While the Wednesday Design Automation Conference keynote
speech by Bernie Meyerson, vice president of innovation at IBM, covered several
topics, what I found most interesting was his strong advocacy of cloud computing in EDA.
He also provided a fascinating inside look at how IBM is using the technology.
Meyerson's talk was a perfect setup for a Wednesday DAC
panel on "IC design in the clouds," which I'll blog about separately at a later
time. Suffice to say that cloud computing is a significant topic at this year's
DAC and one of growing interest within the EDA community. It's also an area of keen
interest for Cadence, given the company's Hosted Design Solutions
offering and its work with Xuropa on verification IP evaluation.
After a discussion about stream computing,
which is an interesting topic in its own right, Meyerson said "there is another
thing coming along that I believe will have a seminal impact on your future.
You will see more of this to the point where it will become dominant in 5 to 10
years. It's cloud computing."
What's driving the movement from in-house datacenters to the
cloud? Cost and power. Server costs are around $50 billion per year, but that's
not the real concern, according to Meyerson. He presented a chart that showed
exponential increases in server administration costs and power consumption.
Meyerson noted, in fact, that 1.2 percent of the world's power output is now
going into server farms and their cooling equipment. Spending on datacenter
power is growing 600 to 800 percent faster than spending on servers.
The bottom line, Meyerson said, is that "IT spending is
growing at unsustainable rates." This is making innovation very difficult. No
longer can two guys in a garage start a major company (a reference to
Hewlett-Packard), he said; today, startups must equip themselves with a lot of
expensive compute hardware. Unless, of course, they decide to use the cloud.
The IBM Experience
Meyerson said that IBM has been working on cloud computing
for 20 years, and is now at the point where all of its EDA software runs on a
cloud infrastructure. "After all, IBM is a global company," he said. "We have
3,000 designers from around the planet. If you start giving them a common
infrastructure, efficiency goes up dramatically."
The IBM cloud, he said, has 20,000-plus cores, controls 150
terabytes of memory, and runs 40,000 discrete jobs per day plus 50 million
simulation cycles. "It's an enormous undertaking for us and it's global. It's
the same thing whether you're in Bangalore, Austin, or Poughkeepsie."
The cloud, he noted, uses different hardware for different
tasks. Optical proximity correction (OPC) requires a supercomputer. Logic
synthesis runs on FPGA engines. But that's all hidden from IBM designers because
it's handled by the cloud.
IBM aims for 80-90 percent CPU utilization in their cloud.
How can that be obtained with a mix of batch and interactive jobs? "The way we
solved it," he said, "is that if you have an interactive job you get priority.
We will unload the batch jobs so the interactive jobs run with no delay. That
turned out to be the right balance."
Meyerson concluded that cloud computing "enables you to say
that I am not going to make this massive capital investment in building out my
IT infrastructure, I'm going to link into it. Then the whole business paradigm
starts to shift regarding how you run your business, where you get your IP,
where you keep your design. That design is precious - to fully design a major
microprocessor, there could be a half-billion dollars of value sitting out
there on somebody's cloud. You do want to ask about things like security, data
integrity, and backups."
"I'm not saying this is easy," Meyerson continued, "but this
is the direction it's going because it gives you an economy of scale you cannot
achieve on your own unless you intend to be an IT company."
There are many challenges with cloud computing, and these
were discussed in the panel I'll write about later. Some EDA applications will be more appropriate for tre cloud than others. But I found Meyerson's
arguments to be persuasive, especially since IBM has shown it can be done.
/* Style Definitions */
mso-padding-alt:0in 5.4pt 0in 5.4pt;
font-family:"Times New Roman";
Videos of the DAC keynote speeches are now available on line.
Gary -- Good point. Xuropa is a pioneer of cloud computing in EDA, and James Colgan was on the DAC cloud computing panel I'll write about shortly. In contrast to IBM, this would be a public cloud approach.
Add to that, Xuropa.com which is lead by James Colgan ... it's a cloud computing service that can massively extend the reach of EDA technical marketing to geographic nooks and crannies like (e.g.) Winnipeg or Gimli, Manitoba, Canada, and probably frustrate a lot of sales and marketing folks' efforts to maintain their silver/gold/platinum status on airlines, hotels and auto rental companies! :)
I missed the keynote, but feel like I was there from your writeup.
It sounds like IBM has implemented a Private Cloud, which is basically adding virtualization (e.g. VMWare) to an existing fixed cluster (20,000 CPUs). This has benefits as you mention, but the real benefit will be to move to a hybrid cloud that keeps a baseline of in house computing CPUs and used the public cloud (e.g. Amazon EC2) for peak needs. And that's where the question of licenses comes in. If you can get 1000 CPUs added for an overnight run, then how do you get 1000 EDA tool licenses for an overnight run? IBM has an advantage that it still does a lot of it's own EDA tools, so licenses are a non-issue I imagine.
Broadcom is doing something similar according to a briefing they gave this week. Since they have quite some influence, I'd expect EDA companies to start to respond soon.