6 ms·

#1 Supercomputer today is Summit - IBM Power System AC922, IBM POWER9 22C 3.07GHz, NVIDIA Volta GV100, Dual-rail Mellanox EDR Infiniband #1 Brain today (unknow

#1 Supercomputer today is Summit - IBM Power System AC922, IBM POWER9 22C 3.07GHz, NVIDIA Volta GV100, Dual-rail Mellanox EDR Infiniband #1 Brain today (unknown but someone like Terence Tao, Edward Witten, Magnus Carlsen, or Donald Trump) https://aiimpacts.org/brain-performance-in-flops/ https://aiimpacts.org/brain-performance-in-flops/ #1 Supercomputer performance: 148,600 TFlop/s power: 10,096 kW 1# Brain performance: 9,000 - 337,000 TFlop/s power: 20 W. Biology (=organic chemistry) has has huge computational power/watt advantage.

AllegedAlec 7y ago
> 1# Brain > performance: 9,000 - 337,000 TFlop/s > power: 20 W. Yeah, not really. These measurements mostly rely on the entire brain firing all at once, while a real brain does not do that.
- Razengan 7y ago
  GP did specify a generous range.
  - ben_w 7y ago
    Not as broad as the range given in the summary at the top of the link, which says 10^12 to 10^28 FLOPS. The lower end (5e10 operations per Joule) is worse than existing hardware: Google’s first generation TPUs were reported as 40 W for 92 TOPS (8-bit) = 2.3e12 o/J, and Apple’s “Neutral Engine” cores are similar (I can’t find specific values for power use, but that was 5 TOPS (8-bit) and therefore plausibly between 1 and 10 watts).
- TeMPOraL 7y ago
  I wouldn't even call them "measurements"; last time I checked, the brain didn't run on IEEE 754.
  - adrianN 7y ago
    I did my own measurements, I got about 0.01 floating point multiplications per second out of my brain.
    - thatcat 7y ago
      It's almost like your brain doesn't even have an FPU
      - 6gvONxR4sf7o 7y ago
        It does! But it's only good for counting up to 3 or 4.
buboard 7y ago
That just means we have to use brain-based computers
asynch8 7y ago
LMAO did you just reference Donald Trump as the #1 brain today? was this a joke?
- ben_w 7y ago
  Jokes about Trump being indistinguishable from a Markov chain trained on his tweets notwithstanding, the difference in natural language comprehension between the smartest and dumbest deciles of humans appears to be quite small compared to be small compared to the difference between AI and the average human. Or it was until recently, it’s hard to keep up.
- jacobush 7y ago
  Textbook neural network local optimum...

AllegedAlec 7y ago

> 1# Brain > performance: 9,000 - 337,000 TFlop/s > power: 20 W. Yeah, not really. These measurements mostly rely on the entire brain firing all at once, while a real brain does not do that.

Razengan 7y ago
GP did specify a generous range.
- ben_w 7y ago
  Not as broad as the range given in the summary at the top of the link, which says 10^12 to 10^28 FLOPS. The lower end (5e10 operations per Joule) is worse than existing hardware: Google’s first generation TPUs were reported as 40 W for 92 TOPS (8-bit) = 2.3e12 o/J, and Apple’s “Neutral Engine” cores are similar (I can’t find specific values for power use, but that was 5 TOPS (8-bit) and therefore plausibly between 1 and 10 watts).
TeMPOraL 7y ago
I wouldn't even call them "measurements"; last time I checked, the brain didn't run on IEEE 754.
- adrianN 7y ago
  I did my own measurements, I got about 0.01 floating point multiplications per second out of my brain.
  - thatcat 7y ago
    It's almost like your brain doesn't even have an FPU
    - 6gvONxR4sf7o 7y ago
      It does! But it's only good for counting up to 3 or 4.

buboard 7y ago

That just means we have to use brain-based computers

asynch8 7y ago

LMAO did you just reference Donald Trump as the #1 brain today? was this a joke?

ben_w 7y ago
Jokes about Trump being indistinguishable from a Markov chain trained on his tweets notwithstanding, the difference in natural language comprehension between the smartest and dumbest deciles of humans appears to be quite small compared to be small compared to the difference between AI and the average human. Or it was until recently, it’s hard to keep up.
jacobush 7y ago
Textbook neural network local optimum...