( DAC'19 Item 2b ) ------------------------------------------------ [02/05/20]
Subject: Cadence Spectre-X skeptics & early sightings is Best of 2019 #2b
SEA OF SKEPTICS: Three days before DAC'19, I scooped that Cadence was going
to launch a new heavily revamped rewrite of Spectre called "Spectre-X".
Well, spies tell me that Anirudh's using this exact same solver technology
to give his elderly Spectre SPICE franchise a big dose of the BLUE PILL.
Yes, grandpappy Spectre is now a hot new young virile Spectre-X that's:
- massively parallel (runs up to 128 CPUs)
- new compact data model (5X smaller than present day Spectre)
- super fast (up to 10X faster than present day Spectre)
- 100% backwards compatible to present day Spectre/APS/RF/XPS/AMS
This new Cadence Spectre-X is Full SPICE accurate -- that is it's NOT
Fast SPICE, it's Full SPICE.
- from Anirudh is launching a new Spectre-X SPICE at this DAC'19
Again, keep in mind that Spectre-X was publicly launched 3 days before the
DAC'19 gathering in Las Vegas. And 2 of those days were a weekend! So what
happens when Anirudh drops a big EDA bombshell that's supposed to change the
SPICE world -- and crush both Mentor BDA AFS and Synopsys FineSim -- and you
then suddenly ask for engineers to react to the explosion but they don't
know the details???
We engineers are skeptical by nature. I'm NOT going to put MY name and MY
reputation on some EDA company's fantastic claims without doing a deep very
tech dive into those claims! And *this* is after SPICE users saw how years
earlier Spectre-APS & -XPS both with similar "10X faster" claims completely
*failed* to crush Mentor BDA AFS. (See DAC'17 #2, DAC'18 #6, DAC'19 #2a)
Some of the users' tamer reactions to the sudden Spectre-X news:
"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"
"We're skeptical about Spectre-X."
"after Spectre APS, we doubt Spectre X. we'll see."
There's more both here in this section and in the Mentor BDA AFS section of
this survey. They're a lot more "colorful". I stopped counting after 50,
but you get the point.
TWO EARLY SIGHTINGS: But beyond all this sound and fury signifying nothing,
as fate would have it, I lucked into 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.
I want to see more benchmarks -- especially direct user benchmarks of
Spectre-X against BDA AFS and/or FineSim -- but this initial user evidence
shows there is likely some "there" there.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen this year? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
Cadence Spectre-X
We tested Spectre-X on three different CMOS designs, running it on
8-core and 16-core CPUs using the MX mode for Spectre-X. The MX mode
is the default mode used in Spectre-APS. For us, the absolute runtimes
aren't inportant, but how -X benchmarks relative to -APS.
Spectre-X speedup
Design over Spectre APS
------ ----------------
SoC full chip (DC) power up 4.7X
48 KB Memory (extracted post-layout) 3.8X
CMOS dual PLL circuit 3.6X
We observed Spectre-X perform even better on a fully extracted memory
circuit due to its higher capacity for post-layout extracted circuits.
Spectre-X's performance scaled really well with the number of CPUs.
The three accuracy modes we used were the VX, MX and CX modes (out of
five possible modes.)
For all circuits in all modes, Spectre-X's performance improvement was
most noticeable (2x) using 16-core CPUs than it was using 8-core CPUs.
There was an additional performance improvement (1.2x) for the 32 core
count runs -- but it was moderate than the big 2x performance jump when
we went from 8 to 16 cores.
Also, Spectre-X runs even faster in the VX mode, but we haven't tested
the accuracy for our circuits in VX mode.
SPECTRE-X ACCURACY
Spectre-X's accuracy in its default accuracy mode (MX) was nearly
identical to Spectre-APS' highest accuracy mode -- but with much
greater (3.6x to 4.7x) speedup and with far less memory use.
SPECTRE-X MEMORY USE
For all three test cases mentioned above, we found that Spectre-X uses
1/3rd the memory that Spectre-APS uses.
This is a big deal for us, because 1/3rd less memory needed let's us
use our existing in-house machines rather than forcing us to buy new
machines. And it'll get even better when we eventually upgrade to
our 64+ core CPU platforms in the future.
Virtuoso Maestro Environment -
We primarily use Cadence Virtuoso ADE now, and have been moving to the
Cadence ADE Assembler/Assembler (Maestro) environment.
We've been moving to Maestro over ADE because
- Maestro enables a much better specification-driven design flow.
- Maestro also accelerates worst case simulations and regression
testing of analog circuits in its integrated cockpit.
We use Maestro with Cadence's AMS and Spectre tools today and we are
looking forward to further integration with SpectreX to accelerate our
specification-driven regressions.
AMS Designer Mixed-Signal Simulator -
We use AMS Designer in our mixed signal environment for all our full
chip analysis; currently with Spectre/Spectre APS and Cadence Xcelium.
We haven't tested Cadence AMS with Spectre-X yet but looking forward
to doing so. (FYI, Xcelium already runs parallel simulations with
multi-core CPUs quite well.)
Conclusion -
We've now completed enough testing to start migrating to Spectre-X,
due to its: 1) 3x to 5x higher speeds, 2) 8- & 16- CPU scalability,
3) 1/3rd memory requirement, and 4) maintaining high accuracy that
we need for our larger and larger circuits.
---- ---- ---- ---- ---- ---- ----
Cadence Spectre-X
We've been benchmarking Cadence SpectreX and Mentor (Siemens) BDA
AFS SPICE simulators. We've been working on the test case for months.
Our engineers are very deadline driven, so our schedule drives our
testbench accuracy/complexity. We want to be able to run the full
circuit rather than reduce our test benches for certain types of
circuits.
