First RISC-V computer chip lands at the European Processor Initiative The European Processor Initiative (EPI) has run the successful first test of its RISC-V-based European Processor Accelerator (EPAC), touting it as the initial step towards homegrown supercomputing hardware. EPI, launched back in 2018, aims to increase the independence of Europe's supercomputing industry from foreign technology …
You know, I'm beginning to think the best thing the current owners of ARM can do with it is flog it to the Chinese at peak share price, before the uninitiated realise that RISC V is about to kill its business, the way Linux killed UNIX. And the best thing UK Gov can do is let them.
Its rather interesting that. ARM has long evolved beyond the CPU in an effort to ensure this didn't happen and offer more value. They do all the IP nowadays, except radios, even the cell libraries of gates and on-die memory generators/macros for many foundries.
So if I was a smaller company that had skills in analog/power or video decoding, or radios, I might buy the CPU, bus, memory, debugger and IO from ARM, and the digital cell library to focus on my strengths.
Another product that ARM offers is risk free, high-performance CPU+cache clusters prevalidated and laid out on many foundries. Just because you have licensed the IP doesn't mean you can read the 10,000 pages of docs, configure it, place it, and verify it without risk - with ARM just hand over $$$ and you know what you are getting.
Time will tell if this evolution will be enough to enable it to compete with RISC-V (which doesn't even have fixed interrupt controller specifications, debugger systems, etc. but they are getting there fast).
Whilst this is clearly a test chip, if it works it could make an interesting SoC for embedded use with all those EU developed accelerators on it (even if they're weak), think Arduino or Pi Pico type devices. 22nm is cheap, it's a small die, so whilst it's not as small and cheap as some of those other devices, it might be more powerful if it does reach that 1GHz speed.
> China, meanwhile, is hard at work on a high-performance RISC-V chip family called XiangShan, and plans to deploy 2,000 RISC-V laptops by the end of next year.
China is also pursuing the Loongson CPU with the "LoongArch" ISA.
Their current offering is a 4 core 2.5 GHz 3A5000 and a 16 core up to 2.5 GHz 3C5000L.
And the internal code names are interesting 'Loongson Zhongke is known as the "first domestic CPU", mostly adopts the MIPS architecture and successfully developed the LongISA instruction set by itself. In late July, Loongson Zhongke launched two processors, 3A5000 and 3C5000L. The code names of these two processors are also quite interesting. The former is "KMYC70"-"70th anniversary of the War of Resisting US Aggression and Aid Korea", and the latter is "CPC100"-"100th anniversary of the founding of the Party". It is really red and specialized, and it is worthy of being a 100% localized Chinese "chip"' - (ref: https://inf.news/en/tech/d914953afc8cf73cff0bff16fd07703b.html )
EPI, launched back in 2018, aims to increase the independence of Europe's supercomputing industry from foreign technology companies
If you did s/Europe/China/, a reader probably wouldn't notice a difference.
The EPI has been funded as part of the European Union's European High-Performance Computing Joint Undertaking, EuroHPC – of which the UK is not a member. Its membership roster is a who's who of European technology firms and academic institutions spanning 10 countries and including France's Atos, the Italian arm of STMicroelectronics, Germany's Infineon and Fraunhofer-Gesellschaft (FhG), the BMW Group, the Barcelona Supercomputing Centre (BSC), ETH Zürich, Instituto Superior Técnico, the University of Zagreb, and the Foundation for Research and Technology (FORTH) in Greece.
So the European tax payer is forced to subsidise big corporations. I think when Karl Marx thought of redistribution of wealth, he meant it the other way around.
Anyway, the problem with that is, these kind of technologies are developed by privileged people and organisations. If you wanted to develop such technology with your small business, you'll get most of your profit taken away to fund your potential competition. All in the name of greater good :-)
Chinese model unlikely going to work in the EU, unless they stop pretending and just no longer bother with elections. Now when they plan something, there is still a little bit of clouding over their heads that the may not be in the next draw and whoever comes next may scrap their work.
Neither what you wrote, nor the post responded to…..
