( ESNUG 494 Item 9 ) -------------------------------------------- [10/25/11]
From: [ Speedy Gonzales ]
Subject: Berkeley AFS benchmarks 5x-10x faster than Cadence Spectre/APS
Hi John,
Berkeley DA recently held a 'Nanometer Verification Forum' at TechMart.
I know how you like real user data benchmarks, so here's my notes on a
speaker there who benchmarked BDA AFS vs. Cadence Spectre and APS. I am
keeping him 'anonymous' since he shared hard data that might anger Cadence.
Please keep me anonymous, too.
The speaker began by saying that sensors are everywhere, in smartphones,
game consoles, and automobiles. He then focused on the characterization
of a sigma delta modulator that is often used with such sensors.
He commented that sigma delta designers often use large oversampling
ratios (>100) to achieve high resolution for these modulators. Sim
times are very long as they are directly proportional to the oversampling
ratio. Designers have a tough choice, they can either:
1. Use a turbo/fast SPICE simulator, with not have enough accuracy,
or
2. Characterize blocks individually and use behavioral models for a
full SPICE simulation. This is useful, but does require accurate
characterizations for every block and is susceptible to missing
hidden interactions among these blocks.
He then compared the SPICE simulation times for Monte Carlo, Transient and
Transient Noise analysises.
MONTE CARLO:
His benchmark results below covered the block characterization approach he
discussed in his option 2 above.
Block type: Comparator Offset
Monte Carlo simulations: 2000
Std Dev (sigma)
Time (hrs:min) of offset (mV) Speed-up
Cadence Spectre 02:56 3:49 1.0
Cadence APS (n=4) 15:53 3:45 0.19x
Berkeley-DA AFS 00:56 3.42 3.15x
Berkeley-DA AFS MCP-* (n=4) 00:39 3:36 4.47x
* - MCP = multi-core parallel
He noted that Cadence APS didn't like the simulation set-up and actually
took 5x longer. (This was an unexpected result.) He also commented that
BDA was a lunchtime simulation, versus all afternoon for Cadence, and more
importantly, the accuracy did not change. He got the same offset
distribution for all these simulations.
TRANSIENT:
He then commented that rather than option 1 or 2 above, designers really
just want an accurate transient simulation, or transient noise simulation;
both of which were possible with BDA.
There were 840 total elements, 10 MHz clock frequency, oversampling ratio
of 500, .18 nm technology. All simulators/simulations were run using a
Quad core Xeon, with 12 GB of memory.
Time (hrs:min) SNDR (dB) Speed-up
Cadence Spectre 101:27 108.8 1.0
Cadence APS (n=4) 20:15 111.0 5.0x
BDA AFS 5:50 114.7 17.4x
BDA AFS MT (n=4) 3:50 114.8 26.4x
Cadence Spectre took over 100 hours, and Cadence APS cut down that sim time
by a factor of ~5X with 4 cores. BDA AFS cut it down by 17X, and with
4 cores. The AFS simulation time was under 4 hours for 26X. The speaker
indicated that these relative numbers were pretty standard for other designs
they had done also.
TRANSIENT NOISE:
The transient simulation results were good for comparisons, but what he
really wanted was transient noise.
Time (hrs:min) SNDR (dB) Speed-up
Cadence Spectre 261:00-* N/A-** 1.0
Cadence APS (n=4) 105:00 97.8 2.5x
BDA AFS 8:46 96.9 29.8x
BDA AFS MT (n=4) 7:58 97.8 32.8x
* - He couldn't wait this long for the simulation to finish, so he
extrapolated this from the time at 5% completion.
** - Since he couldn't complete the run, he didn't get an SNDR result.
The speaker said that Berkeley-DA AFS saved him 4 days for each transient
SPICE run, and 10.5 days for each transient noise SPICE run. He added
that since he ran these simulations multiple times, BDA actually saved him
many times that amount. He commented that he also did corner SPICE runs,
which he would not have been able to run without BDA.
- [ Speedy Gonzales ]
Join
Index
Next->Item
|
|
