4 ms·

Interesting idea, however in practice: https://pcpartpicker.com/products/cpu/#g=0 https://pcpartpicker.com/products/cpu/#g=0 (779 products) vs https://pcpartpic

Interesting idea, however in practice: https://pcpartpicker.com/products/cpu/#g=0 https://pcpartpicker.com/products/cpu/#g=0 (779 products) vs https://pcpartpicker.com/products/cpu/ https://pcpartpicker.com/products/cpu/ (1369 products), and this is mostly just desktop/consumer CPUs - servers are much less likely to have one, and I don't even know where to start researching timesharing GPUs on VMs - most people just do passthrough, usually as a part of a dedicated gaming or transcoding build. One thing I've observed, is that Electron apps all tend to ship libEGL.so, libGLESv2.so, libvulkan.so, etc rather than depending on system Mesa, so the part where the application tries to take advantage of a GPU by any means necessary (if possible) is already happening, it's just it appears to be Chromium-shaped rather than GCC-shaped.

bonton89 3y ago
Intel actually had GVT-g which let you do 7 time slices of the iGPU and give them to VMs. They abandoned it after 10th gen however it was never available on any of their dedicated GPUs.
rbanffy 3y ago
> servers are much less likely to have one That’s a chicken and egg problem: there is no server software for it because there is no server hardware to run it on and there is no incentive to build the hardware because the software doesn’t exist. But that’s an x86 thing. The IBM Telum processor is as server-grade as it gets and has a huge fraction of its real estate dedicated to an inference accelerator. Of course, it’s not a GPU because mainframes tend to not have those, but it shows that if there’s a need (real-time fraud detection) someone will build a product.