ARM Not Just For Macs: Might Make Weather Forecasting Cheaper Too (nag.com) 41
An anonymous reader writes:
The fact that Apple is moving away from Intel to ARM has been making a lot of headlines recently — but that's not the only new place where ARM CPUs have been making a splash.
ARM has also been turning heads in High Performance Computing (HPC), and an ARM-based system is now the world's most powerful supercomputer (Fugaku). AWS recently made their 2nd generation ARM Graviton chips available which allows everyone to test HPC workloads on ARM silicon. A company called The Numerical Algorithms Group recently published a small benchmark study that compared weather simulations on Intel, AMD and ARM instances on AWS and reported that although the ARM silicon is slowest, it is also the cheapest for this benchmark.
The benchmark test concludes the ARM processor provides "a very cost-efficient solution...and performance is competitive to other, more traditional HPC processors."
ARM has also been turning heads in High Performance Computing (HPC), and an ARM-based system is now the world's most powerful supercomputer (Fugaku). AWS recently made their 2nd generation ARM Graviton chips available which allows everyone to test HPC workloads on ARM silicon. A company called The Numerical Algorithms Group recently published a small benchmark study that compared weather simulations on Intel, AMD and ARM instances on AWS and reported that although the ARM silicon is slowest, it is also the cheapest for this benchmark.
The benchmark test concludes the ARM processor provides "a very cost-efficient solution...and performance is competitive to other, more traditional HPC processors."
Cheap why? (Score:2)
Depends why it's cheaper and what htey mean by cheaper. If it uses less silicon die area, and maybe uses less electricity .. that's one thing. But if it's "cheaper" because it is priced lower that could turn out meaningless .. a function of supply/demand. If you buy the other one in bulk whoever's selling it might be able to offer you a better deal.
Re: Cheap why? (Score:2)
Re: Cheap why? (Score:5, Insightful)
Not every compute task needs a Xeon Gold or AMD EPYC with 64 cores. There are a ton of tasks where ARM is a lot better with power, such as DNS, NTP, light DB work, NAS I/O, Redmine, internal Web servers, etc.
My issue with ARM is the lack of a unified boot process. There is no such thing as UEFI or (blecch) BIOS booting, so loading an OS onto a Raspberry Pi is a completely different thing than loading an OS onto another SBC. Partially because most ARM chips have a lot on the die, and Linux kernel blobs for the drives can be dicey to fetch, and are almost never open sourced. It would be nice to have an SoC standard with open source kernel stuff, so Linux can boot directly without being virtualized (as with the Raspberry Pi and ThreadX), or loading vendor specific blobs, which likely will not be available for any recent kernel version.
ARM also needs some better hardware virtualization, because what the architecture excels at are lower performance , but yet vital tasks.
Re: (Score:1)
ARM has an entirely different market than Intel. You quite correctly identify Intel as the all-around performance winner. That's true, however if you care for only one thing, you could get away with designing your CPU for just that one thing, that's why AMD is popular right now, they are great with games, and that's about it, because they were designed to give the best gaming benchmarks. Similar reason why Apple is choosing their own ARM design, because they are very knowledgeable about their kernel and sof
Re: (Score:2)
ARM has an entirely different market than Intel. You quite correctly identify Intel as the all-around performance winner
AMD and Intel have been trading blows in that regard since the Am586, an internally RISCy x86 processor that was little-used but which proved that AMD could present a credible threat to Intel. While the K6 was not very good at being a Pentium, the K7 was a performance leader for some time. And AMD's current chips beat Intel's in several benchmarks.
Intel provided the best minimum and maximum gaming frame rates for years. AMD is now able to go toe to toe with them again.
As for ARM in HPC, if your problem is
Re: (Score:1)
Xeon's still beat AMD's server range on many scientific benchmarks. For single floating point, such as games, great, but not for serious applications. Also the triple memory controller per Xeon chip and multi-core concurrency issues of the AMD design give Intel the boost on many HPC and workstation configurations.
Re: (Score:2)
[insert ignorant babbling about instruction sets]. All nonsense
Why exactly is it "all nonsense"? You can screw up good utilization of those "masses of power hungry cache, pipelines, cores" with a shitty ISA, so that "small percentage of the chip for the instruction set" can have a significant negative impact.
Re: Cheap why? (Score:2)
SBBR specifies UEFI. Qualcomm uses UEFI to boot both Windows 10 and Android.
That's RPi's fault for using an 'embedded' processor.
Re: Cheap why? (Score:2)
Re: (Score:1)
Yes, in a time when mobile computing was still far, far in the distance. Now, the mobile world has shaped the processor world, Intel doesn't need to operate in that space because we still need high performance general purpose processors, but the world has adapted to be able to rent a fab for just 100k-1M units and then move on to another design.
In raw power, ARM cannot compete with Intel on price. In customization, it can.
Re: Cheap why? (Score:2)
Fujitsu A64FX chips say otherwise.
