EDA simulation on GPUs

[trinibwoy]

http://www.eetimes.com/news/semi/showArticle.jhtml?articleID=216500149

We've been working with Acceleware and Nvidia to port our EMPro and ADS Transient-Convolution Simulator products to GPUs.

Software developers have been helping us port the code, and chip designers have been our lead beta site. They've achieved a 14-fold improvement in simulation time.
Not quite 26 hours to 26 minutes, but we're getting there.

—Larry Lerner is R&D senior manager at Agilent Technologies, EEsof EDA division.Interesting. GPGPU is getting serious....

I have no clue how EDA simulation could be parallelized but wouldnt the lack of a cache diminish the GPUs memory bandwidth advantage somewhat?

[silent_guy]

I have no clue how EDA simulation could be parallelized but wouldnt the lack of a cache diminish the GPUs memory bandwidth advantage somewhat?
These are analog simulations. So above all, they require lots of flops and bandwidth to solve large (sparse) matrices. Much easier to parallelize than digital simulations which are much harder to coerce into SIMD friendly behavior.

[CarstenS]

Sounds like this would be the ideal field for a real-world show-off of AMDs RV7x0-GPUs, doesn't it? Makes me wonder why they edit: AMD! didn't take the chance - too few engineering ressources perhaps?

[Davros]

So what is an EDA simulation

[silent_guy]

Sounds like this would be the ideal field for a real-world show-off of AMDs RV7x0-GPUs, doesn't it? Makes me wonder why they didn't take the chance - too few engineering ressources perhaps?
How?

As long as AMD lacks an equivalently powerful and well supported tool chain and a unified marketing voice that shouts off the roof what they're offering, most people won't even know it exists.

VC's are heavy users of the Google test: if one of their portfolio startup companies (or their products) don't show up in the first 3 links when doing a search, they're in big trouble.

Give it a try: you want to learn about either player's GPGPU offering. How do you start?

AMD: I first tried 'brook+', which brings you to a couple of wikipedia entries and a 2 year old power point presentation on AMD's website. Type 'brook' and not a single AMD site shows up. 'firestream' lists a graphics card (3rd link) which gets me to where I want if I click on the generic 'software' quick link. I could have tried to google the equally bland and generic 'stream computing', but I honestly had forgotten that this is now the umbrella under which their GPGPU solution is operating.

Nvidia: it may be a silly name, but it's uniquely identifiable. You enter 'CUDA'. All relevant links and the most relevant one on top.


I suppose things will change once there's an OpenCL tool chain...

[silent_guy]

So what is an EDA simulation
Electronic Design Automation

These days, pretty much any electronic widget that's designed (chip, FPGA, PCB) goes through one or more simulations to see if it will work.

Digital simulations to check if the logic is correct.
Analog simulations for everything analog (on-chip and off-chip).
Power simulations
Noise simulations on chip and on PCB.
EM simulations to prevent interference with other devices
etc etc

[CarstenS]

How?

Sorry, I meant AMD. They should've gone to Agilent promoting the strengths of their hardware to be used as a showcase.

[Dave Baumann]

AMD: I first tried 'brook+', which brings you to a couple of wikipedia entries and a 2 year old power point presentation on AMD's website. Type 'brook' and not a single AMD site shows up. 'firestream' lists a graphics card (3rd link) which gets me to where I want if I click on the generic 'software' quick link. I could have tried to google the equally bland and generic 'stream computing', but I honestly had forgotten that this is now the umbrella under which their GPGPU solution is operating.

Its called 'ATI Stream'. Search on that.