Hacker News Reader

Comment by throwaway_yy2Di | Hacker News Reader

"Today, all of the machines dedicated to mining Bitcoin have a computing power about 4,500 times the capacity of the United States government’s mightiest supercomputer, the IBM Sequoia, according to calculations done by Michael B. Taylor, a professor at the University of California, San Diego."

This is obviously wrong. Where did it come from?

View on HN

Back in the GPU mining days people would say that since their GPU could do either 300 MH/s or 2 TFLOPS that every MH/s is equivalent to x GFLOPS. Then you multiply that by the total hashrate and get a ridiculous number and compare it to the Top500. As others have pointed out, this thinking does not apply to ASICs at all but journalists cannot resist it.

Even more hilarious is the fact that SHA-256 doesn't involve floating point operations at all in the first place.

I would guess from Michael B. Taylor, a professor at the University of California, San Diego.

And UCSD is home of the San Diego Supercomputing Center (SDSC), a major site for National Science Foundation computing research. So rather than 'obviously wrong' I'd weigh this assessment as 'presumptively right' unless/until other better evidence or testimony is presented.

I don't know if it's wrong, but it is a meaningless comparison since bitcoin mining is done with special purpose hardware that could never, for example, run a linpack benchmark.

I imagine they are comparing raw operations per second (in some measure) but the comparison is totally meaningless.

I'm not sure of this specific calculation but in general they come from the network's total hashrate compared with the known floating point operations per second of the top 500 supercomputers. Not sure how flops has been equated to hashrates though.

Nowadays it makes less sense since bitcoin hardware is ASIC/custom silicon designed & optimized for hashing as opposed to floating point operations making the comparison biased.

The numbers are: 9 petahashes/second (entire Bitcoin network [FTA]), and 20 petaflops/second (IBM Sequoia [1]). If you equate a SHA-2 hash with a floating point op, the ratio is 0.45, not 4,500.

[1] http://www.top500.org/system/177556

This site (http://www.bitcoinwatch.com/) seems to do a 'flops' type conversion from the hashrate on the bitcoin network.

Currently at 108239.28 Petaflops which is 5400x the performance of the IBM Sequoia. This is approximately the quoted amount of 4500. While I don't agree with the idea that a hashrate can be converted into flops because its more bitwise opts others are drawing similar calculations to UCSD.

Also a bit more of a discussion on this: https://bitcointalk.org/index.php?topic=7675.20

Thanks for the numbers, but a SHA-256 hash takes a lot more than one "flop".

Do you think that is a sensible equivalence?

A Radeon 7990 gets around 1.2 gigahash/second and 8.2 teraflops.

The ASICS are of course skewed towards hashes from there, and I guess the supercomputer could be more skewed towards floating point than the GPU, but a large factor seems more likely than a fraction.

You are making a terrible apples vs. oranges comparison.

A bitcoin hash is about 1900 32-bit integer operations [1]. This is comparable to 1900 32-bit floating point operations, not one(!). Because most CPU/GPU, when they can execute N integer ops per cycle, can also execute N floating ops per cycle.

So the ratio clearly is in the ~1000 order of magnitude.

[1] https://en.bitcoin.it/wiki/Why_a_GPU_mines_faster_than_a_CPU...

"A bitcoin hash is about 1900 32-bit integer operations [1]. This is comparable to 1900 32-bit floating point operations, not one(!)."

That's the amount of work you do in a software implementation of SHA-2, running on a CPU or GPU. This abstracts the hardware complexity inside the chip, which is very relevant when we're going into ASIC's. SHA-2 is bit shifts, masks, and binary adders: it's much simpler than a floating point core.

A hardware comparison would be something like ops/(transistor * time). A recent Bitcoin press release [1] claims 500 billion SHA-2 hashes/second out of >1 billion transistors: ~500 hashes/(transistor-second). A 7990 has a peak throughput of 8.2 teraflops SP [2], out of 8.6 billion transistors: ~1,000 flops/(transistor-second). Both are on 28nm processes.

[1] http://finance.yahoo.com/news/uniquify-hashfast-announce-wor...

[2] https://en.wikipedia.org/wiki/Radeon_HD_7000_Series#Southern...

Reply on news.ycombinator.com