Slashdot Mirror

According to a report by The Register, Fujitsu will use ARM to build an exascale supercomputer, which is called Post K.

According to a report by InsideHPC, the supercomputer will debut in 2021, which is the same year in which Aurora will debut.

According to another report by InsideHPC, the ARM supercomputer will be tuned for high performance in machine-learning applications.

Fearing an economic Pearl Harbor, Washington quickly asked Cray to develop an answer to Post K. Aurora is the answer.

According to a report by The Register, Fujitsu will use ARM to build an exascale supercomputer, which is called Post K.

According to a report by InsideHPC, the supercomputer will debut in 2021, which is the same year in which Aurora will debut.

According to another report by InsideHPC, the ARM supercomputer will be tuned for high performance in machine-learning applications.

Fearing an economic Pearl Harbor, Washington quickly asked Cray to develop an answer to Post K. Aurora is the answer.

You're entirely right that the memory subsystem is 90% of the battle for most server workloads once you exceed ten cores.

For integer workloads with unreasonable parallelism and unreasonable cache locality (that Intel's AVX doesn't already handle almost ideally), I'm sure this design will smoke Intel on the thermal management envelope, a nice niche to gain Qualcomm some traction in the server mix, but hardly a shot heard around the world.

And Qualcomm better be good, because Intel will soon respond with Omni-Path Knights Hill—perhaps also larded with HBM—that could probably take on the same workload between power sprints (less power efficiency in the CPU itself—which isn't always the main power draw—and probably more flexible as part of a tidy one-vendor-rules-them-all server mix).

I'm all for vendor diversity, but let's not get ahead of ourselves thinking that 10 nm levels the playing field, sucking down the data aquifer through a double-wide handful of drinking straws.

Yes, core count matters, but size matters even more when it comes to the hose.

—

Looky looky, the bow moveth:

Intel announcements for AI: Nervana 100x faster than GPU, Knights Crest & Mill 4x faster, SKL mid-17

Kx Streaming Analytics Crunches 1.2 Billion NYC Taxi Data Points using Intel Xeon Phi

I suppose you guys will shred me for this, but I took on this topic in my SCI-FI story, "The Observer Effect" where a scientist attempts to prove the existence of god through a kind of "cat in the box" experiment using a large cluster spun up on AWS. You can imagine that it doesn't end well. http://insidehpc.com/2013/11/o... For my day job, I write tech news at insideHPC. Forgive me for sharing my hobby here, but I think it has something to say about this topic. -Rich

I've seen estimates of the brain's computing power ranging from 38 PetaFLOPS, which Tianhe-2 is approaching, to 6.4 ExaFLOPS, which was the estimated computing power of all the computers on Earth in 2007. I suspect the actual figure is closer to the latter than the former.

And yet you need an ever increasing super computer to track and simulate them???

In the old days, we just detonated a bomb or two if we needed experimental data. Since the test-ban treaty, however, numerical simulations have been used in lieu of physical experiments. I suspect that the accuracy of numerical simulations is a closely guarded secret, and the DOE hasn't yet decided that the present state of the art can't be improved upon.

There is a rival benchmark, Graph 500 Roadrunner isn't on it. Neither is Cielo.

But, it's intended to simulate a different sort of problem set.

And yet another perspective comes from Intel’s John Gustafson, a Director at Intel Labs in Santa Clara, CA, “The answer is simple: Graph 500 stresses the performance bottleneck for modern supercomputers. The Top 500 stresses double precision floating-point, which vendors have made so fast that it has become almost completely irrelevant at predicting performance for the full range of applications. Graph 500 is communication-intensive, which is exactly what we need to improve the most. Make it a benchmark to win, and vendors will work harder at relieving the bottleneck of communication.”

The Case for the Graph 500 – Really Fast or Really Productive? Pick One

Just because the Chinese popped in there with #1, doesn't mean that we're losing our domination in the industry all of a sudden. Plans are well on the way for the next generation of supercomputers. We're talking about 20 petaflops and up,. . .

Google powers its servers off 12V DC with inbuilt batteries on each server. The savings in power costs make this worth it. http://insidehpc.com/2009/04/02/google-unveils-its-super-secret-server-design-dc-and-batteries-built-in/

- What about a super computer made out of FPGAs ?

It's been done.... more than once, or twice.

New FPGA-based Supercomputer in Scotland
SGI Builds World's Largest FPGA Supercomputer

According to insidehpc, Oracle has stopped developing Lustre and developers "have reportedly been encouraged to apply for other positions within the company".

A group of Lustre users already created OpenSFS on October 2010 to continue developing Lustre.

... consider the human brain -- a massively parallel 3D processing structure. The brain has an estimated processing power of 38*10^15 operations per second (according to this reference), while consuming about 20 W of power (reference)...

Good point. I believe I have solved Moore's Law in computing for some years. I need shovels, accomplices, and every Roger Corman movie.

First is heat. Volume (a cubic function) grows faster than surface area (a square function). It's hard enough as it is to manage the hotspots on a 2D chip with a heatsink and fan on its largest side. With a small number of z layers, you would at the very least need to make sure the hotspots don't stack.

I'm not saying your point is entirely invalid, however, heat isn't necessarily a problem if you can parallelize the computation. Rather the opposite, in fact. If you decrease clock frequency and voltage, you get a non-linear decrease of power for a linear decrease of processing power. This means two slower cores can produce the same total number of FLOPS as one fast core, while using less power (meaning less heat to dissipate). As an extreme example of where this can get you, consider the human brain -- a massively parallel 3D processing structure. The brain has an estimated processing power of 38*10^15 operations per second (according to this reference), while consuming about 20 W of power (reference). That is several orders of magnitude more operations per watt than current CPU:s have.

SGI claimed $526M in debt in their bankruptcy filing. However, Rackable apparently WON'T be assuming this debt! Go figure. Here's an explanation.

With multi-core processors becoming more and more prevalent, R's developers should remedy this as soon as possible.

Ask and ye shall receive

A bit of googling leads to this PDF which I got from this link. These are more descriptive than the ZFS slide.

Posted on InsideHPC blog, 48TB in 5U. 26TB isn't too much for this monster.