( DAC'18 Item 4a ) ------------------------------------------------ [01/23/19]
Subject: Avatar/AtopTech's big comeback in digital PnR is #4a "Best of 2018"
CHOICE IS GOOD: A whole bunch of digital PnR users were quite happy to see
that AtopTech ... err... correction ... make that "Avatar" was back in the
EDA game. There were two groups of users commenting on this:
1. the older earlier ATOP users who were generally happy that their
PnR tool of choice could now compete on a technology level instead
of on a "who has the sleaziest lawyers level".
2. and a number of potentional *new* Avatar users who love the idea
of having a viable 3rd choice in PnR over the CDNS vs. SNPS
duopoly -- especially now that MENT Olympus and Nitro-SoC are gone.
(01/30/2018 Edit: Mentor says Olympus and Nitro-SoC are NOT gone.)
ON THE TECHNICAL LEVEL the old guard users like Aprisa's performance, TAT,
and ease of use. One user even directly compared Avator favorably over
Synopsys ICC2. But the lowest node any existing users have succesfully
used Avatar at is 16/14nm.
What we need to see next is how Avatar plays in the emerging 7nm battle.
The first obvious hurdle is Avatar must get 7nm certified by TSMC in a world
where Cadence Innovus and Synopsys ICC2 already are.
Next, Avatar will have to explain that how as a point PnR tool player it can
work in a 7nm space where timing/extraction/noise/IR-drop integration is
becoming more and more important.
And it's one thing to get a bunch of tire-kickers sort of looking at your
"newish" PnR tool; it's another thing altogether getting those tire-kickers
to actually switch away from CDNS/SNPS and into to your Avatar tool.
But if anyone is going to do pull this off against the CNDS/SNPS duopoly,
it'll be Charlie Huang and Chi Ping Hsu, two old time EDA PnR veterans who
have a ton of experience fighting uphill EDA battles like this.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
QUESTION ASKED:
Q: "What were the 3 or 4 most INTERESTING specific EDA tools
you've seen this year? WHY did they interest you?"
---- ---- ---- ---- ---- ---- ----
COMMENTS FROM *PRIOR* AVATAR (ATOP) USERS
Avatar Aprisa.
We use two PnR tools at my company -- Avatar Aprisa and Synopsys ICC2.
We've done projects at 28 and 14 nm.
Our primarily focus is on runtime after post-route -- mostly ECO steps.
Aprisa's turnaround time is always faster than ICC2.
Aprisa has a built-in DRC engine that is very good, including at 14 nm.
I am highly confident in the DRC correlation between Aprisa and our
sign-off tool (Calibre). If we can clean our DRC in Aprisa, there
should be not many DRCs left for Calibre.
My feedback on other Aprisa features:
1. Easy to handle MCMM. Note that it does take more runtime
when you switch scenarios.
2. Handles slack-driven CTS well.
3. Support of AOCV/SOCV/POCV/LVF. (For now, we still have some
issues with POCV/LVF.)
4. Supports H-tree, Mesh, and Multi-point clock tree
methodology. (We currently still use own techniques here.)
5. We are happy with the built-in timing capabilities. The
correlation was good between Aprisa and PrimeTime sign-off
at 28 and 14 nm. Although at 7 nm, we still have issues
with getting Aprisa to correlate with PT-SI.
6. For area and routability, Aprisa and ICC2 are about the same.
Aprisa showed better results in some designs than ICC2, but
not on some others.
7. Aprisa has "double-patterning-technology" aware routing that
it handles well. (I also apply additional techniques to get
our final best results.)
Aprisa's ease of use is very good. All commands and its GUI are easy to
learn and use. The Avatar AEs are very nice and willing to support us.
Now that Avatar bought Atoptech, I believe they can compete with the big
EDA companies. Avatar will of course need to invest more in R&D to
improve their tools to do so.
---- ---- ---- ---- ---- ---- ----
We use Avatar/Atoptech Aprisa for place and route.
We've had multiple tapeouts with it. For high-speed and challenging
designs such as in networking, we depend on it to get us through.
I don't have the exact data, but my estimates are:
- Aprisa's runtime is about 20% better.
- Our turnaround time with Aprisa is about 50% better than the
competitive P&R tool. We can get the job done with Aprisa.
Feedback on some Aprisa features:
- We've used Aprisa's leakage-power based optimization and dynamic
power-based optimization and have found both to be useful.
