AMD Now Powers 121 of the World's Fastest Supercomputers (tomshardware.com) 22
The Top 500 list of the fastest supercomputers in the world was released today, and AMD continues its streak of impressive wins with 121 systems now powered by AMD's silicon -- a year-over-year increase of 29%. From a report: Additionally, AMD continues to hold the #1 spot on the Top 500 with the Frontier supercomputer, while the test and development system based on the same architecture continues to hold the second spot in power efficiency metrics on the Green 500 list. Overall, AMD also powers seven of the top ten systems on the Green 500 list. The AMD-powered Frontier remains the only fully-qualified exascale-class supercomputer on the planet, as the Intel-powered two-exaflop Aurora has still not submitted a benchmark result after years of delays.
In contrast, Frontier is now fully operational and is being used by researchers in a multitude of science workloads. In fact, Frontier continues to improve from tuning -- the system entered the Top 500 list with 1.02 exaflops of performance in June 2022 but has now improved to 1.194 exaflops, a 17% increase. That's an impressive increase from the same 8,699,904 CPU cores it debuted with. For perspective, that extra 92 petaflops of performance from tuning represents the same amount of computational horsepower as the entire Perlmutter system that ranks eighth on the Top 500.
In contrast, Frontier is now fully operational and is being used by researchers in a multitude of science workloads. In fact, Frontier continues to improve from tuning -- the system entered the Top 500 list with 1.02 exaflops of performance in June 2022 but has now improved to 1.194 exaflops, a 17% increase. That's an impressive increase from the same 8,699,904 CPU cores it debuted with. For perspective, that extra 92 petaflops of performance from tuning represents the same amount of computational horsepower as the entire Perlmutter system that ranks eighth on the Top 500.
Different requirements (Score:4, Interesting)
Supercomputers have different requirements to most other machines.
Performance per watt and per Euro are the most important things. They are interlinked too, since electricity costs to run these things tend to be considerable.
It's interesting that AMD is doing so well here. Their parts are leagues ahead for Intel's for efficiency, but also ahead of any ARM parts that are generally available too.
Re:Different requirements (Score:4, Interesting)
Performance per watt is generally important across the board. However it's currently not possible to build a supercomputer on a single chip, so where supercomputers differ is on having large numbers of high speed processors interconnected and able to work together efficiently on whatever workload the system is designed for.
The latter point is also important, supercomputers are generally built for a specific purpose and the workload may suit one type of processor or interconnect over another.
Re:Different requirements (Score:4, Informative)
> supercomputers are generally built for a specific purpose
Yes and no.
Some are (weather forecast, CFD). But many supercomputers are owned by academic institutes which can have a wide variety of users and workloads.
And then, there's the military. You'd be a fool to believe the benchmarks they run on their supercomputers (and are given to the hardware vendors) match the programs they actually run ;-)
Re: (Score:2)
Re: (Score:3)
I don't think they are ahead of arm parts: they rely on GPUs which give much more FPU performance per watt and higher memory bandwidth. The ARM machines do it all on the CPU. They are quite exotic A64fx processors with HBM2 memory and on board networking however.
In terms of generally available: well if you're in the market for a supercomputer, Fujitsu will sell you one!
Re: (Score:3)
Solaris had some outstanding innovations, like their version of troff was light-years ahead of everyone else's
I literally lol'd here, thanks for that.
Nothing stops anyone from using Solaris to build clusters, nor indeed has anything ever done so. My first sysadmin job involved a distributed processing system for running Verilog, SPICE, Cadence, and other related tools in an office full of SparcStations, where users sat at all the machines...
Re: (Score:2)
Supercomputers have different requirements to most other machines.
Performance per watt and per Euro are the most important things.
These are the most important things in personal computing, too, especially since so much of it is mobile now. But performance per watt matters even on the desktop, because it lets you make the computer smaller and cheaper. Performance per currency unit used to be far and away the most important statistic, but another thing which changed that is that even a budget CPU is really fast now by reasonable standards (i.e. "lets me do the stuff I want to do in a timely fashion".)
Re: (Score:2)
It's interesting that AMD is doing so well here. Their parts are leagues ahead for Intel's for efficiency, but also ahead of any ARM parts that are generally available too.
AMD produces both CPUs and GPUs. It should be noted that AMD Epyx CPUs have done really well on the Top 500 far better than the AMD Instinct GPUs. Most of the AMD CPU systems use Nvidia GPUs. However, it should be noted that AMD GPUs have done quite well near the very top of the list.
And in 30 years... (Score:4, Interesting)
Re:And in 30 years... (Score:5, Informative)
Will that be the case? Yes! I thought no but I was wrong!
1983, 40 years ago, the state of the art was, well, high end workstations 4MHz 68000, at an astounding price. More personal machines might use a 65C02 or 6809 for a price. In practice many still had the 6502 or Z80. The Cray X-MP was the top supercomputer at 800MFlops (was a 4 CPU version made? by then)
1993, 30 years ago, the state of the art PC was the 66MHz pentium. There were faster machines like a POWER 1 workstation from IBM or a DEC Alpha. Useful if you're feeling the weight of your wallet. Any of those utterly destroy the competition from 1983. Top supercomputer was 143 GFLOPS.
2003, 20 years ago, the state of the art PC was the brand new Opteron 64 bit CPU on the $$$ end or more like a 2.2GHZ P4 or similarly clocked Athlon XP. Either way hilariously faster than the OG Pentium. Even a single core opteron would beat the Cray from 1983, and it went up to 4 way. Certainly vastly faster than the 30 year old CDC 7500 at 35MFlops. Top 1 is 35TFlops (peak).
