Europe's First Exascale Supercomputer Will Run On ARM Instead of X86 (extremetech.com) 40
An anonymous reader quotes a report from ExtremeTech: One of the world's most powerful supercomputers will soon be online in Europe, but it's not just the raw speed that will make the Jupiter supercomputer special. Unlike most of the Top 500 list, the exascale Jupiter system will rely on ARM cores instead of x86 parts. Intel and AMD might be disappointed, but Nvidia will get a piece of the Jupiter action. [...] Jupiter is a project of the European High-Performance Computing Joint Undertaking (EuroHPC JU), which is working with computing firms Eviden and ParTec to assemble the machine. Europe's first exascale computer will be installed at the Julich Supercomputing Centre in Munich, and assembly could start as soon as early 2024.
EuroHPC has opted to go with SiPearl's Rhea processor, which is based on ARM architecture. Most of the top 10 supercomputers in the world are running x86 chips, and only one is running on ARM. While ARM designs were initially popular in mobile devices, the compact, efficient cores have found use in more powerful systems. Apple has recently finished moving all its desktop and laptop computers to the ARM platform, and Qualcomm has new desktop-class chips on its roadmap. Rhea is based on ARM's Neoverse V1 CPU design, which was developed specifically for high-performance computing (HPC) applications with 72 cores. It supports HBM2e high-bandwidth memory, as well as DDR5, and the cache tops out at an impressive 160MB. The report says the Jupiter system "will have Nvidia's Booster Module, which includes GPUs and Mellanox ultra-high bandwidth interconnects," and will likely include the current-gen H100 chips. "When complete, Jupiter will be near the very top of the supercomputer list."
EuroHPC has opted to go with SiPearl's Rhea processor, which is based on ARM architecture. Most of the top 10 supercomputers in the world are running x86 chips, and only one is running on ARM. While ARM designs were initially popular in mobile devices, the compact, efficient cores have found use in more powerful systems. Apple has recently finished moving all its desktop and laptop computers to the ARM platform, and Qualcomm has new desktop-class chips on its roadmap. Rhea is based on ARM's Neoverse V1 CPU design, which was developed specifically for high-performance computing (HPC) applications with 72 cores. It supports HBM2e high-bandwidth memory, as well as DDR5, and the cache tops out at an impressive 160MB. The report says the Jupiter system "will have Nvidia's Booster Module, which includes GPUs and Mellanox ultra-high bandwidth interconnects," and will likely include the current-gen H100 chips. "When complete, Jupiter will be near the very top of the supercomputer list."
Pi-equivalents? (Score:3)
So, how many Raspberry Pi 4's would that be equivalent to?
Re: (Score:2)
Literally.
Jülich and Munich are both cities in Ge (Score:2)
Jülich Indeed has a superconputing center.
Munich probably has several, public and private business owned (one is named after the mathematician Leibnitz and is run by TUM, a university)
So which ohne is it?
Re: Jülich and Munich are both cities i (Score:2)
Re: (Score:2)
I'm pretty sure JA~1/4lich nor Ita^(TM)s are real places. I'm also pretty sure you both of you need to get better phones (or use a non-restricted computer), and that Slashdot needs to upgrade its SlashCode.
For the time being, use plain ASCII.
The CPUs are a sideshow (Score:2)
LinPACK is running on the GPUs ... CPUs are just there for the shittier programmers, who will only get a tiny fraction of the peak FLOPS out of the system.
Re: (Score:2)
In peak parallel or supercomputing times, the x86 families were ok, but not great. These were the wannabe CPUs. Times have moved on, but it seems the reason people use Intel x86/64 compatible chips for supercomputers is because that's all anyone knows anymore. Monoculture madness. The modern chips have problems in super computers: the CPU in your PC is actually a big stack of chips with multiple layers, there's a lot there that gets in the way of performance and it's there for compatibility, etc. Eve
Re: (Score:2)
Right remind me how much RAM there is on the best GPU's? Some problems just don't work well or even at all on GPU's because they have limitations that CPU's don't. If you actually worked in HPC you would know that and not make assine comments.
