Fujitsu's billion-dollar ARM supercomputer delayed by up to 2 years Fujitsu's monster ARM-powered supercomputer, the Post-K, will miss its 2020 deadline. The $910m project has stalled because its engineers need more time to design, test and perfect the machine's processors, we can confirm. Fujitsu was hired by Japan's boffinry nerve-center RIKEN to produce a roughly 1,000 peta-FLOPS …
Only in Management Land can one both work on the bleeding edge and still be on-time and on-budget. Because the bleeding edge can just be looked up in books, right, how hard can it be.
These programs are "industrial policy" anyway - get new researchers into the field, keep the knowledge alive, build the bases to attack the next problem earlier than the country next door. The money will flow (even it it has to be printed). It's all in good fun.
(Anybody remember the "Strategic Defense Initiative"? It started off with the idea of building Intelligent Machines (always good for Cold War machism), then when it became clear that that was kinda out there, quickly veered into supercomputing and integrated communications. Success!)
Not sure the issue will be with the process side of things. Supposedly, Apple already has 10nm products from TSMC. We might even see them announced later today...
Note that TSMC may or may not be actual 10nm - their 16nm was 20nm with FinFET gains to approximate 16nm (http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm). In which case it may be a 14nm SoC base process with FinFET giving equivalent performance to 10nm in comparison to the 14nm SoC base process.
Which leaves the issue being Fujitsu's chip design... And most likely the performance they can extract from it at present.
"Not sure the issue will be with the process side of things. Supposedly, Apple already has 10nm products from TSMC. We might even see them announced later today..."
In which case it's possible that Apple are Bogarting the 10nm capacity, and Apple could well have bought out Fujitsu's slice of the pie to cover their own yield shortfall... ;)
It certainly is true that when one is trying to achieve things on the edge of what is possible, there are going to be unforeseen delays. I hope that everything goes right, and it gets out the door, working perfectly, as soon as possible, though, so as to encourage the adoption of instructions for processing longer vectors elsewhere, not just in a select subset of supercomputers. If this kind of vector processing were used by gamers, then it would be available even for COTS-based supercomputing as well, a very good positive feedback loop.
"The Post-K system will be used to model climate change, predict disasters, develop drugs and fuels, and run other scientific simulations."
Being 8 times faster than the world's current leader, it is obviously intended to become the world's first piece of sentient silicon. Welcome your new master.
I could make up a super computer by connecting together a BUNCH (10k or so) Raspberry Pi's and calling it a "super computer". Even with all the wires and power supplies, it would be less than $1,000,000 pretty easily.
It IS ARM, and available TODAY!
Now to get it programmed correctly...
You can call your bunch of ARM processors a supercomputer, but whether they can be used for much will be down to how well you can distribute tasks across them which will come down to I/O and memory bandwidth. I mean, what point is having 10,000 or more CPU's available to you if the first 100 or so have finished the tasks you have distributed to them before the last 5000 or so processors have received any work?
You can break the bunch of processors down into nodes, but you can do that with other processors too.
I suspect, the advantage of Fujitsu's ARM's won't so much be in the ARM core as the offload processors that accompany it which is why they can't use off-the-shelf products. Potentially what Fujitsu need is the next die shrink to get the performance they need from each SoC to make this project worthwhile...
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
