GTX 770 and Compute Capability 3.5

Hello.

There is defined (https://developer.nvidia.com/cuda-gpus) that GeForce GTX 770 has Compute Capability 3.5 (Dynamic parallelism…).
I see only 3.0 ! What is wrong ?

deviceQuery
Device 0: "GeForce GTX 770"
  CUDA Driver Version / Runtime Version          5.5 / 5.0
  CUDA Capability Major/Minor version number:    3.0

M.C>

What driver version are you running?

Binary: NVIDIA-Linux-x86_64-319.23.run
nvidia-smi -q | grep “Driver Version”: Driver Version : 319.23

did you try the CUDA toolkit 5.5 release candidate yet? does it make a difference?

It seems like the link has incorrect information:
GTX 770 is GK104 based (cc3.0)
GTX 780 is GK110 based (cc3.5)

Ouch, I hope you didn’t purchase the 770 based on the incorrect information you found on the nVidia page…

Yes, I buy.

CUDA toolkit 5.5 does not help. It is the GK104 problem.

Device 0: "GeForce GTX 770"
  CUDA Driver Version / Runtime Version          5.5 / 5.5
  CUDA Capability Major/Minor version number:    3.0

I suppose that new released GTX 770 has some step up in chip that support cc3.5. Also accessible marketing with release GTX 770 says - the same feature as GTX Titan and GTX 780 !
I see that https://developer.nvidia.com/cuda-gpus is now corrected.

:-(((((

http://en.wikipedia.org/wiki/CUDA should be modified too.

Thanks for raising this inconsistency. The CUDA GPU website has been updated and I will look into the other sites.

Hopefully your prompt notification will help others - and we appreciate that, I have already emailed you, I want to ensure you are able to continue your CUDA Development work, so lets talk.

Thanks again

It’s already 6 years, but I’m still an owner of both GTX 770 and GTX 740m.

Surprisingly, GTX 770 has the compute capability = 3.0, so I can’t use the most current version of PyTorch library. What’s even more interesting, is a fact that GTX 740m has the compute capability = 3.5 which allows me to use that particular version of PyTorch (1.0.1).

How is that two cards released in the same year (as far as I know) has different compute capabilities and cheaper one has a higher version? What’s more, it’s stated that GTX 740m should also have the compute capability = 3.0 in the link provided by OP. GTX 770 runs on CUDA 9.0 and GTX 740m on CUDA 9.1. Both cards use a nvidia driver version of 390.116.

According to Wikipedia, there are two variants of the GTX 740M: The initial release was based on the GK107 die (compute capability 3.0), while the later variant is based on the GK208 die (compute capability 3.5). The latter is a smaller die (118 mm² vs. 87 mm², according to the TechPowerUp database) and therefore cheaper to manufacture.

I think this is not the only instance where the same GPU type has been based on two different dies. It makes sense for NVIDIA to list the compute capability of the less capable variant in such cases.

The wisdom of offering GPUs with the same name that are based on different hardware under the hood is less clear. Presumably it has to do with the desire to offer a GPU at a particular price/performance point at minimal cost, maximizing profit.

Thanks for an exhaustive answer. So, I’m now surprised why they manufactured, in fact, two different devices under the same name. It doesn’t change the fact that GTX 770 seems to be useless nowadays for current frameworks but I can still use a mobile GTX 740m.

This isn’t really surprising, as explained by Njuffa, specially if you also consider for example the GTX 750 Ti. It was manufactured a couple of years after the GTX 770, it has inferior specs to the 770 and it is slower than the 770. But it is a Maxwell (cc 5.0). The GT 1030 is slower than a 770, but better for programming since it is a Pascal (cc 6.0) and supports newer features.

And though it is 6 years since the GTX 770 was launched and the evolving cycle was pretty fast, I have a MBP from 2012 with a GT650M and 1GB, very inferior to the 770, and it is my go-to system to test work before shipping. But one day everything comes to an end.