Re: (Score:2)
x86 is the shitness. ARM is the new hotness.
ARM is the old hotness. It's been around for 38 years.
Risc-V is the new hotness.
Re: (Score:2)
ARM is the old hotness. It's been around for 38 years. Risc-V is the new hotness.
If you're gonna talk about ARM in general, you need to talk about RISC in general - and RISC has been around since the 1960s.
Re: (Score:2)
x86 is the shitness. ARM is the new hotness.
ARM is the old hotness. It's been around for 38 years.
Risc-V is the new hotness.
The United States Government has decided that Risc-V is too "Chineses Friendly" to be utilized by anyone else. Therefore, it's hotness is not allowed here. Because I like to take all my tech advice from people born when computers were giant rooms filled with mathematicians.
Re: (Score:2)
"Chinese Friendly" sadly means it's not proprietary enough. It's a ridiculous argument, and it echoes of the old cold war days where open sharing of technical knowledge was often illegal. At least in the cold war the fears were more about the potential use of doomsday weapons and the like rather than mere trade protections.
Next up, the US government will warn against the use of Linux because it's being widely used in China, North Korea, Russia, and (gasp) Venezuela!
Re: (Score:2)
Next up, the US government will warn against the use of Linux because it's being widely used in China, North Korea, Russia, and (gasp) Venezuela!
The only reason this hasn't happened yet is most of Microsoft's lobbying money has gone to other pet causes. Once they get their ducks in a proper row, I expect this type of reasoning to start being spouted. Can't wait to get strung up over some of the articles I had in Linux Journal decades ago.
Re: (Score:2)
Irony, Microsoft once sent me a Linux book, free. I am suspicious that they thought I was an influencer or something, in the early 90s before that word existed. A time when they were much more Linux and Unix hostile than now that they use Linux extensively in their back end and support a Windows Subsystem for Linux
.
Re: (Score:3)
I think one of the reason is that historically, ARM has been the best performance per watt. x86/amd64 was the best performance one could get... well, with the exception of POWER and SPARC which I've not read about much [1]. With ARM improvements, server performance can be "good enough". Especially if the server is mainly I/O bound.
[1]: Would be nice if those architectures were still in the game.
Re: (Score:2)
"with the exception of POWER and SPARC which I've not read about much.
Would be nice if those architectures were still in the game.
Where is MIPS these days?
I've not heard anything about them in over a decade but I recall hearing their assembly language was supposed to be one of the best for low-level coding whereas x86 is a relative Brainfuck
Re: (Score:3)
Someone who better knows RTLs these days can correct me, but AFAIK, the actual assembly language doesn't really matter that much these days, due to caching, branch prediction, and other magic. I do wonder if starting with an older CPU makes it easier to learn assembly, versus just starting with a modern CPU architecture and teaching how to use that, because assembly language programming on x86 is way different than, say an ARM CPU, with little and big endian something to watch out for.
MIPS [eejournal.com] is now part of t
Re: (Score:2)
Assembler does matter if you want the performance. The libraries are going to be making use of it. But the biggest challenge in supercomputer has often been the "dusty deck problem", meaning how to get performance out of code that was not designed for the supercomputer. Ie; the code written in Fortran by an academic in 1970 and on a dusty deck of punch cards, the newer researchers don't understand the code and aren't really wanting to rewrite it all (being researchers and not highly trained programmers w
Re: (Score:3)
Where is MIPS these days?
I believe the Chinese Loongson [wikipedia.org] processors are based on MIPS (and recently have started to include some features from RISC-V). China even has a supercomputer based on those chips.
Re: (Score:2)
Where is MIPS these days?
Suffering for lack of investment, completely eclipsed by ARM, and with no reason to exist. SGI wasn't big enough to keep its development going BITD so it was showing its age even while it was still being put into machines. That's why SGI tried to go PC, but unfortunately for them they did it with Windows and that was an idiotic move. You have to go commodity and mass market to make money in that ecosystem.