- Aprisa has a built-in DRC engine, and the correlation has been
very good. There have been cases with new technologies where we
find differences, but Avatar's R&D team has been very responsive
in addressing and closing on the issues.
- Aprisa's global router does a good job in keeping track of this
and making sure post-route correlation is good. We have not had
any correlation issues with Aprisa between pre-route and
post-route.
This is even more important in FinFET nodes where metal thickness varies
a lot in a metal stack. It's key to set the right flags pre-route and
post-route to get the best correlation.
Avatar taking over Atoptech and bringing in new investment to Aprisa is
fantastic for us. The big industry players' AE support and R&D access
sucks for us. Avatar's AE support is outstanding, in-spite of our
designers spread out globally. And their R&D turnaround time to
address critical issues is also very good.
We need smaller EDA players to address the challenges of P&R especially
in the lower nodes, high speed designs and low power designs.
---- ---- ---- ---- ---- ---- ----
We use Aprisa (from Avatar/Atoptech). It can be difficult to compare
Aprisa's turnaround time and run time directly to other tools because of
multi-threading and its simpler licensing model. Aprisa will use all
available CPUs when it can, so for a single license, or small number of
licenses scenario, it's great.
TAT is also a function of how difficult the design is. If the design is
not too complex, Aprisa's run time/QoR is generally equal to the other
tools.
However, as the design gets more complex, the other tools sacrifice QoR
for runtime, while Aprisa still gives high QoR for a faster run time.
For our more complex blocks, we can usually get through Aprisa in half
the time of other tools.
General Capabilities
- Placement: Aprisa typically does a better job in placement
compared to other tools. The CTS is fast, and gives good
results (better than some, not as good as others), and the
post_cts opt and routing/routing opt are plusses as well.
We use MCMM, and EM fixing, and have had good results.
- Path-Based Analysis (PBA) timing: We're just starting to use
Aprisa's PBA timing, and we don't have much feedback on this
yet.
- Routability: We've had very good success with routability.
Overall, the designs are cleaner, with less shorts than other
solutions. We generally don't have any surprises once we get
through CTS.
- Ease of use: A big plus for Aprisa is that it's very easy and
straightforward to use. The only issues have been the commands
changing due to the legal issues. One thing we really like is
that they have an extremely useful clock tree browser. Unlike
other tools, it's fast, very intuitive, and gives very good
results. Customer support is good considering how small the
team is. They jump on problems and help find solutions
quickly.
Our designers have the choice of using multiple tools for PnR, and we
use Aprisa for the majority of our block level designs. We plan on
continuing to do so.
We're happy that the Avatar team will now continue to support and
develop Aprisa. Competition in this space is important, as is having
high quality, easy to use tools.
---- ---- ---- ---- ---- ---- ----
Avatar/ATopTech Aprisa.
APRISA'S ADVANTAGES:
1. Superior QoR compared with major P&R competitors
We have done some benchmarks with other major EDA vendors using our real
designs. We found:
- For the same timing spec, Aprisa achieved a smaller area
- DRC results from Aprisa correlated with Calibre much better
- Power is better in general, thanks to the smaller area and
less detoured routing, though not as significant as other
metrics.
2. Fast closure rather than fast turn-around time
I was amazed at how well Aprisa's placer and router worked together.
This has not been the case for other tools I've used. The benefit is
that Aprisa has very accurate congestion prediction during the
placement stage. That derives accurate timing prediction as well.
So, if the user sees that the feasibility of the prototype floorplan is
low, they can fix the problem early. In addition, Aprisa usually yields
a much lower DRC count at the end of routing compared with its
competitions.
After a reasonable floorplan is provided, Aprisa is not the fastest tool
to complete a P&R run among its competitors. However, it uses the least
amount of iterations by far to close a design. The difference is quite
significant and consistently applied to all our test cases.
3. Minimum porting effort for existing designs
Aprisa is highly compatibility with other vendors. e.g. It's command set
instructions that are compatible with Synopsys. We were able to port
our existing design flow to Aprisa with little effort.
4. Excellent support from very experienced AEs
In the EDA industry, applications engineers are one of the most
important assets, if not the most important one. A capable applications
engineer can figure out their customer's true requirement and drive
their R&D to provide solutions exactly for that.