Re: (Score:1)
As I said, it depends on your measure of performance. It's great for a specific subset of calculations (floating point), but with only 16 PCIe lanes you'll be hard pressed to make a working workstation/HPC configuration out of it where you need a set of GPU and 100Gb network.
Re:Cheap why? (Score:4, Informative)
Depends why it's cheaper and what [they] mean by cheaper.
The article is comparing computational costs per hour for executing a Weather Research and Forecasting (WRF) Model on AWS of various architectures. The result they got was that ARM took longer but with the lower cost it still cost less.
The question now becomes, "how does AWS come up with their pricing" because that's what this claim is based on.
Re:Cheap why? (Score:4, Insightful)
The reason why it is cheaper is because Amazon controls the price of all three. And they are heavily discounting the price of Gravition instances right now because they want people to switch to Gravition. Don't be surprised if one day >50% of AWS workloads are running on Gravition the price will slowly creep up to almost the same amount as Xeon instances.
Gravition is nothing more than a trap designed to make it harder to run your workloads on non-AWS clouds or on-premises, just like most of the AWS APIs. To be fair, Azure is also moving in this direction.
How about making weather forcasting more accurate (Score:2)
How about making weather forcasting more accurate?
Its already free with ads
Re: (Score:3)
There have been incremental advances in weather forecasting for a LONG time. The joke back in the 1960s was that it wasn't accurate even 1-2 days ahead. Now people get upset that it's not predicting the weather correctly weeks out.
Re:How about making weather forcasting more accura (Score:5, Insightful)
People also suck at interpreting probabilities. If you say "10% chance of rain on X day", and then it rains, people say you're a liar and you got the prediction wrong. No, they predicted a 10% possibility that it would rain, that's entirely different to saying it definitely won't rain at all.
People are so bad at probabilities that you could probably tell people that there's a 45% chance of something happening, and a large percentage of people will only plan for the 55% outcome then get angry when the 45% one does.
Re: (Score:2)
To not mention that some newscasters etc might not do good readings as well.
Re: (Score:3)
Re: (Score:2)
People only remember the times it's wrong and they got caught without an umbrella.
Re:How about making weather forcasting more accura (Score:5, Informative)
Re: (Score:3)
Re: (Score:2)
I prefer increasing the weather forecast and dropping the cleavage.
Re: (Score:2)
Your ideas are intriguing to me and I wish to subscribe to your newsletter.
Re: (Score:2)
weather.gov —it's free without ads, too.
Re: (Score:3)
How about making weather [forecasting] more accurate?
To do this you need more and better satellites. That's actually something that's been recently improved, so we're doing that too.
Re: (Score:2)
This.
And I suspect that the quality of the forecasts depends on a lot more than the FLOPS per kWh that the hardware can supply. It's as much about the process and software [nytimes.com]. I find it hard to believe that we (the USA) can't afford as good or better machines than the EU has access to, politics willing.
Re: (Score:2)
I think the quality of the models matters quite a bit too. Research isn't just running some software.
Re: (Score:3)
Special property of Navier Stokes Equations. (Score:4, Interesting)
Conversely the state of points outside the zone of influence will not influence the central point. CFD, or computational fluid mechanics places a grid of points in space, and starting from some initial condition calculates the how this will evolve in future. So as long as the time step used in this calculation is made to be so small, each point in grid can influence its neighbor, but not neighbor's neighbor, the calculation involves very few points, and it can be done simultaneously for ALL the points in the grid. The time step restriction is called the CFL condition. And the ability to update/calculate the state of fluid in the entire grid simultaneously gives fantastic parallelization capability for the Navier-Stokes solvers. Update calculation is quite simple, just five equations (mass, momentum and energy conservation, momentum counts as three because its a vector), about 8 points, the code can fit into a tiny processor, the kind used to render pixels on the screen. Well organized data and the ability to parallelize.... thats why GPU based time domain solvers work very well for Weather and fluids.
This is true only for some, not all, simulations. Time domain electromagnetics and some accoustics problems fall into this category. Other physics based simulations are not parallelizable as easily.
So now ARM is just a Mac thing? Really? (Score:1, Insightful)
Apple users of today are all made dumber by these self flagellating mindless cult-tards that can only see anything Apple does as the only way.
I'm done with my rant, I'm going back to tinkering the tons of ARM devices I own, none of which are tainted by Apple's logo.
Re: (Score:3, Insightful)
ARM on Macintosh is a tech journalism thing. For a special type of tech journalist, no less: Mac enthusiast tech journalists.
They are a special sort.
Intel lost when the screwed up the Pentium IV (Score:1)
I am surprised it hasn't taken over the Datacenter (Score:2)
Lower power saves money in massive scales. Intel was leading the way 5 years ago before AMD and Microsoft licensing forced businesses to buy faster more wattage with lesser cores.
ARM now has Windows Server support and Linux had it for ages. Now if VMWare ESX can make an ARM port ...