( DAC'20 Item 1b ) ------------------------------------------------ [02/11/21]
Subject: Cadence Spectre-X users say 2x speed-up real is Best of 2020 #1b
2020 PROOF OF SPEED-UP: In my "Best of 2019" survey, I reported that two
sightings of early Spectre-X users. At that time, and probably because it
was a surprise "new matrix solver! tiny mem footprint! super fast!" attack
on Ravi's AFS SPICE two days before DAC 2019, most users were skeptical of
Anirudh's speed/accuracy claims. From back then:
"I don't believe the Spectre X claims. I've got to test it first."
"Right now we have more questions about Spectre-X than answers."
"skeptical on spectre x"
There were more than 50 "doubters" back then. But that all changed when my
2019 EDA user survey also found two early Spectre-X user benchmarks.
"The first user benchmark compared Spectre-X vs. Spectre-APS, where
Spectre-X saw a ~3.6x to 4.7x speed-up and used 1/3rd the memory
Spectre-APS uses.
The second user said he's been working on a Spectre-X vs. BDA AFS
40 million element testcase for months -- and right now they're
"both neck-and-neck" in terms of speed and accuracy."
- early Spectre-X skeptics & 2 benchmarks is Best of 2019 #2b
Those two Spectre-X quick & dirty benchmarks mildly placated some of early
doubters, but in truth all it did was whet the appetite for more benchmarks.
So I'm happy to report that this year in my "Best of 2020", I have 4 more
hands-on Spectre-X user benchmarks to add to the list. (In short, these
4 users report Spectre-X is 2x faster than Spectre-APS, and 7x faster than
the original no frills Spectre.)
MYSTERY SPICE BOOTED: on this first Spectre-X vs. Incumbant SPICE benchmark
even I don't know what vendor's SPICE got kicked out by Spectre-X. All I do
know is the booted Incumbent SPICE was NOT a Cadence SPICE.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen in 2020? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
Sorry, John,
My hands are tied about what I can talk about, but I want to say that
our use of Cadence Spectre-X is increasing as we qualified it to replace
our incumbent SPICE simulator -- which I can't name.
Our mgmt has decreed that our designers can never take a step backwards.
So, to win their acceptance over the incumbent, Spectre-X had to:
- maintain and/or improve accuracy
- provide clear performance benefits
Since I know you love hard data, John, the one number I can share from
our internal benchmarks is Spectre-X is 3x to 4x faster over what our
incumbent SPICE was.
And, yes, we've fully transitioned over to Spectre-X because of this.
We keep some of the old incumbent licenses around only for legacy work.
Our primary simulation domain is mixed-signal components like PLLs.
---- ---- ---- ---- ---- ---- ----
We ran Cadence Spectre-X on a full custom design with millions of
transistors.
We benchmarked it against Spectre APS for similar high accuracy levels.
- Performance: Spectre-X was ~2x faster vs. Spectre-APS
- Accuracy: Spectre-X met our requirements
Additionally, Spectre-X got a 5x-6x speed up on 8 cores verse 1 core.
---- ---- ---- ---- ---- ---- ----
Cadence Spectre-X SPICE Simulator
We compared Cadence Spectre-APS and Spectre-X for our post-layout ADCs,
PLLs, and RF ICs.
We ran Spectre-X with 16 cores, as we're already using Spectre-APS with
16 cores. Our designs have millions of elements.
Speedup
- In high accuracy mode, Spectre-X was 2x faster than Spectre-APS
- This saved us a lot of time, as our typical Spectre-APS runs
are two days.
Accuracy
- We use Spectre-APS for our golden simulator accuracy mode.
We've had many tape outs and silicon with it.
- Spectre-X's accuracy overall was good. Even so, it didn't have
sufficient accuracy for our post-layout high-speed timing needs.
Other Factors
- Spectre-X converges fine.
- Spectre-X is very similar to Spectre-APS in terms of how you
use it.