SpectreX vs. Mentor AFS benchmark
We ran a post-layout SPICE simulation of a ~40 M element design as
our test case. (We can't detail the design because it'll give us
away.)
We ran both SPICE simulators on 8 cores in the Cadence Virtuoso ADE
environment. They were compatible in terms of outputs and inputs.
1. The runtimes for the two SPICEs have been neck-and-neck.
But, there is more upside potential for Spectre-X while using
multi-threading. We tested both on 8 cores, and
- AFS is supposed to cap out at 16 cores of multi-threading.
- Cadence claims that Spectre-X can go to 128 cores.
2. Both SPICEs converged.
3. Both Spectre-X and BDA AFS had good accuracy. So far, the
results match what we expected for both SPICEs. (We do not
yet have silicon confirmation.)
I only have this one solid test case under my belt, but I was blown away
by the improvement that Spectre-X brings to Spectre.
SpectreX -- More automatic tuning, fewer configuration options needed
Cadence made a change that we like for Spectre-X. They reduced the
configuration options, with more now done automatically.
- This is good because having a complicated SPICE simulator
configuration can produce potential error points.
- Spectre-X tries to solve this by autotuning the simulator
internals with one global external setting: "+preset".
- This makes Spectre-X easier for us to use and get the best
speed at a given accuracy
- Another upside that simpler configuration brings is that if
the simulator has backend (under-the-hood) changes later,
it's not constrained to the legacy settings.
Accuracy mode correlation
You can correlate the five Mentor AFS "afsmode" accuracy modes to the
five Spectre-X "+preset" accuracy modes.
- Note that historically, there was a global option in both
simulators called "error preset", which had three settings:
liberal, moderate, and conservative.
- Spectre-X now eliminates "error preset". AFS does not. So
technically, AFS now has 15 different modes (5x3) to correlate
to (vs. Cadence's 5 modes) when benchmarking both tools.
Although having fewer modes with SpectreX (5 vs 15) may seem like less
control, Cadence claims the new under-the-hood algorithms will now
autocorrect for those issues. Although this feature would be very
desirable, we have not yet fully tested it.
Spectre-X definitely has strong potential. (I need to run additional
tests to be able to recommend it.)
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "[optional] What were the WORST tools or biggest EDA/IP
LIES you heard this year? Give details/explain why."
---- ---- ---- ---- ---- ---- ----
I'd call Spectre-X this biggest lie this year.
---- ---- ---- ---- ---- ---- ----
Want to do Spectre-X benchmarks before I nominate it as biggest lie.
---- ---- ---- ---- ---- ---- ----
We're skeptical about Spectre-X.
---- ---- ---- ---- ---- ---- ----
Spectre-X
---- ---- ---- ---- ---- ---- ----
after Spectre APS, we doubt Spectre X. we'll see.
---- ---- ---- ---- ---- ---- ----
doubt spectre x
---- ---- ---- ---- ---- ---- ----
I don't believe the Spectre X claims. I've got to test it first.
---- ---- ---- ---- ---- ---- ----
Spec X sounds too good to be true.
---- ---- ---- ---- ---- ---- ----
Spectre X sounds fishy
---- ---- ---- ---- ---- ---- ----
maybe Spectre-X as biggest lie?
---- ---- ---- ---- ---- ---- ----
Spectre-X got a lot of fanfare at DAC.
when we get home we'll see what it's real story is.
---- ---- ---- ---- ---- ---- ----
Anirudh's Spectre-X new matrix solver has us intrigued.
It might be real. It might be fluff.
We'll see.
---- ---- ---- ---- ---- ---- ----
Spectre-X
We'll check it out in 6 months after others have tested it.
---- ---- ---- ---- ---- ---- ----
Beckley would kill me if I didn't at least try Spectre-X.
---- ---- ---- ---- ---- ---- ----
wait-and-see on Spectre-X.
---- ---- ---- ---- ---- ---- ----
Right now we have more questions about Spectre-X than answers.
---- ---- ---- ---- ---- ---- ----
Our local Cadence sales guys are saying that Spectre-X is the
next big thing coming in SPICE. We'll see.
---- ---- ---- ---- ---- ---- ----
"Spectre-X -- the hottest new breakthrough since sliced bread!"
---- ---- ---- ---- ---- ---- ----
We asked our CDNS FAE to explain the new matrix solver in Spectre-X.
He couldn't.
---- ---- ---- ---- ---- ---- ----
Your big pre-DAC scoop (ESNUG 587 #3) got our experts to start
a first look at Spectre-X in Las Vegas.
Our jury is still out. I'll tell you how their eval works out
when I know.
---- ---- ---- ---- ---- ---- ----
Not hearing good things about Spectre-X.
---- ---- ---- ---- ---- ---- ----
Our question isn't how fast Spectre-X is, but can it converge
on our entire design?
---- ---- ---- ---- ---- ---- ----
Can't wait to benchmark Spectre-X
---- ---- ---- ---- ---- ---- ----
skeptical on spectre x
---- ---- ---- ---- ---- ---- ----
Spectre-X is 10x faster on exactly what types design?
---- ---- ---- ---- ---- ---- ----
biggest lie is spec-x until it's tested
---- ---- ---- ---- ---- ---- ----
We were happy with plain old Spectre licenses. Now they want us
to pay top dollar for fancyass Spectre-X licenses.
---- ---- ---- ---- ---- ---- ----
Related Articles
Mentor BDA AFS speed/accuracy/convergence wins Best of 2019 #2a
Cadence Spectre-X skeptics & early sightings is Best of 2019 #2b
Empyrean ALPS GPU crushing Cadence Spectre-APS is Best of 2019 #2c
Mentor Symphony takes on Cadence AMS Designer is Best of 2019 #2d
Join
Index
Next->Item
|
|
