Nividia is sunseting CUDA support for a range of GPUs
28 Jan 2025 5:22:27 UTC
Topic 232022
(moderation:
Maxwell, Pascal and Volta:
https://www.theregister.com/2025/01/27/nvidia_deprecates_cuda_support/
Just means you will have to
)
Just means you will have to scavenge for Legacy drivers until none are to be ever found again.
The gpus will still be usable as long as the applications still support them.
to add on, it seems like the
)
to add on, it seems like the drivers will still support them (at least for a short while, but they wont include any new CUDA features that might come in CUDA 12.8 and in the future. so they will be stuck at whatever features are available in CUDA 12.6 and lower. which isnt really a huge deal. most of the apps here and other places are not even using the latest features to require the latest CUDA driver anyway.
_________________________________________________________________________
7900 XTX FP64 = 1.918 TFLOPS
)
7900 XTX FP64 = 1.918 TFLOPS (1:32)
7900 PRO FP64 = 1.916 TFLOPS (1:32)
4090 FP64 = 1,290 GFLOPS (1:64)
6000 ADA FP64 = 1,423 GFLOPS (1:64)
5090 FP64 = 1.637 TFLOPS (1:64)
Meanwhile:
R VII FP64 = 3.360 TFLOPS (1:4)
Titan V FP64 = 7.450 TFLOPS (1:2)
GV100 FP64 = 8.330 TFLOPS (1:2)
Why the big middle finger to local workstation high precision compute in the past 8+ years?
FP64 isnt a major factor here
)
FP64 isnt a major factor here at Einstein. it's needed and used to some degree, but usually not a large part of the calculation where performance scales in a direct relationship to that. Einstein cares most about fast memory and a lot of GPU memory bandwidth. those cards you listed also have HBM memory, so the good performance from them is from the memory more than the FP64 performance.
not sure what the Radeon VII has to do with Nvidia drivers discussion though.
these cards will not stop working. you just keep the latest driver available. the drivers right now support the most product generations than nvidia has ever before in the past. but it's understandable that they have to draw the line in the sand somewhere. they did the same thing with dropping Kepler support a few years ago.
_________________________________________________________________________
Ian&Steve C. schrieb: FP64
)
Thank you. Enlightening. Then the benchmarks (in boinc app) make much less sense for Einstein. A test file for all, would be much more telling. Could someone at Einstein set up such a standard file?
Cuda is clearly from the past, and need to be exchanged to OpenCL as an open standard. Nvidia is going even away from OpenCL, to push their own propritary format. In my opinion, OpenCL should be the standard for crunching.
not sure what you mean by
)
not sure what you mean by CUDA being "from the past". it's still the present and actively developed, and it's almost always much faster computation than opencl equivalents. even AMD cards become faster when you run CUDA on them (with ZLUDA). CUDA is superior.
if you want to setup a task as a benchmark, you can just copy all the information about how to run tasks (read the client_state.xml file to piece together how the apps/tasks run and their respective command line arguments, there's a lot of them) and run the task offline outside of BOINC.
_________________________________________________________________________
Ian&Steve C. schrieb: not
)
Okay, i believe you. This was only my uninformed opinion. You seem to have much more knowlege about this topic, than me.
Thank you, for the answer!