One problem is that the core-years for RSA-250 were in 2017 Skylake Xeon single core terms.
If you can rent a dedicated 96-core Epyc for $1/hour (cheap dedicated host), the combination of IPC improvements (>2x) and core count make 7,010 "skylake core-years" cost only $175k, not $5M. On-demand cloud servers (which cost more than $1, maybe $5/hour) probably make the GPU cheaper, but it's closer than the author says.
I think this is the first time an "I made an agent swarm do something" blog post was actually written by a human, what a breath of fresh air.
3 hours ago [-]
big_toast 2 hours ago [-]
"I estimate that this factorization cost about 4,900 GPU-days, or 13.5 GPU-years, which is about $400k at current market prices"
"factorization ran at no marginal cost on spare or fragmented compute that couldn’t be used for other purposes"
Interesting use of stranded compute.
schoen 1 hours ago [-]
I'm confused by this because I was involved in various distributed computing projects from about 1997 to about 2001 (as a person running compute notes for them) and from about 2001 to 2019 (as a person helping to administer a distributed computing related prize), and in the early part of that era we routinely talked about idle computer power as "wasted" because of the idea that the computer might as well be used to compute something rather than sitting idle. This may have been very credible in 1990s devices that consumed a roughly comparable amount of power regardless of what specific computation they were performing, but all modern devices have extremely variable power consumption depending on the load. You can easily feel this as devices have fans turn on or get hot when the CPU is loaded, and in many cases you can easily query the CPU with software to find out how its power consumption or clock rate or other factors get adjusted based on computational load.
This means that the idea that idle compute would have gone to waste is just no longer true on modern devices.
Now there is certainly compute that couldn't be sold to a paying cloud customer because it's too fragmented in some sense, but it still has some amount of energy cost, and, in a data center, corresponding cooling cost attributable to the marginal heat production. How can one actually say that there is literally no marginal cost at all? I just can't imagine a device that literally has the same power draw regardless of load factor!
9 minutes ago [-]
mkmk 4 hours ago [-]
This is so well written as to be refreshing
bleepblap 4 hours ago [-]
Who is Devin and why do they listen to the author?
If you can rent a dedicated 96-core Epyc for $1/hour (cheap dedicated host), the combination of IPC improvements (>2x) and core count make 7,010 "skylake core-years" cost only $175k, not $5M. On-demand cloud servers (which cost more than $1, maybe $5/hour) probably make the GPU cheaper, but it's closer than the author says.
Normally we'd downweight a follow-up [1] but this is a good article and arguably adds SNI [2] in its own right.
[1] https://hn.algolia.com/?dateRange=all&page=0&prefix=true&que...
[2] https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...
"factorization ran at no marginal cost on spare or fragmented compute that couldn’t be used for other purposes"
Interesting use of stranded compute.
This means that the idea that idle compute would have gone to waste is just no longer true on modern devices.
Now there is certainly compute that couldn't be sold to a paying cloud customer because it's too fragmented in some sense, but it still has some amount of energy cost, and, in a data center, corresponding cooling cost attributable to the marginal heat production. How can one actually say that there is literally no marginal cost at all? I just can't imagine a device that literally has the same power draw regardless of load factor!