You don’t know whether he is American. Anyway, I’m not, I’m from the U.K.
I also think that it is foolish for the EU to subsidise this project. But honestly, it’s their money. I don’t see any problem *in principle* with government investing in infrastructure and “goods for the people”. Governments of many countries have historically shown both good judgement and effficient execution, despite the neocon propaganda. Nearly half of all the U.K.-originated FTSE companies have their roots in something that government built. ARM is one of them. ARM wouldn’t exist without investment by an effective arm of government in the BBC microcomputer, intended to educate children.
However, it so happens that the RISCV and EPI projects are both rather useless, and over fifteen years behind the times. This is the problem I have: the EU is famously poor at identifying sensible use of resources. And this isn’t a coincidence. The EU *always* considers whether something is strategic and/or politically expedient, without any attention at all to either technical side or execution. Often self-consciously so: in engineering projects, the stated political goal is usually “development of capability”training of newly qualified engineers, and basic proof-of-concepts. Not usable products, nor even credible roadmaps. This is a screaming problem.
Not entirely off-topic, have a look at the whole “side of a bus, £15bn gross vs £8bn net EU contribution” thing. I am a Brexiter, and I would have been absolutely happy if our contributions had been £15bn net. Or £25bn net. I just don’t care. What I care about is *whether the money was well-spent so the investments benefited the UK*. It’s not a zero sum game. You can’t claim we’ve only given £8bn, if we received the £7bn as Taylor Swift CDs. I want to see some evidence that projects initiated by the EU have been of some value.
I’ve worked on dozens of EU projects, and managed about ten at a senior level. I have *personally* overseen expenditure of well over £35m total on those projects. I estimate the outcome benefit of the sum of it all, ten years later at….ummm, zero. I can’t find any record at all of anything that came out of any of them. Even the webpages are dead links.
We hit all our KPIs, exceeded our deliverables. But the product that the EU contracted us to build to build was useless. Not a single user out in the real world. Not. A. One.
@Justthefacts
"I can’t find any record at all of anything that came out of any of them.".
I am not all that convinced about the honesty of your comment but could it be you should have worked a bit harder or perhaps searched a bit harder.
PS. Britain was a part of the EU, perhaps you should look at the British failing more so than those 27 other memberstates.
Perhaps the real failure the "EU" made was really to let the British with guys like you waste money like that achieving nothing.
PPS. what makes you think RISC-V is useless, and over fifteen years behind the times.
PPPS. That comment was as dumb regardless if he was American or British or whatever.
And with your knowledge what would you recommend in stead.
EPI is *at least* fifteen years behind the times.
Here’s an article from only last year
https://www.nextplatform.com/2020/01/27/european-processor-initiative-readies-prototype/
Amongst other points, you will see that this test-chip was supposed to tape out on TSMC 6nm. It actually taped out on 22nm. That’s *three silicon generation slip* just within the last year.
It’s a supercomputer test chip with a performance of one teraflop, when I can buy an NVidia A100 with 150 teraflops.
You need to understand what EPI project looks like in practice. *Forty* partners, each with their own little bit of IP to push. Here’s one: Kalray.
https://www.kalray-bourse.com/fr/
Formed in 2008, funded by various French R&D contracts and “national champions” like Safran.
And in thirteen years, not one single commercial product design win. Zero.
And the EPI is a *test chip*. No chip in the history of ever has worked fully correctly on the first spin, and it takes a lot of time and money to get the yield to anything economic. So, there’s at least 2-3 years between tapeout of prototype to full volume production, which is what NVidia A100 is. And that’s for an experienced commercial company like Intel, NVidia. ARM. I would be gobsmacked if EPI can do that in less than 4 years.
@Justthefacts
So what, do nothing because somebody is ahead, or try to catch up.
Buy from abroad or try to produce a more domestic industry to buy from and sell to the rest of the globe.
What would have come of Airbus (or China) if they had thought like you.
Are you actually revealing the history of Britain without thinking of it.
No. Absolutely not “do nothing”. If you read my initial post, you will see I am very *pro* government intervention and investment. Historically the U.K. did very *well* in that.
