Hacker News Reader

Comment by CyberDildonics | Hacker News Reader

This is from a paper and any decent paper says what hardware their benchmarks were run on. No one is benchmarking on hardware that isn't accessible to the general public.

"We measure timings for rendering in Unreal and for all denoisers except NBG on an NVIDIA RTX 3070. Our PyTorch implementation is unable to run at real-time frame rates. Therefore, we use NVIDIA TensorRT to generate reasonably optimized CUDA kernels for inference of all neural network layers in our method."

I don't know why you would think a rasberry pi factors in to cutting edge graphics research at all, it makes no sense.

View on HN

Makes perfect sense, Pi4 has a GPU after all and runs all the modern "AI" software.

It doesn't make sense. Based on what limited info is out there, the Pi 4's cpu is around 15 GFLOPS, and the GPU is roughly 32 GFLOPS. The cpu in my 4 year old laptop hits at least 58 GFLOPS single core and 216 GFLOPS multicore, on top of the 400 GFLOPS gpu. Considering that laptop cpus are much more accessible than Raspberry Pis (the 4gb model apparently going for $200 on amazon) and they're more performant, it makes no sense to benchmark on a raspberry pi. That doesn't even get into the fact that CUDA code (which most AI work uses) won't run, and the raspberry pi's GPU ecosystem appears to be a nightmare to work with [1]

[1] https://news.ycombinator.com/item?id=31147393

These GPUs have more memory than the whole rasberry pi. Why would someone use a computer that is going to be at least 1/100th the speed of some typical commodity hardware? Do use a Pi4 for everything you do? Are you posting this from a Pi4?

Also this paper is from Intel so this is extra ridiculous.

Intel spent billions on Edison trying not to acknowledge the Pi, I know. But it's here to stay now. And nobody needs to be impressed about numbers when the very platform changes all the time. How many Vaxen is the Intel setup the equivalent to?

This is so incoherent I'm not even sure if you understand the context of this thread. There is nothing you wrote that has any connection to reality here.

This is research by Intel written in CUDA which only runs on nvidia cards. It is about filtering as fast as possible on higher end consumer hardware, because it being made for games. It is about complex filtering of hardware ray traced renders that use lots of extra information to suppress noise as fast as possible.

There is no part of this that makes sense for a $35 15 watt computer to be involved. If you suggested at a meeting about this project some people might chuckle thinking it's a joke, the rest would probably try not to make eye contact out of embarrassment.

Reply on news.ycombinator.com