GPUs with the same compute capability have the same set of basic capabilities and are binary compatible at the machine code level; they differ in the amount of hardware resources that implement these capabilities.
We know it very well, but I buy myself during the year a cluster computer with ryzen, but I hesitate between 4 GPU 1660 super or 4 GPU 2060, I do machine learning and as I read contradictory remarks, I am somewhat lost, so if I am told Turing = Turing and too bad for the Tensor Core optimized for Tensor Flow (which I use but not only …)
I see what you are saying. Best I can tell, all Turing architecture chips have compute capability 7.5, but reviews of the GeForce 1660 (Super) indicate that it doesn’t come with Tensor Cores or Ray Tracing cores (emphasis in citations mine):
If true, that is a very confusing move on NVIDIA’s part. There may be handwave-y talking points from NVIDIA marketing why tensors cores are not considered to be part of the compute capability but I haven’t seen any.
It seems that if you want/need tensor cores, you want a GeForce 20xx.
thank you, I take your remark into consideration, for a year I hear everything and its opposite, read blog, white paper, chain bench, but nothing that takes into account the ML and pro who not the means of past on Titan RTX, I must not be the only one in this case, I leave the discussions open, have a good evening
Compute Capability (CC) is a scam to push people to buy newer graphics cards… CC should be stackable so if I have 2 GPUs of 5 then I should have a total of 10 CC… Otherwise Nvidia will just keep people to buying new GPUs when the old ones are perfectly good. This is equivalent of the iPhone scam - where users have to keep buying new iPhones when in fact the older ones work perfectly well.
Compute capability is just another name for “GPU architecture” with the addition of an “onion model” where each new architecture supports the features of previous architectures at PTX level. While evolution of GPUs has slowed down a bit, it still proceeds at considerable speed. Because of that, there is generally no binary compatibility between GPU architectures.
There is a cost associated with maintaining support for a given GPU architecture which cannot be borne forever. The question then becomes: when should a platform be deprecated and taken out of support. Currently, NVIDIA supports architectures down to CC 5.0 (Maxwell; first shipped February 2014: 11 years ago) which seems like a reasonable cut-off for 99% of real-life scenarios.
Even projects like Linux will eventually stop supporting old hardware. For example, support for 386 processors stopped in 2012 after 20+ years.
Nvidia handles this much better than most other tech companies.
There are PTX compilers integrated into the graphics driver to optionally make code compatible to new architectures.
And the compute capabilities are mostly stable, some few new features each generation. Even as @njuffa mentioned, the active compatibility goes back a bit more than 10 years, you can write code that compiles with today’s hardware as well as the one 15 years ago.
Your reasoning of 2 GPUs of CC 5 = CC 10 is wrong.
The CC does say nothing about speed or performance, but about features.
So 2 GPUs of CC 5 = CC 5.