Player three has entered Cray's supercomputing game: First AMD Epyc, now Fujitsu's Arm chips

Cray has said it will build a family of supercomputers for government research labs and universities. The kicker? The exascale machines will be powered by Arm-compatible microprocessors. The HPE-owned biz has partnered with Fujitsu to roll out the beefy big iron. Fujitsu will supply its homegrown A64FX processors – understood …

[Anonymous Coward]

Nominative determinism

“The most demanding computing work at Los Alamos National Laboratory involves sparse, irregular, multi-physics, multi-link-scale, highly resolved, long running 3D simulations,” said Gary Grider, deputy division leader of the HPC division at Los Alamos National Laboratory on Wednesday. “There are few existing architectures that currently serve this workload well.”

So the bloke in charge of their grid computing is called Gary Grider?

[Chairman of the Bored]

It's come a long way...

... since the humble Acorn RISC machine. Have a well-earned pint.

[Anonymous Coward]

Re: It's come a long way...

But will it run Zarch?

[TonyJ]

Re: It's come a long way...

Hmm even if it could run Zarch, I suspect I still couldn't play it all these years later! Just could not get the hang of that game.

[Anonymous Coward]

Re: It's come a long way...

Indeed.

I reckon that in another decade or two this will be run-of-the-mill and we'll be reading about some new cellphone running these processors on El Reg one day.

(if we don't blow ourselves up beforehand)

By that time, the Chrome web browser and other apps will be so bloated that the new phone will be just as slow as todays devices as the HBM RAM is trying to keep up.

[Chairman of the Bored]

Re: It's come a long way...

Bloat... Yes, software is a gas. It will expand to completely fill any volume.

A new hire observed my gray hair and similarly aging HP-12C and -48S calculators and started yapping about the glories of modern software. I told him, "Yes, when I started this game in the 80's, PCs were new. Really new. The only real applications I had were games, word processor, database, spreadsheet, presentation software, and a comms package so I could get my email and hit the bulletin board service - kind of like a web browser, but without the decent pr0n"

And he starts saying, "Uh, today we now have. Hmm. Better pr0n?"

"Definitely"

[defiler]

Not even a codename

And yet it's shipping next year?

It's got to be tough in the supercomputer game. For something so complex to be developed in such a short time. A new car might take 10 years to develop, but if it takes 10 months to debut your supercomputer then it's hopelessly outclassed all over again.

Wow.

Also, yay ARM - I saw an interview with Steve Furber who said that the only reason they developed the ARM processor at Acorn was because they know somebody else (one of the big CPU players) would build a RISC chip, so they wanted to understand all the ins and outs of them so they could make an informed choice. And before anybody (Motorola, Intel, National Semi etc) had stepped up, Acorn had working silicon. So they ran with it.

[Korev]

Re: Not even a codename

These computers take years to develop. Cray and competitors will have a huge amount of unreleased chips to play with and access to roadmaps and specifications that you and I would never be allowed to see.

[Jc (the real one)]

Re: Not even a codename

They are available now to the right people - we have them running in our development lab

Jc

[amanfromMars 1]

Re: Not even a codename

They are available now to the right people - we have them running in our development lab .... Jc (the real one)

What sort of right people, Jc? Are some Sectors in Lab Developments by ESPecial Delivery Personnel Invite? Such is akin to an Immaculate Capture with Captivating Potential to Plunder.

Novel Noble Sectors of Outstanding Particular and Peculiar Interest to Much More than just a Self Anointed and Appointed Few.

El Reg .... That is white hot news .... free to share and air.

Don't worry about putting the proverbial spanner in global works with the product, it is just what IT needs to Quick Start SMARTR AI Progress. And if ever questioned about matters beyond a current present comprehension, refer all inquiries to source, for that is where all the answers certainly reside.

[defiler]

Re: Not even a codename

Yeah, I guess, but at the same time that means you're developing against somebody else's speculation. And they're developing at breakneck speed to match a third party's speculation.

It's not unlikely that one group's speculative plans will hit some kind of wall that holds them up, or means they have to change course, so then everyone leaning on them has to respond in kind. I have no doubt that the job gets done, but it's still impressive and must be pretty high-pressure.

[StargateSg7]

Re: Not even a codename

We got our Super-Workstation/Super-Server chip design for you to peruse when its released soon enough...

