Hacker News Reader

Comment by pierrebai | Hacker News Reader

AMD already has a fix, it requires a BIOS update for the motherboard. They have not given details, but some experts think the fix is merely an increase in the power delivered to that part of the chip. Their conclusion stems from teh fact that the FMA3 bug disappeared when the chip was over-clocked because the over-clocking required increasing the voltage to the chip. So the BIOS update probably adjust the voltages up a bit.

http://www.fudzilla.com/news/processors/43166-amd-confirms-r...

View on HN

Since they are pushing how efficiently the chip are regulating power to various parts of the chip to keep power draw to a minimum it sounds like a likely cause.

Assuming most of the testing budget is spent on the most common and likely instruction sequence patterns is probably not unrealistic. Still a bit odd that such a short sequence of repeats of the same instruction failed.

Which got me thinking.

At more than 2000 opcodes for x86-64 and legacy opcodes, testing all 3-sequences with one input is more than 8G sequences, to cover that and (guesstimated average) 64 bits of input per opcode is a staggering 10^29 number of instructions, or 2T CPU*year at 4Gops/s. Not easy to test all that, even if it would be off by a couple of magnitudes in the right direction!

Might be that valid 2-sequences is the most that can hope to be even close to exhaustively tested while at the same time covering some significant fraction of the operand space ?

The fact that modern CPU's work at all given their complexity in so many different areas looks pretty darn close to magic, even from a pretty close distance.

At least we try, with SMT solvers and such. Surprise surprise, math works!

A power issue would also plausibly explain why disabling SMT avoids the problem, since disabling SMT powers off a bunch of stuff (a whole thread context does need a bit of power), and, in a multi-threaded scenario, tends to reduce core usage thereby reducing average (! so likely not relevant) power draw as well.

And (probably more importantly) reduces overall IPC by around 30% generally. The whole point of SMT these days is to have continuous work for the ALUs despite recent cache misses.

Reply on news.ycombinator.com