Yen Explains Sun's Chip Strategy

Published: April 12, 2007

by Timothy Prickett Morgan

David Yen, the long-time chip guru from Sun Microsystems who is often moved around by the company's top brass to troubled business units, has taken the helm of a new, independent Microelectronics group. Sun has announced its first OEM customer for chip technology, and that was a good excuse to get Yen on the phone and ask him what the real deal is behind Sun's moves.

As we previously reported two weeks ago, Sun broke its chip design unit away from its System group to form the Microelectronics group. Yen, who has been shepherding Sun Sparc processors from the whiteboards to the motherboards for the past decade, was removed from that position to take over Sun's storage business in the wake of its acquisition of StorageTek. Having killed off a bunch of products and integrated the sales and tech teams of Sun and STK, Yen has now been asked to go back to the chips side of Sun and turn it into a business unit that has more than one customer (in this case, Sun itself) and offers products and services that leverage Sun's substantial expertise in chip design and its resurgence as a thought leader when it comes to chips, peripherals, and system designs.

A year ago, Sun took some steps to reposition its chip business while Yen was running the storage business, and it seems highly unlikely that anything was done without talking to him first, even if he was not running the Sparc chip unit. Perhaps the most significant step was to create an open source version of the multithreaded "Niagara" Sparc T1 processor. All of the key design specs plus the hypervisor microcode and the documentation for the Sparc T1 have been made available through the OpenSparc project.

Some people are noodling around with the OpenSparc specs. Canonical, the creator of the Ubuntu Linux distribution, used the OpenSparc materials to do a rapid port of Linux to the Sparc T1 chip--something that would have taken a lot longer had Sun not given this material away. In the past incarnation of Sun, it would have been hard to imagine the company actually courting a Linux port. And, it is inexplicable why Sun isn't interested in courting Red Hat and Novell to do ports to the Sparc T1 chips ASAP and to have Linux running on its future "Rock" enterprise-class, multithreaded processors, which are due next year.

The first true chip spin-off from the OpenSparc project was announced last October, when the "Sirocco" S1 chip from Anglo-Italian chip maker Simply RISC started shipping. The S1 chip was created by taking a single T1 core (the real T1 chip has eight Sparc cores, each with four threads) and implementing it with a Wishbone system-on-a-chip interconnect bridge. Simply RISC is hoping that the S1 chip is adopted for mobile phones, PDAs, and other embedded devices.

Sun could, of course, be hoping for a little of that action itself, and Yen made no bones about the suggestion that Sun might want to do a variant of the Sparc T1 aimed at cell phones, PDAs, and other handheld devices. "You never say never these days," he explains, adding that what Sun is most interested in is increasing the Sparc base. "But Sun is not a job shop. We are not going to design a Sparc chip for HDTVs. But we are willing to do something that might be less financially rewarding to grow the Sparc base." As an example, Yen said that it might make sense to carve out a single core Sparc chip with a single thread and create a field programmable gate array (FPGA) co-processor for systems out if it. "Sun has over 1,000 chip engineers, and we are trying to leverage the substantial design and packaging expertise we have in the chip business."

The deal that Sun announced since Yen took over the Microelectronics unit has nothing to do with the Sparc, actually, but rather on a multithreaded ASIC, code-named "Neptune," developed by Sun for supporting multiple 10 Gigabit Ethernet network interface cards that Sun has licensed to Marvell Technology Group, one of the big players in the manufacturing of networking and communications chips and peripherals that brought in $2.2 billion in sales in its fiscal 2007 ended this January of that year. (Marvell is in the middle of a stock options review and has delayed its quarterly filings for the past three quarters, and has grown smartly in the past several years as wired and wireless PC connectivity has taken off in the consumer and corporate PC markets.)

Marvell has its own chip design teams, and is a licensee of the ARM RISC chip that was acquired by Digital Equipment and after many passes around has ended up at Intel as the XScale chip. Marvell's own ARM variant, called Feroceon in its most recent incarnation, as well as Intel's XScale, which Marvell helps Intel design, are used in all kinds of electronics--PDAs, printers, cell phones, and so forth--where power consumption is a bigger issue than raw performance, but where there has to be a baseline of acceptable performance, too.

