IIRC they claim 2.5-3x times more performance using a Tesla than using the CPUs in their workstation. Ignoring load time.

Their CPU numbers almost certainly take SIMD into account.

I'm doing cryptography research, and some of my colleagues have been considering building a similar "desktop supercomputer". The speedup there looks more reasonable: a single high-end GPU should be worth maybe 5-10 quad-core CPUs; it costs double and uses double the power, but it's easier to put a dozen of them in a single PC. The numbers aren't as good as for big matrices of floats, but that's because we're doing integer operations and GPUs aren't optimized for those. (But then again, crypto problems tend to set the standard for embarrassingly parallel problems.)

Anyway, the new "box-fulla-GPUs" supercomputers sure beat the heck out of the previous generation of cheap scientific compute cluster: a hundred PS3s running linux.

No, that's petrified grits you're smelling.

Long-term exposure to nicotine isn't good for you either, kind of like long-term exposure to caffeine. It at least causes heart and stomach problems. But it's much more addictive than caffeine.

It costs on the order of $10k to sequence a single genome. But you wouldn't do it for every cancer patient. Instead, you'd do it for a couple hundred cancer patients, and study the results. You'd hope to find a few dozen common mutations which indicate which treatment to use. Checking a cancer for a few dozen known marker genes is considerably easier than sequencing an entire genome.

Suppose that a faulty MegaCorp device burns Middle Class Joe's house to the ground. Joe tries to sue MegaCorp, but all he can afford is an average lawyer working on contingency. MegaCorp sends the dream team. If Joe wins, then MegaCorp has to pay him and his average lawyer. Sounds good: Joe gets more money. But if he loses, MegaCorp bankrupts him because even though his own lawyer is working on contingency, he has to pay MegaCorp's dream team. This provides a strong disincentive to bring suit.

It's even worse if the defendant automatically pays when losing. In this case, if the MAFIAA sues you, you defend yourself and you lose, you'll have to pay not only the ridiculous statutory damages, but also attorney's fees.

Of course, there are ways to fix this. The most obvious way is for the judge to have discretion on whether the loser pays, but caps on the fees might work almost as well.

The Indians already have a word for 10 million cores.

This isn't quite true for copyrights, at least in theory. EMI has a claim on basically any work which is derived from a significant portion of a Beatles song. They don't have a claim on work which is similar to a Beatles song, unless it is derived from that song.

Copyright law is about provenance, not content. If someone who has never heard a Beatles song (nor looked at the score of one, etc) composes and records an exact replica of "Yellow Submarine", EMI doesn't own it. But good luck convincing a judge or jury of this.

So if this guy is "psychoacoustically simulating" a Beatles song, he's illegally preparing a derivative work.

The upside in this is that the lights still work when the controller is down. They don't go flashing red, stay red, turn off or something worse.

Actually, it might be closer to the chance of 3 meteors colliding in midair in your living room.

Actually, this is a birthday attack. The point of a birthday attack is that with n samples, you have n(n-1)/2 possible collisions. Usually people call this n^2/2 or "about n^2".

When probabilities are so small, they more or less add. So the odds are about n^2 / (2*space) that you find a collision with n objects. So long as this number doesn't get close to 1, the approximation is accurate enough. (You could try to evaluate the formula you gave, or something very much like it, but the factorials are so large you'd have to approximate anyway.)

If you choose n close to 2^128, the probability becomes close to 1 and you have to choose a better approximation to find it. This gives you a way to find a collision with meaningful probability if you can hash about 2^128 random numbers and store the hashes. Obviously, this is not going to happen (yadda yadda boil the oceans yadda), but it means that it takes "about 2^128 effort" to break SHA256.

On the other hand, if you choose n = 8M, this gives about 8M^2 / 2^257 or 2.76e-64 probability of finding a collision. This is about the same probability as two meteors colliding in midair in your living room.

If SHA256 behaves randomly, the odds are considerably lower than that. The odds are on the order of n^2 / 2^256, where n is the number of blocks. Here n is about 8*10^6, so the odds are between 10^-64 and 10^-63.

If you only assume that SHA256 is collision-resistant, the odds might be more like the 10^-32 you suggested, but this seems unlikely. If such a problem were discovered, it would cause SHA256 to fall out of favor. People want their hash functions to behave randomly. (You could use a universal hash function and guarantee odds of 10^-64. But this would be a security disaster if an attacker somehow recovered the key.)

Still, many people will not accept heuristic or probabilistic solutions to deterministic problems, because they don't trust the heuristic and/or don't want to increase their chances of failure.

Just imagine a Beowulf cluster of those!

Actually, 2^64 bytes is exactly 16 exbibytes. But of course, that's the address space. You have to have some of it reserved for physical addresses, and some for the graphics card, some for memory-mapped I/O, and a whole bunch more for Stupid VM Tricks.

Still, not gonna happen anytime soon.

VIA Nano is 64-bit. Dunno how its price/performance/power compares to an undervolted desktop CPU or cheap laptop CPU. It's definitely faster, more power hungry and more expensive than an Atom. But like the Atom, it's also definitely available in Mini-ITX.

True, but they might sell it to another would-be customer for $500,000. Which you might not be happy with.