informa
/
7 min read
Commentary

Interview: Intel Talks Multicore Processor Trends, Chip Tech

Get ready for a new battle in the multicore wars. AMD is planning to ramp up production of its next-gen 32-nm processor in mid-2010. This follows on the heels of Intel's revelation that it'll be first to ship 32-nm parts with its Westmere chips late this year. What jumps out at me most here isn't that this is great news for consumers -- it is -- but that there's a big element of saber-ratt
Get ready for a new battle in the multicore wars. AMD is planning to ramp up production of its next-gen 32-nm processor in mid-2010. This follows on the heels of Intel's revelation that it'll be first to ship 32-nm parts with its Westmere chips late this year. What jumps out at me most here isn't that this is great news for consumers -- it is -- but that there's a big element of saber-rattling going on. That's because the 45-nm chips shipping from both vendors are still new themselves and far from obsolete. Read on for my interview with Intel's Stephen Smith, where he discusses new chip-making technology, server processors, and netbooks.On the desktop side, those 45-nm parts include the Core i7 family from Intel and AMD's Phenom II processors. For servers, AMD is shipping 45-nm Opterons and Intel has its Xeon line.

Nevertheless, when it comes to processor marketing, it's never enough to emphasize what you've got. The competitive imperative necessitates a constant game of one-upmanship, where the future road map becomes a big focus. Which is fine; we're all interested in next year's model.

What I think is different this time -- as opposed to the battles of 2004/5 when the first dual-core parts launched -- is that we've reached a technology watershed in chip design and manufacturing. Until a few years ago, chipmakers were reaching the end of the line in terms of the additional performance improvements they were able to extract from their designs. This was because of the limits of the materials used in the chips. Then Intel came up with a new material -- called Hi-k plus metal gate transistors -- and AMD partnered with IBM to push forward into 32 nm.

Now, it's a whole new ball game with a clear runway to a future filled with not just quad-core processors, but six- and eight-cores and more.

OK, all this is by way of an excuse to post an interview I did a couple of months ago with Stephen Smith, VP of Intel's Digital Enterprise Group, where he gives some serious perspective to several of these threads

Wolfe: Why is this new chip-making material important?

Smith: We've been building a similar transistor structure for almost 40 years. We've gotten a great 40 years out of that, but we needed to fundamentally reengineer the materials of the transistor to continue scaling. And with this new Hi-k and metal gate transistor, we can actually continue deliver scaling -- meaning smaller physical size -- and more performance. And we have the next three or four generations of silicon worked out. So every two years, we expect to get a doubling of transistor density, and some progression in performance per amount of power we spend.

So this new transistor was critical in enabling us to continue to move on Moore's Law in a way that's power efficient.

Wolfe: In 2004/5, dual core was new, and then about a year ago we went to quad core. So what is the sweet spot of the market now?

Smith: What we see becoming commonplace among all computers, except for the very lowest cost, is dual-core capability. Dual- or multicore is a more energy-efficient way to deliver performance.

The people who are most demanding in performance have progressed to quad core. We have very rapid adoption in servers and we have strong adoption in high-end desktops and workstations. We've seen quad core migrate to what I'd call mainstream price points by now, but more value-priced computers are still dual core.

The big thing we're talking about now [this was in December-AW] is our second-generation 45-nm, with Nahalem. Core i7 is its product brand name. That's a design that's fully optimized from the ground up for 45 nm and will give us an even better quad core because we have two threads per core -- hyperthreading -- so even more parallelism within the same thermal constraints.

We have integrated the memory controller onto the Core i7, and we have three channels of DDR3 memory. That gets us almost three times the memory bandwidth that we had in our prior generation desktop platform. What we're doing is systematically looking at the platform and resolving any bottlenecks. So in this generation, we talked about memory, 3X bandwidth, huge performance increases.

Wolfe: What about on the server side?

Smith: At the processor level, we're offering primarily quad core in our volume server products, which are targeted for two-socket servers. And those are probably well in excess of 80% of server volumes. The majority of what we sell is quad-core 45 nm.

For the larger servers -- the ones with four sockets -- we're shipping the Xeon 7400 Series with six cores per processor. And then when you build a four-socket server, you have 24 cores running that the OS has as a physical resource. This is the best platform to use for virtualization because you have all those physical resources available to virtualize.

What also have on our road map production committed for the second half of 2009 for a Nehalem processor for high-end servers. It's called Nehalem EX, for expandable and scalable servers. It's an eight-core design with four Quickpath links. Quickpath is a scalable interconnect which will allow us to build higher-capability processors and systems. So for example an eight-core Nehalem has two threads per core, which gives you 16 logical cores. And with those four Quickpath links, you can get a direct link amongst all the processors. Therefore you have a 64-thread machine with a direct physical link from each processor. We've also shown the concept that, with one hop, you can build an eight-socket machine.

Wolfe: What else is on tap?

Smith: We are working on a future product called Larrabee. It's an array of cores. We expect to ship that first as a discrete graphics capability. But it's also an architecture that's going to be very useful for scalable compute applications.

We want to move workloads to an off-chip accelerator. That's the Larrabee approach, and it's going to be much more accessible to the programmer because it's an array of x86 cores and there are millions of people in the world who know how to program those cores.

Wolfe: Tell me about the netbook side, where you have your Atom processor.

Smith: For Atom, we took a completely different approach. Rather than have our mainstream architecture scale even further, when we move to this new class of device, we need to move to sub-5-W power. We took a clean sheet of paper approach. We designed a machine that's targeted to get down to sub-watt thermal performance levels that you need for a mobile Internet device or I'll call it a few watts that you need for a netbook or for a fanless embedded machine. So we used the latest technology, 45 nm, precisely because that gives us the best energy efficiency. But we did a ground-up, purpose-built machine that we could configure as either a single-core or a dual-core product. Uses the existing chipset infrastructure and has been very rapidly adopted by these netbooks.

Wolfe: Is Atom stealing sales away from laptop processors?

Smith: Adoption has been faster than our original forecasts and we actually had to react to put in place more assembly and test capacity to get them out the door. Netbooks have moved very quickly. It really hasn't had any impact on our traditional notebook business. We see it as a new category and a new line of business for us.

Please leave a comment below or write to me directly at [email protected].

Subscribe to my blog's feed, here.

For a mobile experience, follow my daily observations on Twitter.

Check out my tech videos on this YouTube channel.

Alex Wolfe is editor-in-chief of InformationWeek.com.