Supercomputers in 2030: Lots of exaflops and LOTS of DRAM
The future – like the past, but much hotter and drinking way more juice
Supercomputers will overcome the predicted crumbling of Moore’s Law over the next few years to show massive leaps in performance between now and 2030 – but will still look pretty familiar to today’s compute power junkies.
In a series of sessions at ISC this morning covering exascale architectures, speakers predicted a rapid scaleup in computer power.
Al Garra, Intel fellow and chief architect of exascale at the chip maker, said in a session on supercomputing in 2030 that five years ago achieving exascale computing was seen as the key challenge, but the industry could not see the path forward.
Now, he claimed, “we know how to get there”, though the industry would have to solve a number of problems on the way.
Going into full crystal ball mode, Gara took the US’s Sequoia supercomputer as a benchmark. That system – switched on in 2012 – had a theoretical performance of 20 petaflops.
Gara predicted two exaflop systems by 2022, offering a 100 times general performance jump on Sequoia. 20 exaflops should be possible by 2026, with a 1000x performance improvement, with 100 exaflops by 2030. This should deliver a 5000x general performance, improvement on Sequoia, or 50,000 for specialist applications.
That 2022 system would be packing 8PB of memory, and running at 25MW, while the 2030 system would be packing 60PB but running at a whopping 80MW.
And what will those systems actually be running? One constant was the presence of DRAM, a 1970s technology which was “not likely to be replaced by 2030.”
New memory solutions could find their way into specific applications, he said, while, he said, new packaging techniques could see DRAM moving closer to the logic.
Fabric architectures could be a constraint he predicted with optics remaining a big challenge. But he predicted: “Silicon photonics will become primetime, enabling a big drop” in the cost of bandwidth.”
In the longer term, he said unless lasers could be developed that were 100 times more efficient, new materials would be needed to lower overall power consumption. Carbon nanotubes would be nice to have, he added.
Micron’s veep of advanced computer solutions Steve Pawlowski, speaking in an earlier session, also predicted DRAM would be a constant, despite there being 30 other alternatives. The problem was, he said, despite engineers’ best efforts, “some of them really want to be a memory, and some of them don’t.”
He predicted that some of the newer architectures could evolve into a third tier of memory between RAM and storage.
At the same as DRAM moves closer to the logic in a system, he suggested some logic would move into memory.
“We’re looking at can we build a super computing system where we can share some of the that capability between the CPU and the memory.” ®