I used to work in high performance computing using Department of Energy supercomputers. You can't run 1990s era HPC applications on a GPU. Nor can you run a full web browser on a GPU. GPUs are not magic go-faster devices. They're really good at executing certain kinds of operations and terrible at others.

Seeing tweets like this makes me sad that Twitter warps even smart people toward writing for quippiness over thoughtfulness.

Blow does this constantly, with his takes on programming languages as well. He almost exclusively comes from a game-development background but constantly makes overly broad statements.

Agreed. He also argued on Twitter that no one should need Linux containers because ELF exists. He ultimately backtracked to something like "well really we just need to go back in time and avoid dynamic linking of any kind and put everything in ELF binaries and then build Kubernetes off of that". Which I don't entirely disagree with--personally ELF kinda sucks and that's not the first thing I would do with a time machine and it also doesn't solve for isolation at all (container isolation is imperfect but I think it's worth something) and the whole content addressability and layer caching thing goes out the window (but Blow would also argue--and I quite agree albeit not so absolutely--that we don't need GB-sized binaries), but I can appreciate the simplicity of a simpler package format.

I can run a https client and ssh on a esp32. There is no reason Element (chat app on electron) would use half a gigabyte of RAM for some text, a unicolor flat theme, and some realtime video call. It's just experiment over experiment, we've lost the art of making simple software.

Sure, you can't run a 1990s-era HPC application unmodified on a GPU, but you can write applications that are functionally equivalent (take the same inputs, produce the same outputs) that do the heavy lifting on the GPU.

The latest DoE supercomputers are mostly GPUs. Summit has around 10 PFLOPS on CPUs and 215 PFLOPS on GPUs.

https://en.wikichip.org/wiki/supercomputers/summit

Yes, I agree with that. But it also takes a lot of software work. Getting decent performance on a GPU for an existing scientific application requires something closer to a full rewrite than the usual porting effort to a novel CPU architecture. That is one of the reasons that people are writing new scientific applications to target GPUs from the beginning rather than just adapting mature older programs to GPUs. (Though there is some of both approaches.)

You can probably run them on a modern CPU, though. An i5-5400K runs at 340 gigaflops... single-threaded.

I agree with the sentiment the web has gotten slow for no reason in some ways, but at the end of the day not everything that has made the web so "heavy" is Js developer self-pleasuring and ad-tech.

A modern browser has PDF rendering, video rendering, 3D, AR/VR, a camera viewer, an RTC platform, a screencaster, and much much more. If you tried to run the equivalents of all of this from back in the day at the same time, you'd also bring those old PCs to a crawl too.

I feel like I see the argument that tweet makes a lot and the answer if you look at what browsers are expected to do now is... yeah actually it does.

-

There is a real problem of what we're doing with all that power sometimes, but the drive was created by real need from users. The wild west days of installing a new piece of software for every single utility were great for technical people, but not so great for making the PC an accessible piece of technology