16.8M Core Graph Processing Beast (2023)

The question that isn’t be asked, but should be, is if this idea is so great then why were none of the several prior generations of similar computers successful? To be clear about where my perspective is coming from, I was one of the people doing graph algorithm research on exotic graph processors many years ago and the old Tera machines mentioned in the article are one of my favorite computing architectures of all time. I am predisposed to liking the tech but I also have a realistic view of the tradeoffs.

These architectures were killed by two separate issues. The first is that very few programmers seem to develop an intuitive sense of how to design algorithms and data structures that are efficient on these architectures. Real-world performance is poor because the software is poor. The second, and more permanent state of affairs, is that along the way researchers developed new algorithms and software architectures that run on ordinary CPUs only a little less efficiently, while taking advantage of the faster and much cheaper commodity silicon. Why would you buy exotic silicon when you can call Dell?

There is a community of diehard graph silicon enthusiasts inside the US DoD who continue to fund the development of processors like the one in the article, so we get a new design every few years. And it is cool tech! But if I were building a business around graph processing, I’d use commodity silicon without a second thought. Not that I wouldn’t love to play with one of these new graph computers just to see what they can do.

The framing is weird: The TLAs that gave Tera enough money to buy the Cray badge and that are funding these spiritual descendants don’t care about AI. They’ve got some core, very secret “needle in a haystack” algorithms that really like chasing computed pointers (and popcount… never forget popcount!)

Zooming in on a boundary section of the Mandelbrot set will fuck it up depending on how fast you want it to. Displaying associated Julia sets from the focus in a fly out will simply enhance the fun!

"That’s just the time we live in now." - no it isn't. That beastie might be handy for something in the future but it isn't "I". It will deliver really fast weather forecasts and that's useful.

Implement the GraphBLAS API of course!

https://graphblas.org/

A relaxing game of Solitaire.

Looking at the block diagrams, it looks more like it's a giant mesh router with a few processing cores sprinkled on top, Salt Bae style. After all, the cores take up just over 4% of the transistors. Not sure if that includes the 4MB per-core scratchpad though.

Run the UniProt sparql service on it ;) I was lucky to test the yarcdata implementation of sparql on an xmt machine around 2012, super UI but single user.

Rasterize Quake