128-bits Wide Combined General Purpose CISC chip with a COMBINED set of CPU/GPU/DSP/multi-value SIMD Vector processing/full MIMD Array Processing cores for Signed/Unsigned 128-bit and 64-bit/32-bit/16-bit/8-bit/4-bit Integers, Floating Point and Fixed Point Values, and separate 32-bit/64-bit/128-bit RGBA, YCbCrA and HSLA pixel math running on GaAs (Gallium Arsenide) substrates starting at 60 GHz (soon up to TWO THz!) clock speeds.

Then on the SECOND part of the super-chip is a DEDICATED hardware-based extended-state boolean logic chip array processor with weighted results including:

ABSOLUTELY_TRUE = 100% certainty to the positive

LIKELY_TRUE >= 67% certainty to the positive

POSSIBLY_TRUE >50% certainty to the positive

IS_EITHER_TRUE_OR_FALSE = Split decision (could be either one!)

IS_NOT_TRUE_AND_NOT_FALSE = non-decision (is neither one!)

IS_BOTH_TRUE_AND_FALSE = special decision (is BOTH true and false at the same time)

POSSIBLY_FALSE <50% certainty to the negative

LIKELY_TRUE <= 33% certainty to the negative

ABSOLUTELY_FALSE = 0% certainty to the negative

INVALID_RESULT = error code 1

RESULT_IS_INCONCLUSIVE = error code 2

ERROR_DURING_CALCULATION_OF_RESULT = error code 3

NO_RESULT_IS_AVAILABLE_OR_CONTEMPLATED = error code 4

STILL_WAITING_FOR_RESULT = Status code 1

RESULTS_NOW_READY_FOR_USE = Status code 2

RESULTS_HAVE_BEEN_ALREADY_USED Status code 3

SKIP_TO_NEXT_RESULT = Status code 4

SKIP_TO_NEXT_RESULT_AND_COME_BACK_LATER = Status code 5

IGNORE_CURRENT_RESULT = Status code 6

The above is IDEAL for processing neural net-based and expert system applications where comparisons and decisions are not always black and white and have Shades of Grey!

.

We've had this super-chip design for YEARS and now we are at Full Tape-Out stage. Since GaAs is printed at around 280 to 400 nm line trace widths, it's a tad easier (if a bit slower!) to etch the entire circuit with multi-electron beam etchers. And to get to 60 GHz clock speeds we just up the voltage and current.

Doping the substrate has always been the GaAs substrate's downfall over CMOS/Si process but I thing we've finally got it right these days!

.

Coming soon to a Best Buy and Amazon Store NEAR YOU!

.

[Tom 7]

Re: Not even a codename

Yay to ARM. Now when can we get a class action for the sale of ARM when it should be obvious it was worth a shit load more when it was sold.

[Peter Gathercole]

Re: Not even a codename

Um. The development is Fujitsu silicon. Fujitsu are an ARM development license (or whatever it's called) holder, so are free to take ARM ISA, core and ancillary execution unit designs and develop them into products.

ARM will be getting it's money from the maintenance of the license, and possibly a small royalty for every core used. But if these SUPEs sell well, they won't get anything more than the license and royalties, so it will not significantly increase their financial worth.

A company like ARM lives more on the value of it's IP rather than the number of devices sold, and this is decided as much by the market as device penetration.

I did not like the Softbank purchase of ARM much myself, but the sale was a result of the shareholders cashing in on their investment, so you can't really complain about the value returned, they obviously were happy at the time.

[fpx]

Multi-Physics

I thought there was just a single set. Anybody can clue me in?

[Mike 16]

Re: Multi-Physics

Well, there are at least Classical Physics (Newton), Biblical Physics (Joshua), Quantum Physics (Barnum), and Political Physics (your choice), off the top of my head.

[Robert 32]

Re: Multi-Physics

HLMWTFY:

"Multiphysics is defined as the coupled processes or systems involving more than one simultaneously occurring physical fields and the studies of and knowledge about these processes and systems."

"In a broad sense, multiphysics refers to simulations that involve multiple physical models or multiple simultaneous physical phenomena."

"The physics refers to common types of physical processes, e.g., heat transfer (thermo-), pore water movement (hydro-), concentration field (concentro or diffuso/convecto/advecto), stress and strain (mechano-), dynamics (dyno-), chemical reactions (chemo- or chemico-), electrostatics (electro-), and magnetostatics (magneto-)."

