4 ms·

It should be well known by now that at the high end performance scales very poorly with power consumption (and also with price). (Same reason why you couldn't s

by em500 4y ago

It should be well known by now that at the high end performance scales very poorly with power consumption (and also with price). (Same reason why you couldn't straightforwardly extrapolate the great laptop performance of Apple's M1 and project them to clobber Intel/AMD on the desktop/server.) But the people who (think that they) really need those extra few % performance also tend to be the ones willing to spend. Personally I think this is fine: the big spenders who buy the top $$$ and TDP parts basically subsidize a lot of the costs for the more price conscious who benefit from all the R&D and still get great performance.

Dalewyn 4y ago
I agree. You buy high end parts either to enjoy it or because you truly need it. You don't buy high end parts to be fiscally efficient, efficiency has always been the realm of mid to low end parts. This applies to nearly everything.
- hayley-patton 4y ago
  It didn't seem that the mid-end was efficient when I last went processor shopping (last June). It appeared that of the Zen 3 processors (which were interesting as I already had an AM4 rest-of-the-computer) only the 5600 and 5900X had "good" prices to me, as the 6-core 5700 was only a bit faster than the 5600 for a fairly large jump in price, and the 8-core 5800 had only two more cores than the former processors, while not being _that_ much cheaper than the 12-core 5900X. Given that I hold onto parts for a while, and that moar cores with shared memory is what I was after (for fuzzing and model checking, mostly[1]) I took the plunge and got the 5900X. Another advantage there was that I would not have to build another computer to use two cheaper processors. But that only works as I don't think I am limited by any other factors like memory or disk bandwidth. I did however limit the processor to a 65W TDP to reduce heat output and power usage, which only slowed down my very parallel programs about 10%, and didn't affect less parallel programs at all. (The duplicated other components of two machines would also draw more power, though perhaps not that much in comparison to processors here.) [1] For a fuzzer you could easily give separate machines multiple random seeds, so really that is no issue. But for model checking/exhaustive search there needs to be a proper form of work distribution. I haven't a clue what the network bandwidth requirement of e.g. TLC, so that might also require faster NICs than one gets on motherboards to not stall processors on the network.

by em500 4y ago

Dalewyn 4y ago

I agree. You buy high end parts either to enjoy it or because you truly need it. You don't buy high end parts to be fiscally efficient, efficiency has always been the realm of mid to low end parts. This applies to nearly everything.

hayley-patton 4y ago
It didn't seem that the mid-end was efficient when I last went processor shopping (last June). It appeared that of the Zen 3 processors (which were interesting as I already had an AM4 rest-of-the-computer) only the 5600 and 5900X had "good" prices to me, as the 6-core 5700 was only a bit faster than the 5600 for a fairly large jump in price, and the 8-core 5800 had only two more cores than the former processors, while not being _that_ much cheaper than the 12-core 5900X. Given that I hold onto parts for a while, and that moar cores with shared memory is what I was after (for fuzzing and model checking, mostly[1]) I took the plunge and got the 5900X. Another advantage there was that I would not have to build another computer to use two cheaper processors. But that only works as I don't think I am limited by any other factors like memory or disk bandwidth. I did however limit the processor to a 65W TDP to reduce heat output and power usage, which only slowed down my very parallel programs about 10%, and didn't affect less parallel programs at all. (The duplicated other components of two machines would also draw more power, though perhaps not that much in comparison to processors here.) [1] For a fuzzer you could easily give separate machines multiple random seeds, so really that is no issue. But for model checking/exhaustive search there needs to be a proper form of work distribution. I haven't a clue what the network bandwidth requirement of e.g. TLC, so that might also require faster NICs than one gets on motherboards to not stall processors on the network.