Sun chip chief: real men can't afford fabs
The Itanic Verses
Sun has responded to increasingly impressive Itanium benchmarks by conceding that the leviathan chip might have a niche. But its microprocessor CTO offered the counter argument that manufacturing economics will favor smaller fabless chip developers, such as, er ... Sun. This is interesting, as it's an attempt to derail the volume economics logic - more is cheaper - that has sustained Itanium through its darkest hours.
We'd actually called to see what Sun made of Intel's LaGrande plans to restrict transactions by embedding certificates in the hardware, but Mike Splain, CTO of Sun Microsystems microprocessor group and Sun fellow, ranged far and wide, and we'll come to LaGrande last.
Fill the fabs
At a press session in San Jose last week, The Fist claimed Itanium was only 25 per cent behind POWER4 despite only a couple of years on the market (and ten years in the dressing room). Splain told us that Intel's business model, coupled to the lack of demand for the very fastest processors, would hurt Chipzilla.
"Our fab partner Texas Instruments buys the same tools from the same people as Intel does, such as ASM Lithography and Applied Materials," he said. With new wafer technologies increasing the capacity of a plant exponentially, recouping investments was increasingly difficult from PC markets alone, he said.
"When the cost is $3.8 billion per fab, you need some kind of business model to keep those fabs full," he said, suggesting non-PC markets such as cellphones. (Coincidentally TI is the most popular platform for high-end phones.)
In fact faster chips were the last thing computing needed, he said, which is a funny thing for a microprocessor chief to say, when you think about it. But he's correct: the choke point isn't the processor but the I/O and memory.
"The faster processor simply waits faster - we're all waiting for something to do."
But hadn't Intel tried to address with Rambus?
"We were never a supporter of Rambus: it didn't address the core issue. If you remember the old bipolar community; they all jumped to ECL, and all said that was the way forward. Then two or three years later that house came crumbling down; and that happened to Rambus, too."
At the same time he rejected the view that chip performance was hitting a ceiling, after RISC and small off-chip caches, out of order execution, and memory improvements:-
" I think we haven't hit the saturation points, we have a long way to go," he said, suggesting that multiple cores and multithreading offered plenty of potential for faster processing.
As you'd expect, Sun is focusing on Itanic's extraordinary power consumption as a weak point.
"You're not going to see Itanium in blades, and people will rack up Banias processors," Splain predicted.
He did concede "Itaniums will have their place", and said it performed well on specific benchmarks.
"It's respectable on linear algebra - but you don't see widespread software support, there's the issue of die size and power consumption, and no one's paying attention to it. Intel went to 64 bits for larger memory and may still extend the x86 instruction set to get to that larger memory," he said.
Sun blows hot and cold on the subject of Hammer - it all depends on the context.
We'd actually called up to hear what Sun had to say about Intel's announcements to embed share denial technologies into the hardware. LaGrande is the hardware companion to Palladium, and even Intel's mobile processors will embed certification information in the chip.
Right or Wrong? Splain said Sun has no plans to ape Intel's TCPA architecture even though it has used unique processor IDs in SPARC for some time. The ID contains information on when and where the processor was manufactured - including its position on the wafer - for quality control purposes. There was no userland interface to this information, he said, which was only readable by the kernel, with no OS call available to obtain that code.
"We're not looking at it in the same degree," he said of TCPA, which is very much a PC industry creation. At the same time Sun was looking how to secure transactions: "at the system level we've looked at much more exhaustive techniques, that speaks to data on the disks, and to how data is moved across network. It doesn't completely relate to what we do on the processor."
Splain can be grilled at an open mic spot on Wednesday at Sun's Network Conference, alongside other Sun CTOs. This is a revival of SunWorld, a three day forum showcasing all kinds of Sun wares, much like Intel's very successful Developer Forum. Although there really isn't much of a SPARC OEM business - and with Java flying under its own flag, with JavaWorld well established as the biggest show in town - you have to wonder why it took Sun so long to revive the idea. ®