$ \sum _{1}^{\infty }\tau$

The unsustainable aspect I try to stress (and apparently didn’t do a very good job with) refers to the energy consumption. Yes, the improvements are impressive as ever when it comes to things governed by Moor’s law, but leak currents are not going anywhere, and in these days of big data, the hunger for processing power easily outpaces what conventional super-computing can deliver with a reasonable CO2 footprint.

That the current trajectory of our supercomputer power consumption is on an unsustainable path should be obvious by simply glancing at this chart.

I disagree. Your data shows that power of an arbitrarily large machine roughly doubled in one decade. Budgets have presumably at least doubled for supercomputing resources in the same period. Performance also increased by ~1,000x! Only when you force the data to fit a quadratic curve does it become “unsustainable”. Also, I don’t believe the #1 supercomputer ever used 0 kW.

I’m not saying d-wave isn’t the next leap forward, but this particularly publication does not prove (or disprove) anything, except that chip A (which is designed to solve this problem specifically) is faster than chip B (which is designed to solve any type of problem).

What would be interesting would be for someone to design ASIC chips (not just GPU) which are also utilized alongside carefully optimized software specifically to tackle this subset of problems. Only once d-wave can prove that they can ‘out-scale’ that system (both in terms of silicon and power) can we assume they’ve done something concretely relevant.

Apparently Matthias Troyer has been working in this space and has been giving talks. Scott will blog about it soon. I’m definitely waiting on his results.

Let me re-iterate though, I’m very hopeful that Geordie and Friends have done it. But we all need to tamper our optimism a little bit. Getting ahead of ourselves just leads to providing fuel to those who want to bring the team down.

Also, I wanted to bring up a small point. When looking at the arvix paper studying the 128 qubit machine, it concluded that Dwave hasn’t built anything that fundamentally alters the algorithmic complexity, but rather just reduces the sizes of the constants.

http://arxiv.org/pdf/1304.4595v1.pdf

“While quantum mechanics is not expected to turn the exponential scaling into a polynomial one, the constants c and a can be smaller on quantum devices, potentially giving substantial speedup over classical algorithms.”

Has dwave claimed otherwise? If all they have done is reduced the constants, then I can see where a carefully optimized system might be able to challenge what they’re doing, even if doesn’t involve quantum behavior.