Thursday, October 24, 2002


One of things that makes HPC (High Performance Computing, i.e. weather, aerodynamic, vehicle crash and nuclear bomb simulation) so expensive is the huge memory systems on these machines. There is an old saying in the field that latency is hard, bandwidth you just pay for. Anyway, the cost of a memory system isn't the memory itself -- Pricewatch puts a stick of 512MB of 200 MHz DDR memory at $137, or $274K for a terabyte. The cost has generally been in the plumbing necessary to connect that memory to the CPU(s)*. You need lots of very fast wires, which has historically required exotic packaging.

Things change. I just read that half of all cellphones are made on circuit boards with enough density to make an HPC architect drool. Chuck the unobtainium and use ALIVH -- this stuff is cheap! It makes me want to design a hobby HPC.

* Well, actually, the prices mostly come from market dynamics. You have consumers (big companies and government labs) with deep pockets and little ability or incentive to ensure they get real value for their money, and producers who try hard to lock their consumers into proprietary hardware and software to avoid real competition. The HPC money well is drying up as (a) defense budgets have been slashed and (b) many customers have realized that racks of commodity PCs will do the job just as well.

Tuesday, October 15, 2002

National Missile Defense

I watched a Frontline program on the National Missile Defense program recently, while waiting to feel tired enough to go to bed.

When Reagan started the Strategic Defense Initiative, the idea was to build a defense against nuclear missiles coming from the USSR. Frontline went over the history of SDI and NMD, but totally missed the fact that the insights and policy turnarounds during those two programs had all happened before. Summary: we've been trying to build ABMs ever since the 1950s. Every time we have a system ready to deploy, either the Congress or the President kills it, and then funds yet another system design. We actually had a working defense from 02-Oct-1975 to Feb-1976, about five months, before that last one, Safeguard, got killed.

Once again, after a few years everyone has realized that defending against incoming missiles is so hard that stopping thousands reliably, the first time, is impossible. Also, the USSR has evaporated. Russia still has the missiles, but we're buying their warheads from them slowly, and they seem less of a strategic threat somehow. So the folks in Washington have changed the emphasis, just as the Johnson administration did in the late 60s. SDI is now NMD, and the idea is to shoot down just a few incoming missiles from "rogue states". They're still trying to do it with "hit to kill" vehicles, basically, hit a warhead travelling 15,000 MPH with another missile travelling 15,000 MPH in the other direction.

Progress so far: since the mid-80s, they've spend $50 billion, they've decided they want to take out incoming warheads by smacking straight into them, but they can't quite hit the target, and they have no idea how to select the warhead from a cloud of decoys bursting from the same missile.

What bugs me most about all this waste is that now that the program goal has been changed, we could achieve that goal by redeploying the Safeguard system we had in 1975! Why screw around with entirely new research and development? I think we should do a cost reduction pass on the Sentinel system design, and then deploy the damn thing and be done with it. Yes, this means I think we should build more nuclear bombs. (The warheads on the Sprint and Nike Zeus missiles are thermonuclear, which gets rid of most of the accuracy and decoy selection problem.)

It would appear that policymakers in Washington have consistently concluded that it is important to do R&D on ABMs, but not important to actually have ABMs. Which makes you wonder if ABMs are supposed to shoot down warheads or keep a bunch of middle class voters in important districts employed.

If you want to find out how nuclear weapons actually work, check the Nuclear Weapons FAQ. It scared the piss out of me. I had previously assumed that nuclear weapons were a bit like a little piece of the Sun, but here. Nope. The Sun is a great big lazy ball of hydrogen with the odd thermal fusion reaction happening here or there. Nuclear weapons operate with an intensity that puts our Sun to shame. I find the notion of forcing reality so far away from it's normal course terrifying to read about.