Fujitsu busts K super through 10 petaflops

I would hate to have to...

... pay energy bill of such a monster.

Reg units

I refuse to accept Petaflops - can we have it in terms of ZX81s instead?

@Steve 48

Excellent idea! Slow mode or Fast mode?

The ZX81 didn't do double precision, it used a 40 bit format (documented in the excellent manual).

I don't know the flops rating of a ZX81, and I should think that the difference between adds and multiplies (done in software) would be much larger than modern hardware, which may complicate comparisons.

For a rough idea, a mandelbrot renderer I wrote in BASIC on a CPC464 (same Z80 running at similar speed, also 40 bit FP) achieved about 166 iterations per second. That was 4 adds and 4 multiplies, giving a whopping 1333 Flops!

Re-writing it in PASCAL tripled the speed, at the expense of reduced precision of 32 bit.

That was still so dreadfully slow that I re-wrote it again in Z80 assembler, bumping the precision back to 40 bit with my own routines. That ran at double the speed again, 1000 iterations/sec. Here I was able to replace a multiply by 2 with a single INC instruction, so I'll only claim 7 ops per iteration for 7 KFlops. I was still running renders up to 2 days at 320x400 resolution.

I have no idea how fast double precision could be done on a Z80, which is what is needed for a true comparison, maybe someone has done it and knows?

K Supercomputer

It's not as catchy as Sparky.

Nice piece of kit, but what I want to know is...

... where does Japan get an extra 10 megawatts to run the thing, when major metropolitan centres like Tokyo need to disconnect customers on hot days, because of capacity constraints since the March 2011 earthquake knocked a bunch of Fukushima's reactors off-line?

(Although, I suppose a major research centre like RIKEN could have its own power supply, separate from the main grid. Does anyone here know if that's the case?)