Slashdot Mirror
ARM In Supercomputers — 'Get Ready For the Change'
An anonymous reader writes "Commodity ARM CPUs are poised to to replace x86 CPUs in modern supercomputers just as commodity x86 CPUs replaced vector CPUs in early supercomputers. An analysis by the EU Mountblanc Project (PDF) (using Nvidia Tegra 2/3, Samsung Exynos 5 & Intel Core i7 CPUs) highlights the suitability and energy efficiency of ARM-based solutions. They finish off by saying, 'Current limitations [are] due to target market condition — not real technological challenges. ... A whole set of ARM server chips is coming — solving most of the limitations identified.'"
No doubt your CPU would win. But when looking at power/price as well, you'd have to pit your CPU against 50 or so ARM chips in parallel. For some solutions, it may be a far better choice. One size doesn't fit all.
As someone who does heavy duty scientific computing, I wouldn't say that "most" of the actual process power is in GPUs. They are certainly more powerful at certain tasks, but most applications run are legacy code, and most algorithms require substantial reworking to get them to run with reasonable performance on a GPU. Simply put, GPU for supercomputing is not quite a mature technology yet. I am personally not too interested in coding for GPUs simply because the code is not portable enough yet, and by the time the technology might be mature, there might be a new wave of technology (like ARM) that could be easier to work with.
You aren't operating in the supercomputing market. There, what matters is the how much processing you can get for how much money. You can always buy more chips, and power usage and cooling are both signficant factors. That's why x86 became dominant in that space. It was cheaper to buy a bunch of x86 chips than to buy fewer POWER chips. In terms of computing power, a POWER7 will eat your i7 for breakfast, but they are ungodly expensive.
This is my signature. There are many like it, but this one is mine.
architecture is complicated. but in terms of ops per mm^2, or ops per watt, ops per $,
cycles per useful op, the x86 architecture is a henious pox on the face of the
earth.
worse yet, your beloved x86 doesn't even have any source implications, its just
a useless thing.
In TFA's slides 10 and 11, Intel i7 chips are shown to be more efficient in terms of performance per watt than ARM chips. However, they're close to each other and Intel's prices are significantly higher.
"I zero-index my hamsters" - Willtor (147206)
Of the last published top500 list, 7 out of the top 10 had no GPUs. This is a clear indication that while GPU is defintely there, claiming 'Most of the actual processing power' is overstating it a touch. It's particularly telling that there are so few as overwhelming the specific hpl benchmark is one of the key benefits of GPUs. Other benchmarks in more well rounded test suites don't treat GPUs so kindly.
XML is like violence. If it doesn't solve the problem, use more.
These ARM cores are halfway between the extremely limited GPU cores and the extremely flexible X86 cores. They may be the "happy medium".
Help stamp out iliturcy.
Also, a lot of algorithms, perhaps even most, rely on branching, which is something GPUs suck at. And only some can be reasonably rewritten in a branchless way.
The creatures outside looked from Alt-Right to Antifa; but already it was impossible to say which was which.
As I understand it, Intel still has the advantage in the performance per watt category for general processing and GPUs have better performance per watt IF you can optimize for that specific environment--both things which have been commented to death endlessly by people far more knowledgeable than I.
However, to me there are at least 3 questions unanswered:
1. ASICs (and possibly FPGAs): Bitcoin miners and DES breakers are the best known examples. Where is the dividing line between where your operations are specific enough to emply an ASIC vs not specific enough and needing a GPU (or even CPU)? Could further optimization move this line more toward the ASIC?
2. Huge dies: This has been talked about before, but it seems that, for applications that are embarrassingly parallel, this is clearly where the next revolution will be, with hundreds of cores (at least, and of whatever kind of "core" you want). So when will this stop being vaporware?
3. But what do we do about all the NON-parallel jobs? If you can't apply an ASIC and you can't break it down, you're still stuck at the basic wall we've been at for around a decade now: where's Moore's (performance) law here? It would seem the only hope is new algorithms: TRUE computer science!
Current ARM processors may indeed have a role to play in supercomputing, but the advantages this article implies don't exist.
Go look at performance figures for the Cortex-A15. It's *much* faster than the Cortex-A9. It also draws far more power. There's a reason why ARM's own product literature identifies the Cortex-A15 as a smartphone chip at the high end, but suggests strategies like big.LITTLE for lowering total power consumption. Next year, ARM's Cortex-A57 will start to appear. That'll be a 64-bit chip, it'll be faster than the Cortex-A15, it'll incorporate some further power efficiency improvements, and it'll use more power at peak load.
That doesn't mean ARM chips are bad -- it means that when it comes to semiconductors and the laws of physics, there are no magic bullets and no such thing as a free lunch.
http://www.extremetech.com/computing/155941-supercomputing-director-bets-2000-that-we-wont-have-exascale-computing-by-2020
I'm the author of that story, but I'm discussing a presentation given by one of the US's top supercomputing people. Pay particular attention to this graph:
http://www.extremetech.com/wp-content/uploads/2013/05/CostPerFlop.png
What it shows is the cost, in energy, of moving data. Keeping data local is essential to keeping power consumption down in a supercomputing environment. That means that smaller, less-efficient cores are a bad fit for environments in which data has to be synchronized across tens of thousands of cores and hundreds of nodes. Now, can you build ARM cores that have higher single-threaded efficiency? Absolutely, yes. But they use more power.
ARM is going to go into datacenters and supercomputers, but it has no magic powers that guarantee it better outcomes.
Also, a lot of algorithms, perhaps even most, rely on branching, which is something GPUs suck at. And only some can be reasonably rewritten in a branchless way.
nonsence, I play Farcry3 on my GPU, and it renders branches just fine thank you very much.