Unfortunately for ideological reasons and propaganda, sometime in the seventies a bunch of people began yelling “British Rail bad……British Leyland bad”, and we stopped.
One of the big problems that the U.K. has, is not having a U.K. equivalent of French COFACE, nor maintaining golden shares of pseudo-privatised companies…..like “Airbus” which if you actually spend several days tracing and trawling the ownership structure you will realise is basically just an arms-length vehicle for the Elysee Palace. I did, because I used to be employed by them, “owned shares” and wanted to know exactly what they were shares in. So I think Airbus is an *effective* use of government control.
The problem is that the Commission is just very, very bad at steering industrial policy. Because they have zero technical knowledge, and take *pride* in it. Unlike, for example, the French civil service who are mostly educated at Ecole Polytechnique and Sorbonne, and are to my experience very smart. Go talk to a German, and they will almost uniformly tell you Ursula von der Leyen was sent to the EU because she was such a bad Minister of Defense in Germany and they had to get rid of her. Literally, that’s what Martin Schulz said. She has a medical degree, which her university *convicted* her of plagiarising over 60% of her thesis, but still failed to strip of her degree.
No, I do *not* think you should try to “catch up”. That is doomed to fail and just dumb. You should *innovate in adjacent technology*. Make other things that everyone else needs to buy, and trade them. The world is *not* a zero sum game. How many times did we try to tell Trump that. You *dont* compete on ground that enemy holds, and *they aren’t the enemy*.
Just a shortlist - nowadays most of the silicon system power budget goes on the interconnect. Figure out how to make low-power optical interconnect on silicon and the world will beat a path to your door.
Most of the silicon area goes on cache which is SRAM. An SRAM cell is 5T. Figure out a way to cross-code an 8-bit byte in 256 states of something less than 5x8 transistors and again….the world will beat a path to your door. There’s just *dozens* of unsolved problems in silicon. “Catching up” shouldn’t be your metric.
Hold your horse , let's check the facts , @Justthefacts :
1. NVIDIA A100 have only ~ 10 TFlops DP (linpack tested -you can use your google-fu...)
2. ARM doesn't tap chips ...they license chip design (broadly speaking)
3. 22nm , 6 nm, 3 nm, 7 nm ..these are only marketing gimmicks- number 4 in TOP500 (june 2021) is Sunway TaihuLight - and Sunway SW26010 is using 16nm FinFET - aka like 22nm or more for TSMC process , and is still no. 4 after 5 years --i think THIS is pretty impressive..
3. Yes is a test chip but at the same time an "technology demonstrator" for Exascale , like Mont-Blanc 2020 demonstrator (based on ARM ) before EPI.
4. EPI is just a block needed for 100% EU internally produced Exascale and post-Exascale HPC systems
5. You are from UK ,not EU , is their money to spend...
6. As for UK technology this the same as UK Empire --"was"
1) is true, but rather irrelevant. Yes I get that it has been traditional for supercomputing to only count DP flops. But that’s fighting last century’s battle. Todays applications mostly want high *half precision* flops. And that’s what NVIDIA is doing.
2) True. So what? I didn’t say “ARM always ra-ra-ra”. I said: horses for courses, and currently there is a clear market leader in every segment. Some it’s NVIDIA, some it’s ARM licensees like Qualcomm,, some it’s microcontroller cores like ATMega, some it’s DSPs from NXP. Question is, what is the market segment where RISCV can actually win? I can’t see it.
3) I’m aware that nm is a metric that can be gamed. However, *this* 22nm is simply very much less dense in mm2 than TSMC6. That’s just a fact.
Leading TOP500…isn’t a bragging right worth winning. What matters is what can actually be achieved by all those flops. Google TPUs, oil exploration supers, datacentre servers etc. Or for that matter, the lowest power embedded processor sipping less than a milliwatt.
.
3 again, mis numbered) Tech demonstrator…..demonstrate for as much as you like. NViDiA A100 also had to be prototyped. That was done two or three years ago, internally without needing a press release. Doesn’t affect that EPI is a prototype from 2005 here in 2021.
