US seeks exascale systems 10 times faster than current state-of-the-art computers

The US Department of Energy is looking to vendors that will help build supercomputers up to 10 times faster than the recently inaugurated Frontier exascale system to come on stream between 2025 and 2030, and even more powerful systems than that for the 2030s. These details were disclosed in a request for information (RFI) …

[Pascal Monett]

10 times faster ?

We're not in 1980 anymore.

Improvements are incremental these days. 10% in the best of cases.

Exascale is not just the CPU, it's the memory, the interconnects, the bandwidth.

I don't know what bandwidth exascale has at the moment, but I'm pretty sure it'll be a while before it gets 10 times faster.

Certainly not by next year but, hey, I'm willing to be proven wrong.

[Yet Another Anonymous coward]

Re: 10 times faster ?

But supercomputing scales, you want 10x today? Just buy 10 systems and put them in the same building.

When you have a 60,000 node supercomputer you aren't considering the bits of a job that don't parallelize

[patrioticparadox]

Re: 10 times faster ?

While I appreciate your opinion, you are entirely incorrect.

When working at the exascale the interconnects, and their proficiency, are paramount.

https://spectrum.ieee.org/frontier-exascale-supercomputer

Anyone can throw data at a fast computer. These days it matters how the systems handle that data. Talking to each other about a datum is non-negligible

[Yet Another Anonymous coward]

Re: 10 times faster ?

Yes, my point is that this isn't like going from a Cray XMP to a Cray YMP in the old days.

If your job runs effectively on a 60,000 node hybrid Xeon+GPU machine, it's like to run pretty well on an 100,000 node version. Yes it's the interconnect technology and cost that drive these machines but there isn't necessarily an obvious hard limit at today's scale

[2+2=5]

Power envelope

> Any such supercomputer will be expected to operate within a power envelope of 20-60MW, according to the DoE.

Damn. I have the perfect solution but it only consumes 5MW.... Guess I'll just have to no bid and try selling to $other_country instead.

[patrioticparadox]

Re: Power envelope

https://spectrum.ieee.org/frontier-exascale-supercomputer#toggle-gdpr

The article is already off base this article. The only truth is that the US cyber command is currently engaging prospects for an even faster computer. Because they always are. It would be dumb to wait for the fastest computer in the world to be completed before you start building the next fastest computer in the world.

Politicize all you want the President, Congress and even the head of some orgs like the FCC but these others dont warrant airtime so they just keep on keeping on.

The United States currently has the fastest super-computer in the world. It's called Frontier. It's at Oak Ridge national laboratory.

They are already building a faster one.

[Chris Gray 1]

uptime?

Has anyone heard any uptime numbers for the Chinese systems? What about for the current western systems? Do we expect the numbers to be similar? They might not be, in which case this could be a case of the west spending a lot of money to chase something that isn't actually real.

[Yet Another Anonymous coward]

Re: uptime?

We don't have to use them, we just have to build them.

We need to have the biggestly computer or we lose

Actually using them would require having lots of highly educated scientists, do you know how expensive that would be? And it would be a terrible waste of all that expensive education if they were just modeling stuff when they could be leveraging dynamic synergies on Wall St

[HildyJ]

Re: uptime?

Indeed. The Big Dick Data aspect of computer design.

[Anonymous Coward]

Re: uptime?

Quite. Software is the real issue, currently.The money would be better invested in funding for an attractive, viable and sustainable career path for HPC software engineers. But no, good software takes time, and waving willies now is more important than getting good use out of willies in a number of years time.