Personally I haven’t used CUDA, but if anybody could comment on its ease of use compared to Vulkan that would be great.
Source? I can't find anything for '"cublas" scheduler'…
Bespoke accelerator FPGA solutions are the future and will probably scale a lot longer than CPU/GPU combinations.
(I use "fixed function" loosely here. It's not a fixed function T&L pipleline, it's meant in the "the circuit is fixed but can execute arbitrary code" sense.)
This isn't my space as I operate on low-latency networking side.
> A more traditional "many 'fixed function' processor" a la GPU will beat it for any algorithm which at least semi-successfully maps to the provided operations. And most math is matrix multiplication anyway, if you squint your eyes and hold your tongue at the right angle ;)
I don't think it is as simple as this. Certainly not for low-latency scenarios or for HPC efficiency where the gains are done my moving as much computation as possible onto accelerator cards and off the CPU.
The cost of moving things back and forth from the CPU can be expensive (data especially). We are already seeing dedicated solutions from Google, Graphcore that beats the pants off the traditional GPU/CPU approach.
In this case being video streaming in from the NIC, encoding/decoding occurring directly on the NIC.
Same to a degree with the bespoke smartssd where if you have some complex encryption, encoding, decoding or data driven application than the processing happens directly on the storage device via the FPGA.
this is a big deal especially since ai tech advanced.