Hacker News Reader

Comment by remoquete | Hacker News Reader

In 1996 you needed 1,600 sq ft of supercomputing hardware to reach the same mark.

https://en.wikipedia.org/wiki/ASCI_Red

For a little less distant comparison, you could buy a Sun E10k in 1997. It was roughly two full racks in size with 64 total Sparc CPUS delivering ~50GFlops.

So 18 of those needed, 36 full racks of space, 1152 CPUs to get to 900GFlops. Each 10k is roughly 39"W x 50"D x 70"H, so about ~244 square feet of floor space for 18 of them, not including space needed around it. Including the space around it, it would be a typical 1BR apartment full of compute.

Also, I assume the M1 does significantly better than 900 GFlops if you don't run it through the browser.

Even more exciting would be a power use comparison, including for the air conditioning and other auxiliary support needs. Exciting times we live in

A single E10k fully populated needs 11,041 watts, not including the needed air conditioning. So 198kW for 18 of them. I don't know how to convert Btus/hr into power, but each E10k needs 37,165 Btus/hr of A/C.

The Macbook Air M1 would be something lower than 30W, since that's what the power supply is rated for.

https://www.wolframalpha.com/input/?i=37165+btu%2Fhour+in+wa...

“since that's what the power supply is rated for.“

Theoretically, that machine could use more power at full speed. That would mean you couldn’t charge it fast enough to keep it running at top speed forever, but that might not even be noticeable because you hit a heat limit earlier.

In practice, I guess the power usage is lower, as https://support.apple.com/kb/SP825?locale=en_US says it has a 49.9‑watt‑hour lithium‑polymer battery. At 30W, you would run out of that in less than 2 hours (but again, at full speed, you probably hit a heat limit earlier)

37165 BTU/h is roughly a quite convenient 10 BTU/s. 1 W is 1 J/s, and 1 BTU is about 1000 J. So the AC would need about 10 kW, or 180 kW for all of them.

All in all the computers in the example thus come out at north of 400 kW for total power use. Yikes!

That's a power reduction by a factor of more than 13,000 in less than two decades!

Put differently: If your electrical power comes from burning coal, the 1997 alternative to the computing power of a single M1 would emit about 8000 kg of CO₂ per day. The equivalent of driving 80,000 km with a very modern gasoline car. The mind boggles!

Are we comparing single precision vs. double precision?

I assume the M1 does significantly better than 900 GFlops if you don't run it through the browser.

What's being measured is loading something on to the GPU and running it there. It doesn't make much difference how it gets to the GPU.

If the current trend continues we will have a chip the size of a plank unit with a 2020 level computing capacity the size of the Jupiter.

Chat applications will be even slower then today

What will soon be the size of Jupiter is node_modules

Bad timing, the industry is about to move towards direct imports and much simpler dependency management thanks to the evolution of JS and likewise the tooling.

If anything node modules will slowly fade away in the coming years. And not just with that node rewrite that links directly.

Any more info?

Thank you for that info, I will consider it in the future. But before that happens and before these systems are widespread, it's going to be node_modules for another while.

And someone will still somehow have a 500MB folder with node_modules in 2037, I just know it.

If only I could give you gold

I opened Discord after 2yrs of not using it, remembering a simple Slack like chat that was slightly better.

I opened it again recently with all my old channels and it looks like they added a hundred new features I have no idea what they do or never would need.

Discord is still a good product but makes you miss the simplicity. Or a simple IRC interface (although I've seen some IRC servers with a million plugins).

This made me curious about GPU performance as well:

GeForce 210 (2009): 39.36 GFLOPS

GeForce GTX 980 Ti (2015): 6.060 TFLOPS

GeForce RTX 3090 (2020): 35.58 TFLOPS

The GT210 is just about the furthest thing from high end. The fastest single GPU model from that generation (GTX 285) is capable of ~700 tflops.

700 GFLOPS, you mean?

Yup, thanks.

Right, the low end models are excessively underpowered and only exists for niche cases or tricking people. -50 through -80 is a good baseline for normal, and even a GTS 250 is 387/470 gflops.

Wikipedia lists it at just over 1TFLOP:

https://en.wikipedia.org/wiki/GeForce_200_series

Maybe it's another case of Wikipedia not being accurate (at this point in time)?

That is actually funny because I sourced my number from Wikipedia too:

https://en.m.wikipedia.org/wiki/List_of_Nvidia_graphics_proc...

So there’s definitely something wrong there.

I like this metric: sq ft of supercomputer-year. It'd be fun to see comparisons of, say, phones, holding supercomputer-year constant.

Now, imagine going back in time to 1996 and trying to explain to people in the supercomputer lab that 25 years later there would be legions of people using portable supercomputers because you need this to build javascript applications for websites

Reply on news.ycombinator.com