( DAC'19 Item 2c ) ------------------------------------------------ [02/05/20]
Subject: Empyrean ALPS GPU crushing Cadence Spectre-APS is Best of 2019 #2c
ALPS-GPU SPEED BURST: We first heard from the Empyrean ALPS SPICE simulator
users in the DeepChip Best of 2018 #6, where they gave detailed benchmarks:
"We benchmarked ALPS vs. HSPICE,
- ALPS accuracy is the same as HSPICE
- ALPS is 3x to 6x faster than HSPICE
The largest design we've used it on had ~250,000 devices."
"We did 27 runs to cover all corners of our circuit. Our baseline was
Cadence APS+HPA. BDA AFS had an average frequency error of roughly
0.1% and was 2.04x faster. Empyrean ALPS had an average frequency
error of roughly 0.2% and was 5.26x faster."
With the original ALPS software beating HSPICE by ~4.5x, Spectre-APS by ~5x,
and BDA AFS by ~2x.
GOING ROCKETICK: At this DAC'19, the big news that Empyrean announced was
they made a SPICE that runs on GPU's.
When I first heard this I had wondered if Uri Tal, founder of Rocketick had
been secretly moonlighting at Empyrean. (When I asked Uri this, he looked
at me like I was drunk.)
Then when I asked Jason Xing, the U.S. CEO of Empyrean, *he* looked at me
like I was drunk, too! Anyway, long-story-short, ALPS-GPU is something that
Empyrean R&D developed on their own from the ground up -- and it's getting
speed-ups of 15x, 18x, 21x, 50x over Spectre-APS.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen this year? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
ALPS True SPICE Simulator
We use Empyrean ALPS to simulate our analog designs (ADC and PMU).
It is a reliable SPICE simulator.
ALPS can support a top-level scale circuit in post-simulation. The
largest design I simulated with it was over 50 million devices.
I work with ALPS in the same way I work with Cadence Spectre. I
use ALPS in Virtuoso or with a command-line interface.
ALPS, compared to Spectre, post-simulation:
- I've generally gotten a 3x to 5x acceleration with ALPS
when compared with Cadence Spectre.
- ALPS has a GPU mode (called ALPS-GT) that I often use.
- ALPS-GPU mode gives me 5x to 10x acceleration compared
with ALPS-GPU mode and 15x to 50x vs. Spectre.
- ALPS also has a good performance in convergence. We've had
cases that passed in ALPS that did not converge in Spectre.
As for accuracy, ALPS' simulation results are consistent with my
expectations and with our other SPICE simulator, Spectre.
I'd recommend ALPS based on its simulation acceleration and project
schedule reduction.
---- ---- ---- ---- ---- ---- ----
I use the Empyrean ALPS true SPICE simulator for our custom design
parts in our SoC -- including PLL, SerDes and IO designs.
ALPS is accurate and fast for pre-layout and post-layout simulation.
As for its capacity, I've used it to simulate a PLL during post-layout
simulation with millions of resistance and capacitance parasitics.
I ran a post layout simulation of a VCO:
- ALPS finished the simulation in 38 min with 0.5% frequency
accuracy.
- Cadence Spectre APS took more than 120 mins -- 3X longer
The ALPS "Smart Matrix Solver" sped up our post-layout simulation even
further. For the post-layout simulation of a multi-band VCO:
- ALPS, with its "Smart Matrix Solver", finished in 10 days
- Spectre APS would have taken 60+ days to finish -- 6X longer
(I estimated the time above based on Spectre APS finishing only 15%
after it ran for 10 days. I had to kill the Spectre APS sim.)
We run ALPS from on Linux terminal. We just use following command to
start the simulation:
linux> alps [-o outputfile] [-format format_type]
I recommend Empyrean ALPS for its fast speed and accuracy for simulation
of large netlists.
ALPS is especially useful for 16nm and 7nm processes. The SPICE model
parameters there are larger -- and layout parasitics are critical to the
circuit -- so accuracy and simulation time is very important.
---- ---- ---- ---- ---- ---- ----
I nominate Empyrean ALPS. I use it for simulation speed up -- as an
alternative to Cadence Spectre APS. I use both SPICE simulators with
Cadence's Virtuoso platform.