I've seen some impotent applications engineers who don't know what their
customers are talking about. They ended up wasting lots of time and
effort of us all.
Not so for Avatar (Atoptech). All 3 AEs that I have worked with are
top-notch. They understood our problems even when there was not enough
info provided. They followed up with viable solutions in days, if they
could not provide it on the same day. I am truly impressed.
5. Prompt and no-nonsense solutions from R&D
In many cases, applications engineers cannot be blamed for sluggish
responses -- R&D should. But in my opinion, Avatar's R&D is second to
none in this industry.
They fix bugs and provide new features very fast. When the request is
impossible to achieve, they provide a detailed explanation.
I remember we had one feature request just like that. They have had
their R&D visit us on site. In the meeting, they explained that the
request had a conflict with another feature, so it was not possible to
do. They then provided alternatives to accomplish the same effect.
This kind of service is quite rare in the EDA industry.
6. Dedicate binary builds tailored for our group
Working with big EDA vendors, lots of customers often suffer a serious
problem: the software patches usually come slowly, and when they do
come, the new binary usually fixes some problems but create a few
different ones.
We know they have lots of customers and they fix lots of bugs and add
lots of features in the same patch. Not all of those are targeting our
problems. That messes up things. Avatar (Atoptech) dedicated a branch
just for us. So, we never had to deal with this problem.
APRISA'S DISADVANTAGES
1. Relatively small customer base
Compared with large EDA vendors like Cadence and Synopsys, Avatar
(Atoptech) is very small. They don't have many customers. They also
don't have enough resources to deal with all different kinds of problems
that come with the different natures of the designs. For example, they
might not have experience to handle mobile chips that require ultra-low
power. So, if you are new to them, it may take some time for them to
ramp up for your particular design.
2. Potential legal battles
In our EDA industry, lawsuits are a survival tactical tool. Synopsys
used it on Atoptech to take them down before. They could do it again
with Avatar. This is a general problem that bigger players suppress
emerging players in ugly ways. As a customer of EDA, we just have to
always be prepared for the worst.
3. Politically, the new company, Avatar may be closely watched due
to the connection with China.
Due to the US-China trade war, this is bad timing for a high-tech
company like Avatar (Atoptech) to rise. Let's see how it develops.
CONCLUSION
From an engineer's point of view, seeing better technology and fair
competition in the EDA means better lives for us. Because in our
demanding semiconductor industry, we have been constantly pushed by
seemingly impossible feature set expansions and crazy tapeout schedules
internally. Externally, foundries and IP providers always challenge us
with incredibly complicated rule decks and product issues.
EDA does not have to be another source of headaches if we select and
work with the vendors wisely, and if there is healthy competition. I
hope to see new company like Avatar (Atoptech) bring positives into our
entire semiconductor ecosystem.
---- ---- ---- ---- ---- ---- ----
Below is a high-level view of how our engineers feel about using
Avatar's Aprisa P&R tool, and how it compares with other industry
standard tools. We didn't run many benchmarks on Aprisa, so we don't
have specific details to share with regards to numbers and value
comparison.
UI/UX:
- Appears familiar owing to similarity with ICC2 & Innovus. Easy
to navigate and find things around.
Ease of Use:
- Tool is easy-to-use and command structures appear similar to
that of ICC2
- Some shortcut keys have complex combination, i.e. 3 keys for
selecting and moving hard-macros in the floorplan
- Main pages could be further improved by adding examples
Ease of Migration from ICC2/Innovus:
- Better support for migrating from ICC2, e.g. RLC parasitic
information can be generated from ICT file but cannot be
generated from Cap-Table format.
The turn-around time from Avatar's customer support AE's was
exceptional. We didn't have to wait for too long for resolutions to
issues reported as compared to customer support from the other tool
vendors.
---- ---- ---- ---- ---- ---- ----
I am a long-time expert in Place and Route. Atoptech came along a few
years ago and grabbed some of the market from Cadence and Synopsys, by
being user friendly, easy to use, and very predictable, from both the
placement and routing perspective, as well as meeting timing and fixing
noise issues. They grabbed some low-end markets plus some of the
market from Synopsys.
However, even though TCL is an industry standard, some of the commands
they used were Synopsys proprietary, so they got engaged in a lawsuit
which then caused them to have to change the commands and somewhat
derailed them from staying current.
