20090924-AoS-NMI-ESL

709 views

Published on

Presentation given to NMI ESL event 20090924 at Engineers House, Clifton, Bristol.

Published in: Technology, Design
0 Comments
0 Likes
Statistics
Notes
  • Be the first to comment

  • Be the first to like this

No Downloads
Views
Total views
709
On SlideShare
0
From Embeds
0
Number of Embeds
3
Actions
Shares
0
Downloads
6
Comments
0
Likes
0
Embeds 0
No embeds

No notes for slide

20090924-AoS-NMI-ESL

  1. 1. Designer Productivity – An Alternative Approach Rich Porter Art of Silicon 1
  2. 2. Art of Silicon ● Founded in 2005 ● Bristol based ● Multimedia centric Silicon IP ● Bespoke IP creation ● Consultancy 2
  3. 3. Background ● What do we mean by ESL? – running application code on large complex SoCs – uP + many peripherals ● Assembling at the top level – application validation – fitness for purpose – most important signoff test booting application code 3
  4. 4. Higher level models ● Some degree of architectural modeling always required ● Recent trend to TLM – but not cycle accurate – either ● not synthesisable ● or silicon inefficient ● Why not use RTL? 4
  5. 5. Silicon Efficiency ● If you print transistors use them! – (operations per cycle * MHz) / area ● Deep logic usually implies low MHz – or high power consumption ● High level models omit low level low detail at this cost ● RTL can give higher silicon efficiency – but at cost of simulation Hz – and development time 5
  6. 6. Does simulation Hz matter? ● YES! delta debug regress commit ● FE engineer will spend all of time here ● What happens in gaps? – ”Multitaskers bad at multitasking” – http://news.bbc.co.uk/1/hi/technology/8219212.st – email – tea, chat, bbc.co.uk, facebook?? 6
  7. 7. Does simulation Hz matter? II ● Regression time increases !!! sim time tapeout ● De-risking late show stopper fixes – When can I make my last RTL delta? ● Manageable signoff regression time 7
  8. 8. Where have all my cycles gone? ● Computers are fast and cheap ● HP DL165G6 Opteron 2.2GHz 12 cores – £1k 1u 12 cores, £40k 40u 480 cores ● MIPS cheap and available! – so how do I use them? ● 480 licenses a problem?! – even @ £1k/year == $$s – or another 12 racks? each year! 8
  9. 9. Verilator ● Open source verilog simulator – cycle based two state simulator ● Users report 2 to 3 magnitudes speedup – over event driven 4 state simulators ● Capacity for large modern SoCs – extremely low memory footprint ● Successfully used by many companies – including local companies 9
  10. 10. Verilator II ● 100KHz * 480 == 48MHz – not single threaded – but granular ● Easily shared via convential queuing system – silicon development – verification – applications ● Run out of capacity? 10 – buy more computers!
  11. 11. Verilator III ● Can't use e, System Verilog, behavioural verilog or VHDL – many testbenches use these – but these can be issues with other platforms too ● Can integrate into other simulator ● Can run standalone – waveforms – instrumentation 11
  12. 12. AoS Experience ● Verilator based flow since inception – 100KHz to 1MHz ● 50 slot compute farm – 2 tier model – > 10 cpus per engineer ● Regression management – efficient logging – web based UI – triage 12
  13. 13. Recommendations ● Get a compute farm ● Use a queue ● Replace CPUs regularly ● Prioritize high value jobs – $ licenses – interactive jobs ● Evaluate what verilator can do for you ● Profile, record and report 13
  14. 14. Questions ● Questions Rich Porter rich.porter@artofsilicon.com 14

×