All my time consumption/performance comparisons below are based on the
same accuracy requirement.
I recently used ALPS on a design with about 4 million devices (w/o
parasitics), along with Cadence's QRC extraction tool to reduce the
netlist size increase due to the parasitics.
- Pre-layout simulation (16 nm & 7nm cases).
Empyrean ALPS' performance was similar to Spectre APS.
- Post-layout simulation. ALPS was much faster than Spectre APS
for post-layout simulation
Cadence Empyrean
Spectre APS ALPS
----------- -------
Post-layout simulation 1.0x 1.5-2.0x
GPU accelerator
Empyrean's ALPS GPU accelerator is kind of a revolutionary feature. It
shortened ALPS runtime by about *4x to 14x* for my two cases (both N7).
Seconds
---------
ALPS 517,615
ALPS + GPU Accelerator 58,690 <-- 9X speedup over ALPS alone
I'd recommend ALPS to other engineers. It's been very efficient for
our post-layout simulation, and the ALPS-GPU accelerator is the feature
I like the most because it gets 18x over Spectre-APS
---- ---- ---- ---- ---- ---- ----
Empyrean ALPS -- it's a high speed and high precision SPICE simulator.
I used ALPS to simulate high speed PLLs both at TSMC 7nm and 16nm.
Our design had approximately 5 million devices, with a working
frequency of 16 GHz.
Empyrean now also has a GPU accelerator to simulate large designs in
even less time. This is especially good for post layout simulation.
When I used ALPS-GPU accelerator:
- ALPS gave me a 3X speedup for post-layout simulation over
Cadence Spectre APS in TSMC 7nm. ALPS-GPU was 7x ALPS;
making 21x over Spectre-APS
- ALPS also had better precision than Spectre APS.
ALPS also has a built-in matrix solver which enables even faster
simulation.
We like ALPS precision and high speed, especially it's GPU accelerator
and RC reduction. I recommend it.
---- ---- ---- ---- ---- ---- ----
We use Empyrean ALPS true SPICE simulator with Cadence Virtuoso.
ALPS is fast and very easy to use, plus gives accurate results while
saving time for iterations.
- The largest design I've run ALPS on is an ADC.
- For post-ADC simulation, ALPS was ~4X to 6X faster than
Cadence Spectre APS.
- Empyrean now has a GPU accelerator that you can use with ALPS.
It is something we'd be interested in looking at.
ALPS' accuracy is very close to true SPICE -- typically within 3%.
---- ---- ---- ---- ---- ---- ----
We use Empyrean ALPS almost exclusively for our products, including one
design with more than 10 million devices.
ALPS' smart matrix solver speeds up our post-simulation even more.
This is especially important for our full R+C post-layout simulation,
as our netlist is so large. Without it, it would be too slow to run
a full R+C post simulation.
We run ALPS on our Red Hat Enterprise Linux servers (e.g. RHEL5 and
RHEL6). It's easy and works well. We just start Aether from terminal,
and then do everything including running simulation with ALPS.
I especially like the save/recovery, CCK check in ALPS, and the
calculator in their waveform viewer. They're really helpful.
In conclusion, ALPS works very well, with a high cost-performance ratio.
But the most important thing is Empyrean's strong support. They've
given us many new ALPS features and solutions based upon our requests.
---- ---- ---- ---- ---- ---- ----
I use Empyrean's ALPS SPICE simulator.
It's compatible with true SPICE and Spectre netlists. I can run it
using commands, or with Cadence ADE.
I simulated a PLL that had about 1 million devices with ALPS.
- ALPS ran 1.2X faster than Cadence Spectre APS.
- There was almost no difference in accuracy between ALPS
and SPICE.
Empyrean also has a GPU accelerator now, that will give even more
simulation speedup. We have yet to try it, but we're curious.
---- ---- ---- ---- ---- ---- ----
Like Empyrean ALPS for our SPICE. Killer capacity, fairly fast, and
much cheaper than AFS or Spectre-APS on a per license price.
---- ---- ---- ---- ---- ---- ----
ALPS-GPU (by Empire Design Automation I think.)
---- ---- ---- ---- ---- ---- ----
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
|
|