Atoptech was very effective for both 28nm down to 16nm. However, 7nm is
a lot more demanding and there are a lot more rules. They fell behind
in technology as they could not compete with multi-processors technology
by Cadence, so they started to lose market share, resulting in them
needing to sell the company off.
Looking forward with the new company Avatar: I'm sure Aprisa is fine
for 16nm and 28nm -- I used it myself on many chips. If they have good
programmers who understand the 7nm technology well and can run placement
and routing in parallel on many CPUs they have a chance. Though there
is another complication which I think will stop them from being
competitive again and that is physical synthesis, which requires a lot
of resources.
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
---- ---- ---- ---- ---- ---- ----
COMMENTS FROM *POTENTIAL* AVATAR (ATOP) USERS
It's good to see Atop back.
---- ---- ---- ---- ---- ---- ----
Avatar. Since Nitro-SoC is dead, we're shopping now.
---- ---- ---- ---- ---- ---- ----
Liked seeing Chip and Charlie revitalizing ATOP now.
---- ---- ---- ---- ---- ---- ----
Avatar.
---- ---- ---- ---- ---- ---- ----
We went to see how Avatar planned to compete at 7nm.
---- ---- ---- ---- ---- ---- ----
Avatar. Competition is good.
---- ---- ---- ---- ---- ---- ----
Happy to see Avatar at DAC. I'm sure Aart is furious.
---- ---- ---- ---- ---- ---- ----
Avatar. Having a third choice is good.
---- ---- ---- ---- ---- ---- ----
We're at 16nm. We're looking at Avatar because the Innovus
and ICC2 pricing is ridiculous.
---- ---- ---- ---- ---- ---- ----
After your Cheesy list, my first stop at DAC was the Avatar booth.
Thanks.
---- ---- ---- ---- ---- ---- ----
1. Avatar
2. Innovus
3. IC Compiler 2
---- ---- ---- ---- ---- ---- ----
Losing Olympus makes us consider Avatar.
---- ---- ---- ---- ---- ---- ----
Avatar if it's more cost effective.
---- ---- ---- ---- ---- ---- ----
Mgmt wants us to look at Avatar.
---- ---- ---- ---- ---- ---- ----
We're 16nm. Possible Avatar.
---- ---- ---- ---- ---- ---- ----
Avatar is in my top 3 to look at. We don't own it. Just curious
now that lawyer stuff is over.
---- ---- ---- ---- ---- ---- ----
I have a friend at Broadcom who tells me Avatar works nice at 16nm.
---- ---- ---- ---- ---- ---- ----
Looking at Avatar.
---- ---- ---- ---- ---- ---- ----
Avatar (Atop)
---- ---- ---- ---- ---- ---- ----
We do 28nm. Time to look at Avatar.
---- ---- ---- ---- ---- ---- ----
Thinking Avatar (maybe)?
---- ---- ---- ---- ---- ---- ----
Competition is good. Avatar.
---- ---- ---- ---- ---- ---- ----
Atop
---- ---- ---- ---- ---- ---- ----
I'd put Avatar in my list.
---- ---- ---- ---- ---- ---- ----
We're curious how Avatar plans to deal with integration issues.
I'm sure Aart won't let them work with Synopsys tools, nor
would Anirudh be that dumb either.
---- ---- ---- ---- ---- ---- ----
Do you have a list of known Avatar tape-outs, John? How many
made it to working silicon?
---- ---- ---- ---- ---- ---- ----
Atop. Oh. Make that Avatar.
---- ---- ---- ---- ---- ---- ----
We plan on giving Avatar a closer look later this year.
---- ---- ---- ---- ---- ---- ----
For cost custing we're considering Avatar.
---- ---- ---- ---- ---- ---- ----
Yes! Avatar! We all win with competition.
---- ---- ---- ---- ---- ---- ----
possibly Avatar
---- ---- ---- ---- ---- ---- ----
I guess Avatar should be in that list.
---- ---- ---- ---- ---- ---- ----
We like ATOP competing on a tech basis only now.
---- ---- ---- ---- ---- ---- ----
Related Articles
Real Intent smacks Synopsys CDC & RDC signoff as #3 "Best of 2018"
Avatar/AtopTech's big comeback in digital PnR is #4a "Best of 2018"
Cadence Innovus dominates Synopsys ICC/ICC2 is #4b "Best of 2018"
Join
Index
Next->Item
|
|
