PACT offers 80 Pentium4s on a 100Mhz chip
Extremely parallel behavior
PACT Corporation will reveal more details about takers for its 'extreme' XPP processor at the Inquest forum in San Jose this week.
The XPP, claims PACT, has turned in performance of 50-60,000 MIPS at clock speeds of 100 Mhz, equivalent to 80 Pentium 4s.
That gives some indication of how radically different the XPP approach is. It's certainly oversimplifying the design to describe it as an array of VLIWs, as we did, and according to PACT CTO Martin Vorbach it's not really a Von Neumann machine at all.
The XPP's clean sheet design is intended to maximize instruction level parallelism - the holy grail of modern microprocessors, which take up much of their time (and no small amount of energy) doing precisely nothing. Or else throwing instructions away.
The XPP processes streams of instructions through a matrix of execution units, reprogramming the units on the fly. Additional logic determines how the instruction stream flows through the matrix - it can even flow backwards.
"We take blocks or packets of 128 or 256 words and stream them through the array," explains Vorbach. "We can do a lot more operations in once clock cycle."
And while IBM soldiers away welding two cores on its POWER4 processor, and Intel experiments with 'virtual' processors on a chip using SMT techniques, PACT has deployed 128 ALU (Arithmetical Logical Units) on the processor. That means very low clock frequencies - as it's doing a lot more real work in each clock cycle. To save power, each of these ALUs can be put to sleep when not in use.
PACT's XPP requires a radically different compiler design. Distinguished creator of the Pascal language, Niklaus ("pass by") Wirth, was involved in the project to provide a preliminary language NML, that compiler writers write to. A port of the GNU C should be available by the end of the year.
"Even in VLIW there are a lot of data dependencies that the compiler cannot solve," says Vorbach. "We do not depend on one register file - there isn't one," says Vorbach. "there's an I/O buffer somewhere, and data is sent from one ALU just to the next ALU that needs it."
The XPP isn't so efficient where the data sets aren't predictable, of course. That's why it won't be aimed at the Chipzilla's gigantic desktop business. But for media and signal processing work, PACT claims that XPP's performance defies the received wisdom about Moore's Law. XPP will allow telco equipment manufacturers to create 3G base stations that support multiple air interfaces on one chip.
PACT is expected to announce that the XPP will be available for licensing, either as a complete device or in part. Specifically, Vorbach said that DSPs may choose to license the chip and implement their own ALUs, as they have better knowledge of the demands of specific industries.
Currently funded by German venture capital, PACT is looking for US funding but an IPO is not on the immediate horizon, according to Vorbach. ®
Slighty Related Story
Project Jackson - why SMT is the joker in the the chip pack