You still see embedded MIPS, but it's always irritating when it happens, you're just wishing it was ARM
Re: (Score:2)
MIPS has chips out there, embedded SoCs and the like. But also some high performance computer chips as well. They're a bit like ARM in that they have design but not the foundries. I haven't actually used one in eons but I occasionally see them as competitors to an ARM or PowerPC system.
A lot of stuff is still out there, surprisingly. SmartCard chips based upon 8051 for example, running a Java interpreter which you'd think would be amazingly slow, and it is slow but it's just a front-end to a cryptographic
Re: (Score:2)
I'm sure it's somewhere. Just a few years ago I wrote a "visible" MIPS assembler and emulator for a Computer Organization and Architecture class a friend of mine was teaching. It only took a day or so. It wasn't a big RISC.
their assembly language was supposed to be one of the best for low-level coding whereas x86 is a relative Brainfuck
MIPS isn't bad, though you'll find a bunch of "pseudo instructions" (that compile to multiple instructions) to make writing in assembly less painful. The instructions themselves are otherwise very regular. There's not a lot of weird stuff there, unlike x86...
It is said that no one rea
Re: (Score:3)
x86/amd64 was the best performance one could get... well, with the exception of POWER and SPARC which I've not read about much [1].
SPARC was something of a special case in the same way ARM currently is. They were both the core of the Fujitsu system which is unusual in not having GPUs. What Fujitsu have is their own very very fast very very wide floating point unit, and a fast, low latency interconnect. Their schtick is bonding the Tofu interconnect to their FPU with a fast RAM bus all on one die. The perfor
Wonder when RISC-V will join the chat... (Score:2)
I wonder when RISC-V will be used in supercomputers. I know that there are a lot of new SBCs coming out with this CPU architecture, and even with backlevel Linux kernels and marginal software, they are showing leaps and bounds when it comes to performance.
Re: (Score:3)
I wonder when RISC-V will be used in supercomputers.
When it becomes better performance-per-watt than the competition.
Re: (Score:1)
I was wondering the same thing.
There are more than enough of these supercomputers to drive demand for a common processor that's geared especially for supercomputer applications.
It seems like it would be worth the effort.
Doesn't necessarily even need to be a RISC-V , a more specialized design focused on massively parallel operation seem like it could be a win.
or not, maybe there's just not that much performance to be gained.
still it would be worthwhile to create a RISC-V with a few features that would make s
Re: (Score:3)
Right now, supercomputers are made of GPUs. The CPUs are there mostly to control them.
GPUs are used because they are massively parallel, operating on the SIMT [wikipedia.org] model.
There already exist RISC-V CPUs with vector units as wide as the SIMD units on ARM and x86.
However, RISC-V's vector extension was designed to be scalable from the start, allowing the same code to run also on very wide vector units.
There has been talks about building GPUs based on RISC-V CPUs with vector units, as they are also capable of runnin
Does it compute? (Score:3)
Or are the ARM cores just there to feed the GPUs?
Re: (Score:2)
In the world of supercomputers everything matters. The GPUs, the interface to the CPU, the memory, the network. Pretending any one part is irrelevant because the GPU does the work is a good way of massively limiting your compute potential.
Re: (Score:2)
Pretending any one part is irrelevant because the GPU does the work is a good way of massively limiting your compute potential.
some great advice there for anyone not wishing to limit their compute potential
Re: (Score:2)
In the world of supercomputers everything matters. The GPUs, the interface to the CPU, the memory, the network.
Pretending any one part is irrelevant because the GPU does the work is a good way of massively limiting your compute potential.
When I said "Does it compute? Or are the ARM cores just there to feed the GPUs?" this is not speaking to relevance, drawing conclusions, making assertions or judgements.
It is a question about the typical expected role of these processors in this system. The article does not say.
As a bonus ... (Score:3)
Europe's First Exascale Supercomputer Will Run On ARM
It will also run OpenWRT and the site will have *really* good WiFi. :-)
And the CPU is from (Score:1)
France! Finally, Europeans buy European.
Re: And the CPU is from (Score:1)
...interesting. Of course, arm is British (japanese owned) so still European.
Re: (Score:2)
Rhea (Score:2)