- Spectre-X has good integration with Cadence Virtuoso.
Spectre-X also supports mixed-signal simulation with Xcelium (via AMS
Designer). You can partition it for different accuracy settings --
e.g., for digital versus analog; this is useful for some circuits.
---- ---- ---- ---- ---- ---- ----
We use Cadence Spectre, Spectre-APS, and Spectre-X. We've had tapeouts
and silicon using the various Spectre's for nodes from 65nm to 7nm.
Our designs are very complex, with 100's of 1000's of transistors,
many frequency ranges, and sampling timesteps from 1 picosecond down to
10 femtoseconds (=10 quadrillionths of a second).
All this increases the amount of data that we must generate, as well as
our simulation time.
So Spectre-X's speed up has been good.
- At their highest accuracy settings, Spectre-X is about 35% faster
than Spectre-APS.
- This saves us time, as our SPICE simulations are multiday runs.
- We use the extra time to run more simulations, which then
increases our confidence in our verification.
We use Spectre-X with 8 cores. Spectre-X is better at parallelizing
larger designs. (While -APS sometimes errors when parallelizing.)
Using Spectre-X is just like using Spectre-APS -- only faster. Its
debugging is just as good as -APS and it's just as well integrated with
Cadence's Virtuoso environment. Which is good because everyone I know
who does analog or RF design uses Cadence's environment.
I'd absolutely recommend it. We've already had a tapeout with it at
12nm and 22nm.
---- ---- ---- ---- ---- ---- ----
We're a long time Cadence house. We're just happy to see that Anirudh
is funding R&D to make Spectre faster. The Cadence management prior to
Anirudh seemed not all that concerned because they had the Virtuoso ADE
monopoly driving Spectre sales.
---- ---- ---- ---- ---- ---- ----
We have a mix of -APS and -X licenses in house.
They do the job.
---- ---- ---- ---- ---- ---- ----
1. Spectre-X for its distributed processing.
2. Virtuoso for its old stuff. Stable is good.
3. Calibre to keep Cadence honest.
---- ---- ---- ---- ---- ---- ----
We've benchmarked Spectre-X to be around 7x faster than Spectre for
our designs. Our Cadence sales contacts say we should then pay 7x
more for it. We just laugh and then tell them to get serious.
---- ---- ---- ---- ---- ---- ----
Spectre-X for being 96% backwards compatible to Spectre/APS/XPS.
(96% isn't 100%, but it's damn good compared to many of the other
simulators backwards compatibilty claims.)
---- ---- ---- ---- ---- ---- ----
Biggest lie?
Anirudh's claim of "up to 128 cores" for Spectre-X. Maybe with
some contrived super simple circuit, yes. But anything production
level at 14nm or lower, Spectre-X craps out at 32 cores or so.
Hardly 128 cores.
---- ---- ---- ---- ---- ---- ----
Spectre-X
---- ---- ---- ---- ---- ---- ----
Our designs are at 90nm still, so original Spectre is more than
ample for us.
---- ---- ---- ---- ---- ---- ----
We've used Spectre, Spectre XPS, Spectre APS, Spectre AMS, etc.
As long as it plays nicely with Virtuoso ADE, we're happy.
---- ---- ---- ---- ---- ---- ----
Virtuoso, Spectre, Star-RC
---- ---- ---- ---- ---- ---- ----
Related Articles
Users choose Joe Sawicki's AFS-XT SPICE gambit as Best of 2020 #1a
Cadence Spectre-X users say 2x speed-up real is Best of 2020 #1b
Empyrean ALPS-GT crushes Spectre-X and AFS-XT is Best of 2020 #1c
EMX/Spectre-RF/Virtuoso-RF combo 75% easier use is Best of 2020 #1d
Siemens Symphony nips at Cadence AMS Designer is Best of 2020 #1e
Join
Index
Next->Item
|
|
