( ESNUG 481 Item 6 ) -------------------------------------------- [05/08/09]
Subject: ( ESNUG 475 #7 ) Cadence Spectre vs Magma FineSim benchmark data
> FineSim SPICE is promising technology in terms of SPICE accuracy with a
> high simulation speed. It's good on large transistor level IP blocks and
> analog macrocells. It is probably the only one on the market fully
> supporting multi-thread and multi-process. We had very good benchmark
> results with respect to the pure SPICE engines and fast SPICE engines
> used at maximum accuracy.
>
> - Pierluigi Daglio of STmicroelectronics
From: [ Speedy Gonzales ]
Hi, John,
I must remain anon on this letter.
We recently evaluated the Magma FineSim SPICE simulator and have decided to
purchase it. FineSim can do multi-processor/-threading runs for accelerated
performance.
Below are some benchmark comparisons of Magma FineSim SPICE versus Cadence
Spectre. We ran a dual CPU machine with quad core Opterons running Redhat
Enterprise 4. Everything except the Verilog A sample was run with four
cores - we didn't explore runtimes on fewer cores.
The largest block we tried was the ADC with extracted parasitic with 90,000
transistors and 300 K capacitors.
SPEED
Cadence Magma
Test Spectre FineSim
Design Size Runtime Runtime Speed up
-------- ---------- -------- --------- --------
Phased Lock 400 trans. 12 hours 46 min 16x
Loop (4 cores)
Front end 2000 trans. 2.6 days 6 hrs 10x
Circuit (4 cores)
(10 uS)
ADC lpe 90,000 trans. 2.5 days 3.5 hours 17x
Cap. only 300K capac. (4 cores)
Small circuit Few minutes Few minutes
in Verilog A - * (1 core) (1 core)
* - Test was run to confirm FineSim supported Verilog A analog
modeling language. It worked.
Additionally, we have been using Magma's FineSim Pro Fast SPICE simulators
for about ~2 years now. (FineSim is Magma's accurate SPICE; FineSime Pro is
Magma's fast SPICE.) Here are Pro's number on our largest design:
Cadence Magma
Test Spectre FineSim PRO
Design Size Runtime Runtime Speed up
-------- ---------- -------- --------- --------
ADC lpe 90,000 trans. 2.5 days 2 hours 36x
Cap. only 300K capac. (1 core) (1 core)
We use Cadence Spectre as our correctness standard. The FineSim results
were essentially the same as the Cadence Spectre results.
It was trivial to rerun Spectre simulations in FineSim. No difficulties
with syntax or convergence. FineSim was trouble free. We didn't need any
support from Magma.
- [ Speedy Gonzales ]
---- ---- ---- ---- ---- ---- ----
From: John Cooley <jcooley=user domain=zeroskew hot mom>
To: [ Speedy Gonzales ]
Thanks, [ Speedy] for the data. Yes, you'll remain ANON. I was just
about to publish your letter but realized you gave no correlation numbers
between Spectre vs. FineSim and Spectre vs. FineSim Pro. Could you give
me those numbers?
- John Cooley
DeepChip.com Holliston, MA
---- ---- ---- ---- ---- ---- ----
From: [ Speedy Gonzales ]
To: John Cooley <jcooley=user domain=zeroskew hot mom>
Hi, John,
You're welcome to the data. I think parallel SPICE is mature enough that
designers should consider it.
I don't have any numbers for you. We felt the Spectre and FineSim SPICE
results were equivalent. By that I mean we had no reason to think one was
more correct than the other.
The fast SPICE situation is more complicated as the accuracy depends on
the settings used, the design simulated, and exactly what metric is being
compared. I don't have a useful number there either.
- [ Speedy Gonzales ]
Join
Index
Next->Item
|
|