[Chairman of the Bored]

Re: Multi-Physics

Good textbook definition. Let me give a concrete example: take high power RF system design.

The actual RF propagation in a system can be handled in a number of ways: direct evaluation of Maxwell's equations, Method of Moments, Finite Difference Time Domain, etc. One output of such calculations is absorption in dielectrics and current densities in conductors. These give rise to heating...

Heat transfer will involve a finite element model (FEM) to determine steady-state temperatures, cooling requirements, etc. Given material mechanical properties and loads, FEM will also tell you how much your structure will distort...

And that changes the boundary conditions on your FDTD electromagnetic sim... Maybe you have to change the design and re-do everything, but regardless you have to re-compute the energy deposition. If it's really high power I might do particle-in-cell codes to figure out whether air or other gas will break down.

And so it goes. Iterate, and try to succeed before you blow through your budget. Make sure the system is affordable and manufacturable. (Monte Carlo over mechanical tolerances... Do not specify pure unobtanium...) The ultimate multiphysics tool will do it all. That doesn't exist. Instead you need specialists ... Usually a team of RF people, mechanical engineers, systems engineers, HVAC, etc ... armed with many tools and a big, scary, expensive, fast freakin' computer. That's real-world multiphysics

[StargateSg7]

Re: Multi-Physics

May I suggest you build your RF or MM-wave cavity resonators and waveguides out of Aluminum Oxide Ceramic and COAT THEM with thin-film layers of a GOOD reflector metal (Beryllium and Tungsten Carbide is best but high nickel content Stainless Steel will do too!) Any decently sized vacuum chamber can be used to do a multi-layer vapour thin film metal deposition onto ceramic surfaces.

The AL2O3 ceramic resonator/waveguide can be formed in a cheap mold and baked in a high-temp kiln with basically NO CNC MACHINING NEEDED. You just need to coat it with metal!

https://www.ceramtec.com/ceramic-materials/aluminum-oxide/

..

That's MY SUGGESTION !!!!

[Ken Hagan]

Re: Multi-Physics

Just sounds like physics to me. At a pinch I'll let you have "physics, without the unjustifiable over-simplifications".

[Chairman of the Bored]

Re: Multi-Physics

Ken,

Quite right. The way I look at it is that Physics is the general case of science. Mathematics is the language of physics. All other hard sciences (bio, chem, electronics, mechanical...) are special cases or simplifications. Unfortunately non-trivial problems are too difficult to handle in their entirety, so must divide problem domains into understandable chunks. Integrating these chunks together is the role of the systems guy/gal

[AJ_Newman]

RIKEN and Co have pulled out all the stops to enable the ARM software ecosystem to scale to proffer high efficacy exaflop capability.

Their extensible Vector processing (128-2048 bit SVE) extension that to the ARM ISA has already proven is viability for existing frameworks.

If the A64FX is able to reduce the electrical power by 3 times compared with the incumber x86 Supercomputer offerings - this will offer Fujitsu a long outlet for their existing and future chip designs.

ARMv8's latest matrix multiplications, and 3d party extensions pave the way for Fujitsu to push the envelope with lower electrical power, more functionally rich cores at 5nm. 3D CPUs and local memory on proximal SoC layers will benefit from a core that is uses less power than Intel and AMDs current offerings.

With Microsoft tapping less capable 'ARM server' designs, and Amazon pouring money into their home grown 'singularity - I wonder if the general availability of Fujitsu's A64FX will dent Caviums or Ampere's dreams?

It will be fascinating to see how Apples (2020) A14 cores compare - especially if MacOS gets its first taste of ARM.

AJ

[John Smith 19]

Instruction set design is tricky.

Too specialized and you only sell one computer.

The challenge at the instruction level is to abstract out the key enablers of a large set of problems (ideally multiple sets) your customers would like to solve so they can be put together in different ways.

That means talking (and continuing to talk) to your target customers before you start design.

Kind of like the thinking behind Unix. Well focused tools that can go together in lots of different ways.

[John Savard]

Vectors Are Good

Well, as I believe that the Cray I was a good architecture for supercomputing, and thus I was excited about the NEC SX-Aurora TSUBASA, I also heartily approve of Scalar Vector Extensions, and I'm glad to hear they're getting more widely used.