The old rule in the systems business is that you need a megabit of network bandwidth for every megahertz of processor clock speed to have a balanced interaction between processors and network interfaces. Processors have hit a wall at around 3 GHz (excepting IBM's forthcoming Power6 chip, which will scale from 3.5 GHz to 5 GHz), but network interface cards are already heading toward 10 gigabits and will soon be running at 40 gigabits and soon thereafter at 100 gigabits. The networks have a lot more bandwidth than the cycle time on a processor. And that means you have to break the network interface up so it can provide adequate bandwidth to each thread and core in a processor through multiple direct memory access (DMA) channels. This, in short, is what the Neptune chip does. Marvell is a chip maker as well as a network interface supplier, and as part of its deal with Sun, the server maker has not only licensed Neptune to Marvell for its own products, but will also be a contractor to supply Sun with its own networking cards.

"The Neptune chip is very big, specifically because it is a general purpose device and not tied to any specific instruction set," Yen says.

While this is all interesting, the real task facing Yen, which I hinted at in the Microelectronics group announcement two weeks ago, is to sort out how Sun will fabricate its future processors and other chips. For two decades, Texas Instruments has been Sun's foundry for Sparc processors, but TI has announced that it will stop creating new chip making processes and running chip fabrication factories--going fabless, as they say in the industry--in the 32 nanometer generation. TI, like Sun, can't afford to make the investments necessary in the chip racket, and so TI is looking for a fab partner for its digital signal processors and the other chips it makes.

The Sun-TI partnership worked well for both companies because Sun needed to have cutting-edge chip fabrication processes to get the highest possible performance out of its generations of Sparc processors. There was some dubs that got flubbed, to be sure, as always happens in the chip business--the UltraSparc-III chips were late to market in the late 1990s and early 2000s when Sun could least afford to lose its performance edge, for instance. And TI needed chips to put its most advanced chip-making processes through the paces before rolling it out to high volume DSP chips.

The good news is that the work on 45 nanometer processes that Sun plans to use and TI will implement is largely done, according to Yen. "We certainly need to work on this issue," says Yen. "The 45 nanometer generation is mostly done. The issue is more with the 32 nanometer generation.

Fujitsu was briefly a second source for UltraSparc-II chips back in the 1990s, and has been late shipping the Sparc64 VI processor used in the shared Advanced Product Line of Sparc-based servers that Sun and the various incarnations of Fujitsu will announce next week. Yen says that Fujitsu's fabs are relatively small, but that there are possibilities. Advanced Micro Devices could be a chip fab partner for Sun, of course, but AMD is a chip process partner with Sun's rival, IBM. And Sun is now readying servers using Intel's Xeon processors and could cut a deal to have Intel fab its chips, as Digital Equipment and HP did for some time. Yen listened to all of these ideas and didn't shoot any of them down or endorse them.

"We need to take a lot into consideration. It is a complicated equation," says Yen.

Yen doesn't think there is much of a chance that the UltraSparc-IV or Rock designs will be licensed, and he is not basing the Microelectronics group's business plan on it. "If IBM or HP want to use Rock, I am all ears," says Yen with a laugh. "And John Fowler, who runs Sun's Systems group may not like it, but that is why the Microelectronics group is independent." Yen does think that there is the possibility that there can be lots of derivatives of the Niagara design, aside from Sun's own future Niagara-2 and Niagara-3 chips.

RELATED STORIES

Sun Boosts Performance of UltraSparc-IV+ Chips

Sun Breaks Sparc Unit Free Again

Will 45 Nanometer Chips Make Two Warring Camps?

Ubuntu to Support Linux on Sparc T1 Chips

Sun Merges Server Units, Taps Key Exec for Storage

Sun Lays Off Workers in Sparc Server Group

Sun Releases OpenSparc T1 Specs, As Promised

Sun to Take New T1 Sparc Chip Open Source

UltraSparc-IV+ Chips Give Sun's Midrange Servers Twice the Oomph

John Fowler Appointed to Head Sun's Network Systems Unit

Sun Reports $760 Million Loss, Shakes Up Executive Leadership

Sun Pulls the Plug on Future Millennium, Gemini Chips

                     Post this story to del.icio.us
               Post this story to Digg
    Post this story to Slashdot