What To Do With 1000 CPUs - The Answers

I recall taking a course called The Counselor Salesperson when I was an AE at Synopsys. The course was very popular across the industry and was the basis for the book Win-Win Selling. It advocated a consultative approach to sales, one in which the salesperson tries to understand the customer’s problem first and provide a solution that he needs second. Sounds obvious, but how often do you encounter a salesperson who knows he has what you need and then tries to convince you that you have a problem?

One of the techniques in the process is called the “Magic Wand” wherein the salesperson asks the customer “What would it be like if …”. This open-ended type of question is designed to free the customer’s mind to imagine solutions that he’d otherwise not consider due to real or imagined constraints. That’s the type of question I asked last week when I asked: What would you do with 1000 CPU’s? And boy did it free your minds!

Before I go into the responses, you may be wondering what was my point in asking the question in the first place.  Well, not so surprisingly, I’m looking to understand better the possible applications of cloud computing to EDA and ASIC design. If a designer, design team, or company can affordably access a large number of CPUs for a short period of time, as needed, what would that mean? What would they be able to do with this magic wand that they would not even have thought of otherwise?

I received 8 separate responses, some of them dripping with humor, sarcasm, and even disdain. Good stuff! I’ve looked them over and noticed that they seem to fall into 4 groups, each of which highlights a different aspect or issue of this question.

“Rent Them Out”

Gabe Moretti had the best response along these lines, “(I’d) heat my house and pool while selling time to shivering engineers”. Jeremy Ralph of PDTi put some dollar value on the proposition, calculating that he could make $8.25M per month sub-licensing the licenses and CPUs. While Guarav Jalan pointed out that I’d need to also provide bandwidth to support this “pay-as-you-use” batch farm.

The opportunity is to aggregate users together to share hardware and software resources. If I buy a large quantity of hardware and software on a long-term basis at discounted rates, then I can rent it out on a shorter-term basis at higher rates and make money. The EDA company wins because they get a big sale at a low cost-of-sales. The customers win because they get access to tools on a pay-as-you-go basis at lower cost without a long-term commitment. And I win because I get to pocket the difference for taking the risk.

“Philanthropy”

One of the reasons that Karen Bartleson and I get along so well is that we’ve both been around the EDA industry for some time (we’ll leave it at that). As a result, we not only feel connected to the industry, but also some sense of responsibility to give back. Karen would train university student’s on designing SOCs. I’d train displaced workers on tools that can help them find a new job.

Even though this is not really a business model, I think it is still something that the EDA vendors should consider. Mentor is already very active in promoting it’s Displaced Worker Program. Autodesk and SolidWorks are giving away free licenses to the unemployed. This type of program should be universal. Using cloud computing resources is an easy way to make it happen without investing in lots of hardware.

(On a side note: PLEASE, PLEASE encourage anyone you know at Synopsys and Cadence to follow Mentor’s lead. Synopsys did this in 2001 and Cadence once had a “Retool-To-Work” program that was similar. I truly believe that both companies have that same sense of corporate responsibility as Mentor has, but for some reason they have not felt the urgency of the current situation. I am personally going to issue a daily challenge on Twitter to Synopsys and Cadence to follow suit until it happens. Please Retweet.)

“Do Nothing”

John Eaton pointed out that it is very difficult to use any additional capability offered as “pumpkinware” if you know it will evaporate within a month. It would take that long to set up a way to use it. And John McGehee stated that his client already has all the “beer, wine, and sangria” they can drink (New Yorkers - do you remember Beefsteak Charlie’s?), so he’d pass. John: Can you hook me up with your client :-) ?

Seriously,  it certainly requires some planning to to take advantage of this type of horsepower. You don’t just fire off more simulations or synthesis runs or place and route jobs without a plan. For design teams that might have access to this type of capability, it’s important to figure out ahead of time how you will use it and for how long you will need it. If you will be running more sims, which sims will they be? How will you randomize them? How will you target them to the most risky parts of the design?

Run Lots of Experiments”

Which brings us to Jeremy Ralph’s 2nd response. This one wins the prize as best response because it was well thought out and also addressed the intention of the magic wand question: what problem could you solve that you otherwise could not have solved? Jeremy would use the resources to explore many different candidate architectures for his IP (aka chiplet) and select the best one.

One of the key benefits of the cloud is that anyone can have affordable access to 1000 CPUs if they want it. If that is the case, what sorts of new approaches could be implemented by the EDA tools in addressing design challenges? Could we implement place and route on 1000 CPUs and have it finish in an hour on a 100M gate design? Could we partition formal verification problems into smaller problems and solve what was formerly the unsolvable? Could we run lots more simulations to find the one key bug that will kill our chip? The cloud opens up a whole new set of possibilities.

__________

I’ve learned a lot from your responses. Some were expected and some were not. That’s what’s fun about doing this type of research … finding the unexpected. I’ll definitely give it some thought.

harry the ASIC guy

1 Star2 Stars3 Stars4 Stars5 Stars (No Ratings Yet)
Loading ... Loading ...

Tags: , , , , , ,

3 Responses to “What To Do With 1000 CPUs - The Answers”

  1. Jeremy Ralph Says:

    Thanks for the props Harry. For the “rent them out model”, one would have to figure out how to federate the users, data and runtime, to ensure privacy.

    Since not many EDA GUIs are AJAX, they can’t run quickly and cleanly in a web browser. That being said, EDA GUIs for synthesis tools are rarely used, and simulator GUIs are only really useful for debugging and non-regression.

    For non-GUI batch jobs, some AJAX interface would be nice to dispatch jobs and script flows.

    For client/server web security/scalability a load balancer with encryption acceleration would be nice. And to eliminate downtime there would need to be some redundancy at the web end.

    The majority of other nodes, after allocating some nodes for the above, could be worker compute nodes for batch type jobs that get dispatched to a job queue, via a slick AJAX web service interface that also uses email for sending messages. These worker nodes would run the tools and would resemble a central CAD compute grid.

    For all the above, the intent is to run applications over the web that were not intended to be run that way, and this makes for over complicated infrastructure and complexity. James from Xuropa with the online labs and the people at Cadence, with their hosted design solution have figured how to run traditional apps over the web. Our approach at PDTi with http://spectareg.com has been to build it as a web-application from the start so it’s a bit different.

  2. Bridging the Gulf: the Embarrassing Truth « Bugs Are Easy Says:

    […] this was a timely post. Just a few days after my last post, Harry  (theAsicGuy) Gries posted an entry speculating that having thousands of computers available would revolutionize EDA algorithms. I […]

  3. Jeremy Ralph Says:

    Commented on Bridging the Gulf’s blog (from link above)…

    I think it would be possible to speed up simulations using many processors and a high-speed interconnect and some gate arrays. Isn’t that what an emulator does? Maybe there is a market for a “verification cloud” with the resources and costs shared between many different companies.

Leave a Reply