( ESNUG 563 Item 3 ) -------------------------------------------- [11/01/16]
Subject: Dean on his new peer-to-peer chip design workspace accelerator
DAC'16 Troublemakers Panel in Austin, TX
Cooley: Dean.
Dean: Yes John.
Cooley: You started a new tool, and I put that in air quotes - "new"
tool - called PeerCache.
Dean: It's not really a tool, but yeah.
Cooley: It's a Napster type thing that you want use...
Dean: Yeah, it's for music sharing, you know for the engineers to
basically, we give them elevator music to improve productivity.
(laughter)
Sawicki: Righteous.
Cooley: Does Apple know you're doing that? They want to compete in that
space. I remember this tool a couple years ago except it was
called "work synch zero"?
Dean: ZeroTime Sync, yeah. (See DAC'12 #10)
Cooley: Are you just rehashing the same tool?
Dean: It's kind of like, kind of like - NO. PeerCache is completely
new technology. (See ESNUG 561 #2)
Cooley: How so?
Dean: My PeerCache uses P2P technology as its core communications
thing. Our older ZeroTime Sync was really great because it
allowed engineers to take the data that was in their design
management system, get into their work space in zero time,
and basically launch tools and get their EDA running faster.
What we discovered, though, is that there's way too much
"generated data" in the EDA tool workspaces for that to be
as effective as we hoped it would be.
It turns out that only about 20% of the data in the design
workspace (when launching the EDA tools) is typically inside
the design management system -- and that 80% of it is in the
form of "generated data".
Cooley: What sizes? What sizes, like 100 Megabytes?
Dean: We're talking 100s of Gigabytes, 100 gigabytes, 50 gigabytes,
250 gigabytes, terabyte...
And so what happened is we had solved the problem, but for
only 20% of the data that they really needed.
What PeerCache does, is it actually solves the problem for
the full 100%. It solves the problem for both the managed
data that's in the DM system, and the unmanaged -- or the
"generated data".
So if you're going to go and run Verilog or SPICE simulations
at 100 corners or more than 100 corners, and you've got to
copy the workspace, the engineers are wasting a lot of time
waiting for all this data to be moved around so they can
launch the runs. You waste a lot of time.
So what PeerCache does is we use a peer-to-peer system and
we can get that copy or replication of that data down to
zero time, like we did with ZeroTime Sync. The observation
I think that we all know is that we're all trying to design
the next generation of computers on the current generation
of computers.
Cooley: Right.
Dean: And we do a lot of great multi-core stuff -- like what Joe
was talking about -- multi-threaded stuff, to get the compute
up.
Then we've actually done farms (which has been renamed to
"scale-out" technology by those folks at Facebook and Google)
which we've been doing, long before they were doing it, to
get the compute even higher.
But we haven't done a similar kind of "scale-out" on the IO
side. So we take filers and rate systems and we gang a
bunch of discs together. But we've never really done an
architecture that "scaled-out" storage such that you could
get the benefits of the farm in the storage kind of environment.
What IC Manage PeerCache is, is basically your own personal
Akamai content delivery network that not only does reads, but
also does writes. And so the technology that we've developed
is peer-to-peer read-write technology, which is probably the
only implementation kind of level of technology in the world.
And it allows you to get "zero time sync" on not only your
managed data but also your generated data.
Cooley: Why wouldn't I just buy something from one of these other
enterprise software companies? They have something like
that, they must.
Dean: I'm sure IBM had, no enterprise software, like Oracle? You
want to get an Oracle database?
Cooley: No, no, you know what I mean, they've got to be populating
workspaces, too...
Dean: No, I don't know of anything, and I've been in that world
for quite a while. The concept is of doing a peer-to-peer
file delivery content system where you can get basically a
full 10 Gigabits of files shoved down that pipe when you
need it, so that your Verilog or SPICE simulation to launch
in 5 minutes rather than 5 hours. No one has that.
Cooley: So latency is the question.
Dean: Yeah.
Dean: The peer-to-peer solution works extremely well for chip design
and chip verification. And the benchmarks that we've done have
shown 3X to 10X kind of speed improvements. So the engineers
aren't waiting for these long copies before they can launch
their EDA jobs.
Cooley: Ok
---- ---- ---- ---- ---- ---- ----
Related Articles
Sawicki on how MENT Calibre DRC/LVS "breaks at almost every node"
Anirudh defends JasperGold's rep and the CDNS-RocketSim purchase
Dean on his new peer-to-peer chip design workspace accelerator
Amit on 55nm non-Gaussian variation starts, plus IoT low VDD issues
Raik on it's only OneSpin vs. Jasper in formal; not SNPS nor MENT
Join
Index
Next->Item
|
|