4) I don’t understand your sentence. But what I do understand is that you are very keen on having something 100% EU produced. And you are just *wrong* to want that. It’s a bad goal. It’s not achievable, but also it’s not a benefit if you could do it. It’s the opposite of Economics 101. Trade is good, because you leverage the relative advantage of each region. If you pursue complete independence from the world, you are Cuba. Doesn’t matter how smart you are, or how far ahead you are when you start, after fifty years of proud isolationism that’s where you will end up.
This is one of the most fundamental things that EU misunderstands about Brexit. It’s the opposite of “little Englanders”. EU thinks a market of 400 million people is big, we think it’s small.
5) Indeed it’s their money. Hurrah!
6) I don’t know what you mean. Obviously the British Empire is no more. And I’m glad about that.
Worth mentioning that for example Belgium’s empire was pretty damn unpleasant. King Leopold the genocidal maniac should never be forgotten. Things haven’t changed that much in #BrusselsSoWhite
London isn’t perfect either.
What’s your point?
I don't want to open a flame war but you asked my point(s) so here we are:
1. DP is relevant because most part of HPC applications need DP. Is true that AI is somewhat tailored for other "flops" but right now , and in foreseeable future (let's think 10-15 years) DP flops will be still relevant.NVIDIA, AMD,Apple, and even Intel add AI-specialized cores (tensor and whatever name) "on top" of normal DP processing not "instead" of DP processing.
2.I was a proponent of ARM in datacenter from a long loooong time , i even used Calxeda and Gigabyte ThundeX2 kit so i am agnostic and pragmatical about ISA's. For me is always right tool for the right job.
So about RISCV my gut feeling (and my technical acumen) tell me this is a good opportunity to truly threatens supremacy of X86 ISA in datacenter on medium term (5-10 years).
3. Top 500 is not a perfect benchmark , no benchmark is perfect and can't be , but still is THE benchmark for HPC suppers....so is still relevant as a baseline.
My point was: Sunway TaihuLight is still on top after 5 years even with older tech..
As for oil/gas/nuclear- private and public suppers - all of them are "calibrated" using HPL Linpack AND other benchmarks . Acceptance tests on super require using HPL Linpack. I installed more than 10 suppers (as a fact next will need to pass acceptance test in a week) so i am more than familiar with acceptance procedures.Even for datacenter servers for critical apps , procedure are to run burnin' tests -yes, linpack based -even Intel have an official one :) , but not for testing raw performance , just a stress-test.
3 bis (my mistake) Please be exact. "The initiative started in 2015, in the aim to produce an exascale supercomputer by 2023. The first phase of the project started in December 2018.[3] In the summer of 2019, the basis of the architecture was decided.[4] In January 2020, the first prototype was presented.[5]"
this is a direct quote from wikipedia page.
So , they have a real working prototype (even on 22 nm GF process) 2 years after deciding on architecture ISA ? -color me impressed , because this is not NVIDIA , or Samsung or a single vendor -this is a large consortium with a lot , and I mean A LOT of organizational inertia.
4.EPI was from start aimed to be a part from "First Completely Open Source European Full-Stack Ecosystem Based on New RISC-V CPU".
Please note: Open Source and Full-Stack.
I don't know about similar project from USA, China, Japan, SK or UK for that matter....
To accuse me (and EU) of isolationism is too much ..
To be able to be self sufficient ( in case of trade wars , broken supply chains , other geopolitical troubles) is just a good EU strategic decision.
5, So we're on the same page.
6.My mistake to for bringing up Empire you baited me with your political comments about France, Germany, Schulz. der Leyen .. , comments which are irrelevant to matter on hands : EPI
Even so , my point is : EU is trying something , UK not so much.
Combining your points 1,2 & 3
There's a real difference whether you are targeting commercial datacentre or academic science.
Datacentre doesn't care about DP, and is something where you might want to invest, as it has a commercial return. Science legacy codes require DP, but don't have a commercial return. FWIW, I assume you are aware that RISC-V FORTRAN toolchain isn't currently supported.
