How much should an Exaflop cost?

Disclaimer:

The point of this post is to shed some light of the many orders of magnitude difference between the cost of today’s supercomputers and the best case raw cost of silicon processing. Please take the article with a truckload of salt. Real supercomputers include a lot more than just refined silicon. The interesting question is how much wasteful overhead we could remove from a supercomputer without really affecting its usefulness?!

Facts:
• Adapteva’s Epiphany E16G301 processor runs at 1GH and delivers 32GFLOPS
• The silicon area of the E16G301 chip is 12mm^2
• Wafers are 300mm^2 in diameter

Assumptions:
• 65nm pricing of $2,500/wafer (hypothetical)
• 90% yield

Conclusion:
Given the above parameters, the raw cost of 1 Exaflop is $15M in 2012. This is 1/675th the cost of a $10B virtual exaflop machine consisting of 100 state of the art 10 Petaflop machines at an estimated $100M a pop and it could be built today.

Andreas Olofsson is the founder of Adapteva and the creator of the Epiphany architecture and Parallella Kickstarted open source computing project. Follow Andreas on Twitter.

 

Posted in Andreas' Blog.

4 Comments

  1. You didn’t take into account the rest of the hardware. Memory, interconnect, cooling, software that makes the 10 Petaflop machine. You’re also assuming your selling the chips at cost price.

    I am not saying Adapteva is not cheaper, but it’s false marketing to compare CPU wafer costs to a fully working supercomputer with software, installation costs, and initial support.

    On my desktops the CPU costs between 1/5 to 1/10 of my desktop cost.

    On the servers I used to purchase in a large financial firm (1,000+ per quarter), the CPU cost was less than 1/150 of the cost of a rack of computers with cheap (relatively) 10Gbit Ethernet connectivity and top of rack switches.

    On larger deployment clusters of distribution switches and storage CPU made up approximately 1/1,000 of the entire solution (excluding end user software costs).

    So assuming you don’t solve the remaining problems around supercomputers, to create a 10 Petaflop machine using your chips would be 150M compared to the competitions 100M

  2. Andreas, I very much like your approach, and also your thinking with respect to many of the issues you’ve addressed in your architecture. However, given the very large number of cores that would be involved, even in a sub-exaflop configuration, I am concerned about single-event upsets (http://en.wikipedia.org/wiki/SEU). How should one avoid, handle, or otherwise mitigate SEUs when their probability of occurrence is very possiblly 1 (100%), over a rather short period which is likely to be smaller than the time required to solve a computation problem?

Leave a Reply Using Facebook or Twitter Account