Slashdot Mirror
White House Warns of Supercomputer Arms Race
dcblogs writes "The White House's science advisors, in a report last week, said a petaflop-by-petaflop race to achieve number one on the Top500 could prove costly and divert money from supercomputing research. 'While it would be imprudent to allow ourselves to fall significantly behind our peers with respect to scientific performance benchmarks that have demonstrable practical significance, a single-minded focus on maintaining clear superiority in terms of flops count is probably not in our national interest,' the report said (PDF). It is urging the supercomputing community to expand its benchmark measures beyond the Top500's Linpack. It says the Graph500, for data-intensive applications involving the rapid execution of graph operations, 'will be more relevant,' but also acknowledges that it will difficult to rely on any one measure."
Supercomputer Race. Unless supercomputers start blowing up or growing arms.
If another country starts to outshine you, try changing the rules.
America's strength used lie in an immense manufacturing culture, and that's given way to "intellectual property". Instead of dealing with tangibles, America is content to sit behind a desk and let the Chinese labour.
We cannot allow a Supercomputer gap!
If you really need to crunch a lot of numbers and are willing to spend a lot to do it, it often makes more sense to develop an ASIC or FPGA type solution. I know the EFF put together a key cracking system for $250,000 that would probably still blow modern supercomputers out of the water for that specific application.
I've had enough abrasive sigs. Kittens are cute and fuzzy.
"Mr. President, we must not allow a mineshaft gap!"
The "race" is not about the hardware. All modern supercomputers are massively clustered, using various shared memory architectures. The technology is commodity level, and even a small sum like $10 million can buy a SHITLOAD of hardware. The challenge, and the point of competition, is the creation of software technologies and algorithms to effectively make use of clustered hardware. It's a question of who has the best minds working on the software. The hardware is a given. People have constructed impressive massively parallel processors using game consoles, after all.
It's the programmers, not the supercomputer makers, who will make the difference in this "race."
1. Tianhe-1A (China)
2. Jaguar (USA, ORNL)
3. Nebulae (China)
4. TSUBAME (Japan)
5. Hopper (USA, LBNL)
6. Tera-100 (France)
7. Roadrunner (USA, LANL)
8. Kraken (USA, UT)
9. JUGENE (Germany)
10. Cielo (USA, LANL)
So, let's see -- half of the top ten are in the USA, two in China, two in Europe, and one in Japan. Granted, China is catching up (rapidly), but if you look beyond the #1 spot, the USA still pretty dominates the overall list. Expand this list out beyond the top ten, and SEVEN supercomputers from 11-20 are also in the USA (11-16 & 18), one is in Russia (17), and two in South Korea (19 & 20). So let's not all freak out here about China stealing the #1 and #3 slots on the list -- the USA still has quite a bit more computational resources than the Chinese,. . .
The advice to the president doesn't change rules for "fastest supercomputer". It tells the president not to be suckered into a supercomputer race measured on only the FLOPS, but rather on more useful performance measures. Because getting sidetracked into less useful metrics to see who's winning the race will waste US resources in winning the race, but not producing the most useful computer. And the US interest is in producing the most useful computer, not in nominally winning the race.
In fact, that report says "let China dominate the Top500, if the US still has the better computers". Which is exactly what I want the US doing, and what I prefer China to be doing rather than leaving the US behind in actual usefulness.
But if you want to get caught up in "the USA is dead" trip that leads into traps that actually would hurt the US if acted on, go ahead. You're not having any effect on the US supercomputer effort.
--
make install -not war
If you honestly think that the US can't cable together thousands of US GPU's in order to set yet another meaningless Linpack milestone, then you are not that bright.
US GPU's? Funny, I thought they where all manufactured in China. Oh let me check my NVIDIA GTX 570 box ... yeah that's right, they ARE made in China.
You only see that in Dan Brown novels because it's too dumb of an idea to be actually implemented. Short of a massive breakthrough in computer speeds that they've somehow managed to keep secret, even all the secret government supercomputers in the world would have a hard time breaking AES-128 or RSA-4096 in a reasonable amount of time.
If the government needs to break somebody's crypto, it's done through side-channel attacks. Anything else is a waste of effort.
Not a typewriter
Ok, not entirely worthless. Linear algebra is used in loads and loads HPC workloads, but Linpack as a benchmark is NOT prepresentative of a typical real-world HPC workload. It focuses on peak flops, leaving behind things like inter-node bandwidth and latency, which are crucial for many important, real scientific supercomputing tasks.
Our CSE department chair recently quotes an article he read. To paraphrase, we're heading to the point where computation is going to be basically free, and what costs all the energy will be moving the data around. This is true for several reasons. One is the recent trend towards near-threshold computing. Ultra-low voltage (i.e. 400mV, when 900mV is nominal Vdd) can save 100x on power. It costs us 10x on speed, but now we can pack in 100x as many nodes into the same power and cooling budget, allowing for a 10x increase in aggregate throughput. But this works best for highly parallel and communication-heavy workloads. Fortunately, for many important areas (bombs, climate simulations, astronomy, real-time raytracing), this is the case. And moreover, people are getting better at parallelizing work.
China does the manufacturing--but the chips are designed in the US by US companies. Those companies chose to locate their manufacturing in China because labor is so cheap there.
None of this addresses the main point, that Linpack isn't a particularly useful metric.