SAN MATEO, Calif. -- A patent granted to PalmChip Corp. on July 29 covers the techniques used to implement on-chip CPU bus structures in nearly all modern system-on-chip designs, the company said. CPU or system-interconnect intellectual-property vendors and most SoC design teams may be infringing the new patent, PalmChip warned.
"Our first phase is to inform the SoC industry that we have been granted this patent," said James Venable, vice president of marketing at PalmChip (San Jose, Calif.). "We want to give companies that are creating on-chip buses an opportunity to examine their technology to see if there is a conflict and then, if there is, to contact us. We intend to be cooperative about this, not punitive. But of course we will defend our patent aggressively."
U.S. Patent No. 6,601,126 describes an on-chip interconnect structure for linking CPUs, memories and peripheral devices. It uses unidirectional rather than tristate buses, with centralized memory controllers and separate low-speed and high-speed peripheral buses.
As such, it describes the CoreFrame technology marketed by PalmChip. But many SoC interconnect schemes, including ARM's Amba and IBM's CoreConnect, also can be implemented with unidirectional, multiplexed paths and separate high- and low-speed buses.
The full implications of the patent are not in the text, which is a detailed description of the CoreFrame architecture and protocol, but in the claims. The patent makes 16 claims, addressing SoC bus structures that use point-to-point unidirectional connections in place of tri-state buses, structures that are organized to use only a single buffer per line, structures that exchange data through a shared main memory rather than making direct peripheral-to-peripheral transfers and structures in which the low-speed bus uses latches and peripherals with clock frequencies different from that of the bus controller.
With its sweeping claims, the patent captures a critical point in the evolution of the on-chip bus. Hence it is applicable to the vast majority of SoC designs today, according to chip designers interviewed last week.
"In the beginning, when designers first started putting IP intellectual property onto chips, there was a tendency to just migrate whatever system bus had been there when the IP was in separate packages," said Thomas Petersen, director of product strategy at MIPS Technologies Inc. (Mountain View, Calif.). "But almost from the moment people started trying to put tristate buses on chips, they recognized that it was a bad idea. The obvious alternative was to use one-directional connections and multiplexers to emulate the bus, and that's what people did."
There were many good reasons not to use tri-state buses on SoCs, even though some designers continued to do so. The bidirectional buses presented a huge headache to test developers. And in advanced processes, bus contentions were likely to damage the interconnect. That led to logic faults being misdiagnosed as yield issues, further complicating the search for the problem.
But the initial rejection of tristate was for quite another reason. "It was actually the FPGA guys who invented the multiplexer approach," said Leo Petropoulos, director of applications engineering at Tensilica Inc. (Santa Clara, Calif.). "They didn't have tristate drivers in their logic cells, so as FPGAs got large enough to need on-chip buses, they had to find an alternative. I first saw it in about 1991."
As it happened, Petropoulos said, the alternative worked very well with current ASIC processes and tools, too. So it was widely adopted by the SoC design community. By the late 1990s, ASIC vendors were strongly discouraging the use of tristate buffers within the chip, and IBM Microelectronics banned them outright. By then the switch to multiplexers and unidirectional interconnect had become nearly universal.
That time line indicates the enormous scope of the technique's useful life " and hence of the patent's claims. But it also raises a serious question for PalmChip. The company filed the patent on May 2, 2000, after the move from tristate to multiplexer-based buses was essentially complete. The patent may be subject to the objection that it attempts to claim structures that were already common industry practice well before the date of filing.
The section of the PalmChip patent on prior art contains a lengthy discussion of the technical needs and market challenges that made CoreFrame an attractive technology. But it does not explicitly cite the widespread use of the sorts of structures described in some of the patent's claims.
In fact, the number of claims and the specific nature of some of them suggest that the patent examiner was concerned about prior art, said Kenneth Allen, a patent attorney with Townsend and Townsend and Crew (Palo Alto, Calif.). "Often, individual claims come from discussions between the examiner and the patent attorney," Allen explained. "The examiner may cite what appears to be prior art, and the attorney will respond with a narrower claim that makes it clear why the prior art is not relevant."
Early in the history of SoCs, any integration of IP onto a single chip was patentable, Allen said. But as SoCs proliferated, examiners began to find that many integrations were obvious, and therefore not patentable. So patent attorneys had to show, through detailed claims, that although the idea might appear obvious, there were certain technical problems, the solutions to which were not obvious.
Even so the results are far from guaranteed. "Examiners are allowed about eight hours per claim," Allen said. "They work from a precise library of prior art, not from a detailed knowledge of everything that has gone on in a field of study."
Allen emphasized that in understanding what a patent was really trying to accomplish, it was important to look not just at the text, but at the claims, and to compare the final draft with the initial one. That makes it possible to see the evolution of the claims through the give-and-take of examiner and attorney.
PalmChip's officers made it clear that they believe the new patent, if properly understood, is free of prior art. "This was absolutely the first use of unidirectional bus structures in a system-on-chip application," asserted CEO Jauher Zaidi. Venable emphasized the SoC angle. He said that while before the PalmChip patent the most common implementation of on-chip buses had been tristate, a variety of techniques had been used in other sorts of chips, such as microcontrollers. But in the SoC arena, PalmChip had priority, he said.
That, some might argue, hinges upon the definition of "SoC." And in this area the PalmChip patent is distinctly vague. The application defines its scope in these terms: "The present invention relates generally to flexible modular integrated circuit embodiments, and more specifically to a structured framework for constructing and interfacing variable mixtures of computer processors, memories, peripherals, and other functional modules on a single semiconductor integrated circuit."
That definition would apply just as well to a family of microcontrollers as to a chip more conventionally thought of as an SoC. This is not surprising, given the lack, even today, of a widely accepted definition. But the question of prior art may put that disputed definition at the heart of the efforts to apply " or defeat " the PalmChip patent.
If the impact of the patent on existing silicon bus architectures and SoC design practices is potentially great, its effect on the future of SoC design appears to be more limited. "Today we are in a predicament where the conventional, multiplexer-based approach to buses is causing congestion and severely straining the layout tools," Tensilica's Petropoulos said. "People are starting to look away from time-domain multiplexed interconnect schemes altogether, and to think about packet-based interconnect between major blocks on the SoC."
In fact, the evolution of interconnect between blocks is following a path very similar to that of the public voice telephone network. From shared, time-multiplexed buses interconnect is moving toward more complex structures. "The idea of a crossbar switch for connecting up blocks has been around forever," observed MIPS' Petersen. "But now chip architects are starting to apply the concept of quality-of-service to on-chip crossbars. And there is a lot of thinking about packetized interconnect as well."
For Petersen, "A lot of the key ideas really crystallized in the HyperTransport standard. There you have packetized communications, virtual channels " pretty much everything you need on-chip. It's not out of the question, if interconnect keeps getting more difficult to model, to see some TCP ideas like error detection, retries and a media-access controller for each block in the design. The overhead would be well worth it if we could have really standard interconnect and plug-and-play IP."
PalmChip, too, is looking ahead.
"In fact, as part of the development of our next-generation architecture," Zaidi said, "we have filed additional patents covering key concepts in the use of crossbar switches and matrix networks in SoCs."
So it appears that PalmChip's claims, and the arguments over them, are far from settled.