Today I attended a webinar put together by XtremeEDA and Cadence Design Systems titled “The Importance of a Complete Methodology for ESL”.  It is meant to be available later for on-demand viewing and the details will be put out by Cadence.

A rather disappointing 29 people were in attendance once the presenters were removed from the list, and some of those in attendance were Cadence people, so not sure what the real number was. Steve Brown from Cadence started the ball rolling with an introduction and the reason why they are working with XtremeEDA.  At the end he also pointed out that this was the first in a monthly series of webinars, and details of the next would be available soon.

My conclusion is that it was worth the hour to listen in for anyone who is just starting to look at the whole ESL area. For those with some knowledge already, there will be nothing new here as it was very high level. The question and answer session was perhaps the most valuable. Read on for more details...

David Black (XtremeEDA)  After he talked a little about who XtremeEDA is and the services they provide. Then he got onto the drivers for ESL. He highlighted the need for HW/SW codesign and agile development techniques. He talked about a 2 foot stack of paper for a specification and the fact they there must have been errors in there. He also highlighted problems with multicore because they create unpredictable systems.

So why can’t you use traditional RTL tools. Basically because they are available too late because it uses implementations and this is too late to make adjustments along any number of axes. Then he outlined the ESL promise starting with a virtual platform and the benefits they provide, then going through high-level synthesis, plus addressing the verification issue.

He then talked about the impact on schedule and the ability to parallelize parts of the flow, especially the ability to start SW validation earlier.

Went through 3 use cases - First case talked about the disconnect between architects and RTL implementation, second study related to hw/sw interaction via co-simulation, and the third was also about the creation of a virtual platform. In this case the processor choice had not been made so a general purpose MPU was necessary.

Then he discussed some of the methodology adoption issues, such as making changes in mid flight. Related this back to the value that XtremeEDA could provide to help with adoption. He made the statement that ESL is all about models – the right models, right abstractions, different needs and uses. He then discussed part of the verification methodology described in the Cadence TLM-Driven design and verification book, and the general methodology in terms of the modified V diagram that recommends a continuous and incremental verification strategy. He noted that this maps well into Agile design methods.

There were several questions raised:

What types of design best for HLS – Used to be DSP types, but this is changing. Now can also deal with control types of design. Synchronous designs only.

How does power get modeled? Static and dynamic modeled roughly and use data from previous designs based on activity. More accurate numbers can be obtained as the models get refined.

When looking to adopt ESL, where should you start? If you try too big a change you may set yourself up for failure. Identify your biggest problem and attack that. This is different for different companies. For some this will be synthesis, for other SW validation etc. Also depends on the time available.

How much time training etc to adopt ESL ? Do an assessment first and identify strategic targets. Some level of training for early team and bring consultants in to help them ramp up.

Issues between SystemC and SystemVerilog? Could be related to the choice of tools as not all tools originally dealt with anything but pin level connections. Getting better today.

Advantages of building cycle based models in SystemC. When do you need them? By the time you have put in the details you are close to RTL and thus come late in the process.  Also can get cycle accurate models from an RTL model (Carbon). However some dissatisfied with the performance increase they obtain.  So only build them as a last resort.

-------------------------------------------------
Brian Bailey – keeping you covered