Hacker News Reader

Comment by scrooched_moose | Hacker News Reader

GPU acceleration in scientific computing. My god in 2011 it was the only thing anyone talked about. Every COTS solver package cobbled together support in less than a year and were promising order of magnitude speedups.

In 2018 it's still hanging around but realizing a 20% speedup on certain classes of problems. Maybe.

View on HN

Relevant war story: I was one of the engineers who initially started looking at GPGPU computing at one of my former employers. This was circa 2007 and people were getting really excited thanks largely to NVIDIA's hype machine. 2007 BTW, was when the first Teslas were making their way into the wild.

I was part of a group that developed signal processing algorithms and my boss's boss pulled me aside and asked me what we ought to do about this stuff. I told him that there are a few applications where this works very well, but it's pretty niche overall and we shouldn't do more than what we'd already done. He was a bit surprised because he thought I'd be one of the kool-aid drinkers given how much I knew about and liked playing with the tech. But he deferred to my judgement.

It turned out to be one of the most significant conversations I've had. I was still a junior engineer fresh out of college, but my bosses realized that I could be trusted put aside naive enthusiasm about the tech and make cool-headed decisions about what was strategically important. And I learned to trust my intuition about what was useful and what was just hype.

> GPU acceleration in scientific computing

> it's still hanging around but realizing a 20% speedup on certain classes of problems

It really varies though. In grad school I felt like god when some molecular dynamics simulations running on my pair of GTX 760s finished faster than the same thing running on my advisor's 96-core CPU cluster.

Point taken, though, that it was over-hyped. There were lots of conference presentations with GPU-acceleration in the title (kind of like deep learning is now).

Wasn't your advisor just using a smaller, more conservative value for h?

Amdahl's law strikes again.

At the same time, I really wish that GPUs would take off in neuroimaging, where 90% of the steps performed in an analysis pipeline are not modeling-related, but rather basic image manipulations / pre-processing (i.e., scaling, smoothing; stuff that GPUs are actually good at).

Didn't that tech shifted to AI/ML? Isn't that considered scientific computing by some measure?

I'm definitely not an expert so I'm maybe missing something here.

"Scientific Computing" was probably too general of a term to use. I focus more on FEA/CFD type problems where it's effectively been a bust. Most of our benchmarks show 8 cpu cores + 1 GPU is roughly equivalent to 10 cpu cores, and a second GPU has little additional benefit.

I'm sure there are other disciplines that have seen a larger benefit.

Reply on news.ycombinator.com