But that performance is all sitting out in the matrix accelerators, nothing to do with RISCV/ARM/x86, and it purely depends on what mix of operations you configured/optimised on that individual chip.
Plus there's the dirty secret that, independent of RISCV / ARM / x86, most science legacy codes can't use the vector accelerators let alone matrix ones. There was a paper where they tested the top 50 science codes, and discovered that I think only ten of them were written to even use BLAS.
I simply disagree on your point "this is impressive because a frankly non-commercial consortium is doing something to compete with the big boys". It may make people feel good, but it isn't a good use of their skills. If this were my job, I would be doing three things, that academic supercomputing types are uniquely placed and skilled to do:
A) Carefully analyse various science codes, without preconceptions, and identify what the fundamental limits are. Design and make custom accelerator IPs for those codes. The return on this, for the top twenty codes, easily exceeds anythng you could possibly achieve in making datacentre-like supercomputers. Japan does it.
B) Figure out ways to automagically optimise memory access; cache prediction etc. Break the von Neumann bottleneck.
C) Bring software resouce in re-factor decades-old legacy codes from an efficiency perspective. Innovate ways to automatically extract algorithmic requirement & unit-tests from existing code to avoid breaking it. *This is such a hard problem, and the rewards are so great for solving it, both immediately to the academic community and to wider society*.
"First Completely Open Source European Full-Stack Ecosystem Based on New RISC-V CPU".
Other countries aren't doing similar, because we assess that this is a Bad Goal.
My analogy is "First Completely Open Source UK Full-Vehicle Automotive Ecosystem, Based on New Armadillo Electric Motor".
Do you see the problems? We both agree that good electric motors are going to be key in the strategic market of electric vehicles, and it would be economically devastating if geopolitics stopped us from getting them.
#1 Armadillo (if they existed) don't make the best electric motors in the world, and while UK can certainly make good electric motors we don't have any unique expertise.
#2 Other motor manufacturers are already in the market, and theirs are perfectly OK, although better ones are continually being designed.
#3 While cars do have design flaws, there is no particular evidence that open-sourcing car design would fix that. And believe me there are *thousands* of engineers who absolutely would spend their weekend optimising Porsche Taycan aerodynamics for free. Fiat Multipla, not so much. Can you see the relationship to supercomputer chips, by the way?
#4 We all know that the designing a car is 90% not about the engine. Having only a good engine is not going to make you a world leader. It's far better to assemble the car than make all the parts, and in fact that's what the major car companies are - assembly plants.
There is this link to Nextplatform from 2018.
The article ends like this.
"Time will tell if Europe will be able to pull off the massive hardware and software development necessary to field a true exascale system. An interesting parallel to the EPI effort is Airbus, another multi-country European collaboration. Airbus has managed to succeed over time versus tough competition from Boeing, selling more single aisle planes than the American company and, more importantly, doing profitable business in a cutthroat industry. EPI is an Airbus-like collaboration addressing even bigger markets, can they engineer the same success?".
https://www.nextplatform.com/2018/10/24/the-plan-for-europes-homegrown-exascale-hpc/
@Gordon Shumway
Interesting question, caught up with what?
You might know that on the whole it's China that is catching up with the west and surpassing too in many fields.
To quote CIA on China.
"Measured on a purchasing power parity (PPP) basis that adjusts for price differences, China in 2017 stood as the largest economy in the world, surpassing the US in 2014 for the first time in modern history. China became the world's largest exporter in 2010, and the largest trading nation in 2013.".
https://www.cia.gov/the-world-factbook/countries/china/#economy
@StrangerHereMyself
"RISC-V is an open standard instruction set architecture (ISA) based on established reduced instruction set computer (RISC) principles. Unlike most other ISA designs, the RISC-V ISA is provided under open source licenses that do not require fees to use.".
https://en.wikipedia.org/wiki/RISC-V
PS. No engines were invented in the USA.
The Register 
Biting the hand that feeds IT
Copyright. All rights reserved © 1998–2024
