non-NVIDIA GPU support ?
talks about how GeForce has become deprioritized by Nvidia, and
The chatter from the grapevine is that we won't see any new GPUs from Nvidia this
year at all — not one — and that's very rare (in fact it hasn't happened in three
decades). This is because Nvidia needs all the chips it can get — and perhaps more
to the point, all the video RAM — for AI graphics cards which are far more
profitable than consumer models.
Soon, Mathworks will be facing a choice: continue to support only expensive Nvidia AI offerings -- or diversify to support alternative GPUs as well.
By the way: Nvidia AI units cost $US7.8 million dollars. https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidias-memory-costs-soar-485-percent-latest-ai-systems-now-cost-usd7-8-million-to-build-memory-now-comprises-25-percent-of-the-total-cost-rubin-gpus-a-mere-usd50-000-apiece
Which alternative GPUs would you most like to see supported?
- Unfortunately, I hear that Apple provides poor support for information on really using their "silicon" GPUs in any way other than Apple's pre-packaged computation libraries. The Apple attitude is apparently that anything that is not already nailed down by documentation is fair game for changing in the future, and that documenting how the GPUs really work would constitute nailing them down, supposedly "destroying" Apple's creativity. Exception: Apple is known to work with major gaming studios (but only the major ones.)
- The Apple silicon series of GPU does not provide any 64 bit operations, so 64 bit support would require emulating 64 bits in software. Nvidia is famous for internally implementing 64 bit support in terms of 32 bit operations, at 1:32 of the speed -- but on the other hand select Nvidia devices operate 64 bit operations at 1:24 or even 1:8 (a small number of devices) through hardware acceleration units. People who need 64 bit operations have the option of shopping very carefully in Nvidia's line to get faster 64 bit processing.
- OpenCL sounds cool and "open". Unfortunately it turns out that a lot of OpenCL operations are optional, so efficient OpenCL libraries would need to be tuned to the exact hardware series.
- OpenCL is not supported on semi-recent MacOS Intel or Apple Silicon series
- I seem to recall hearing that OpenCL is no longer supported by Nvidia either