2013, 10 years ago. The state of the art PC was a 3.5GHz Core i7 Ivy Bridge, 4 core, or maybe the much maligned AMD FX 8350. You could get up to 64 AMD Opteron cores in a server for $, PCs in that they're x86 arch, but not PCs in the sense of, well, personal computers. Either way, smashes the 30 year old super compuetr, doesn't beat the 20 year old one. Top 1 is 33 PFlops (peak).
2023 Ryzen 9 7950, maybe a Threadripper if you're rich (talking about PCs, not super high end). Can hit into the teraflops on the CPU in LINPACK. It's a great CPU, an multiples faster than 2003, but like 2-3x on the single core speed and 4 the cores, so maybe only 12x faster. PC from 2013 is eminently usable (this machine is 13 years old!). But we can add a GPU now and get 83 TFlops beating the 20 year old super computer!!
But it's slowing down, but not quite by as much as I thought. I suspect that by 2033, the top CPU wont hit 35TFlops by itself, maybe 20TFlops , possibly less. On the GPU side, it's already in the bag. The next one wil be tricky though.
Re: (Score:2)
That FX-8350 was great, but if you were into gaming, it was only pretty good. That was good enough for me for the money, though. I'd still be using it if the motherboard hadn't died.
Re: (Score:2)
It always offered good bang for buck, and wasn't slow either. From what I recall certain tasks were pretty slow compared to intel at the time, but others it was trading blows an on parallel workloads it was very good. Certainly for Linux workloads they were good.
I think people love flipflopping between extreme opinions. AMD is now KILLING Intel. Intel RULE AMD ARE SHIT!!11one1!1. Those morans at AMD cant design a cpu for shit. GET A BRAIN! You know that sort of thing. Substituting angry, extreme opinions fo
Re: (Score:2)
I think most of those people just don't really have experience with both. Those of us who have sampled both as far back as the 386 era (386DX40? Yes, please, and thank you AMD) have seen them trading blows all along. Although a clear trend has emerged since Meltdown and SPECTRE appeared on the scene, with mitigation always far more costly on Intel since they were objectively bad about security (checking access controls AFTER access? good plan there guys.) I used to use Intel when I cared about reliability,
Re: (Score:2)
That FX-8350 was indeed an absolute price:performance monster when it came to multi-threaded/-process workloads. Nothing could touch it anywhere near that price point. I had the Phenom II X6 1045T before it, likewise.
Indeed it was! Hot, though. I had (well, ran they weren't personally mine) a cluster of Phenom II X6, though I don't recall the version. There's actually an ask slashdot about it from me before it was built. So, I did get to, not just imagine, but see a beowulf cluster of those!
That was one ins
Re: (Score:2)
I had a big cooler master dual fan quad heat pipe cooler for my FX. Now I have a 1600AF, and run the stock cooler. It's mostly quiet, and I definitely can't hear it over my 1070.
Re: (Score:2)
I've got a hotter one (39?? the 2 gen old 12 core one), didn't build it myself though. I opted for one of those integrated watercooled units. Block clamps to the CPU, then the radiator and fan goes on the case. Probably overkill! Nice and quiet though. The fan has nifty rubber mounts.
AMD is still ahead for this gen (Score:4, Interesting)
I recently did some comprehensive performance/price comparisons of VMs from 10 cloud providers [dev.to], and while public clouds don't employ the exotic interconnects etc that supercomputers use, the base is still AMD EPYC and Xeon Scalable in both cases, hence several of the conclusions still apply.
Specifically, it seems that Intel's latest, Sapphire Rapids (recently available from Google Cloud), just catches up (or at most very slightly passes) AMD's previous gen, the EPYC Milan in performance. From the EPYC Genoa benchmarks we know it's another step up from Milan (it will be available from Google Cloud later this year), so Intel will need at least one more gen to actually catch up. Intel is lower in performance/price over most providers too, so I suspect their power consumption is not better either. Oh, and while Sapphire Rapids seems to be very competitive with Milan in single-threaded performance, if you enable HyperThreading and run parallel threads, AMD gains an extra advantage (that's across all gens for both manufacturers).
Of course I was only benchmarking generic workloads that were based on what the company I work for is running (web services), so I could see very specialized tasks behaving differently.
The other interesting outcome is that for performance/price, ARM solutions seem to be on top, so you'd expect this to start translating into supercomputing designs. Perhaps it hasn't happened yet because these designs take years and also because ARM solutions until recently had a lower per-core performance, so having e.g. twice the cores adds complexity/cost/overhead even if they end up using less power.
However, Apple with their Apple Silicon have shown you can have cores as fast or faster than the best from AMD/Intel, so this should be changing. Yes, those are not server CPUs, but Amazon's Graviton3 is and in my benchmarks showed performance competitive to AMD/Intel.
Re: (Score:2)
From the EPYC Genoa benchmarks we know it's another step up from Milan (it will be available from Google Cloud later this year), so Intel will need at least one more gen to actually catch up. Intel is lower in performance/price over most providers too, so I suspect their power consumption is not better either.
This seems to mirror the situation in consumer/gaming computing [tomshardware.com]. While Intel can beat AMD offerings in some single core use cases, Intel is worse in multi-core and efficiency. For example Intel 13900K beats out the 7950X in some single core tests but loses in multi-core tests. But it requires at least 15% more power.
Re: (Score:2)
However, Apple with their Apple Silicon have shown you can have cores as fast or faster than the best from AMD/Intel
AAAaaaAaaaAaaAAaaAaaAAaaAaaAAA.
NO!
Fujitsu showed that with ARM cores this time round a full YEAR before Apple released the M1. ARM of course showed it first time round, but then fell behind.
Mine too (Score:2)
When I built a supercomputer roughly surprisingly in the top 50 or 100 (I forget), 20 years ago, it too was powered by AMD. Why? By far cheaper than any other solution.