This entry was posted
on Tuesday, April 7th, 2009 at 9:23 pm and is filed under EDA.
You can follow any responses to this entry through the RSS 2.0 feed.
You can leave a response, or trackback from your own site.
10 Responses to “What would you do with 1000 CPUs?”
I’d have a party, bigger than Denali’s, with drinks on me. That is assuming that I could sub-license them out to other companies. For argument’s sake let’s say I could get $10K/yearly subscriber and timeshare each license between 10 paying subscribers on average. That’s $100M/year by my math ($10K/subscriber*10 subscribers/license *1000 licenses).
My current client has FAMs with the major EDA vendors. Like everyone, we run our expensive EDA software on fast and plentiful servers.
Thank you for your generosity, but we have already paid for all the EDA software we can use next month. I’ll just turn off the 1000 CPUs and take the electricity. Your offer does include electricity, right?
Harry, would you be generous enough to add a “fat” internet connection to this bill for support the jobs coming in…
With that I’d host these CPUs & Licenses in a “pay as you use” mode.
Since the flow for any task (verif, back end …) is particular for any organization they should be able to push the data required to run a batch mode task (with their scripts controlling the flow) onto this farm and then pull back the results. They just pay for the time bound usage of these resources.
How about this. I’d require licenses for the following tools: simulator, synthesis, logical equivalence checker (LEC). With the simulator and synth tool I’d create a netlist for a performance critical IP but not really optimize the performance…. that’s what all the CPUs are for. The way you code the RTL has big impact on the performance…
… so I’d create some AI to use all the CPU resources, tweak the different aspects at the RTL level, and ensure that it passes LEC against the un-optimized netlist. Then I’d have a very high performance IP.
I used to know a very smart RTL developer who used this technique, but he did iterations of tweaking the RLT manually, synthesizing, LECing to create some very optimized RTL and corresponding netlist with the specific synthesis tool. If only the synthesis tools were smarter this wouldn’t be required