Intel made headlines last year when it announced that it would fab 22nm products for FPGA designer Achronix. This week, the company has unveiled an additional 22nm partner. Santa Clara will also be building parts for Tabula and will build that company's new Spacetime microprocessors. Tabula claims that its new chip "uses time as a third dimension to deliver unmatched capability and affordability. Tabula achieves this breakthrough by combining the Spacetime hardware that dynamically reconfigures logic, memory, and interconnect at multi-GHz rates with the Spacetime compiler that manages this ultra-rapid reconfiguration transparently."
That's so meaningless it hurts. It's a reconfigurable FPGA that presents a lower clockspeed to the end user and performs tasks "in between" the clock cycles. Think of a chip that "looks" like a 400MHz CPU but is actually running at 1600MHz and can dedicate different clock cycles to different tasks and you've got the basic idea. The fact that Intel is handling the manufacturing is the more interesting part of the story; Intel states it has more than just these two customers but can't currently release their names.
Intel's Sunit Rikhi discusses Tabula's technology and Intel's manufacturing roadmap
It's the types of customers that Intel is signing up that gives the best insight into the company's strategic shift; both Achronix and Tabula are small-scale FPGA designers. If Intel was trying to fill excess fab capacity, it would go after big names and more significant contracts.
The company tightly manages technology transitions and its overall fab utilization is above 90%. Money isn't an issue; Santa Clara has lined up record-breaking quarter after record-breaking quarter. Furthermore, pure-play foundries make very little money on their cutting-edge foundry technology. TSMC's Q4 2011 conference call revealed that the company earned 27% of its revenue on 40nm production and 2 percent at 28nm. 65nm production accounted for 30% of its sales, with the remaining 41% produced on nodes more than five years old. If Intel wanted to make a serious play for foundry business, it would start by converting its legacy fabs not its cutting-edge technology.
We suspect Intel's decision to do limited foundry work is driven by several factors. First, there's the issue of validating new process nodes. Historically, new processes have been tested on simple structures like SRAM or NAND flash. As manufacturing has become more difficult, there's been concern that these structures are no longer effective models for transitioning from node to node. FPGA's are more complex than either SRAM or NAND but are less challenging than a chip like Ivy Bridge. As such, they may represent an effective mid-range product.
The other advantage of fabbing chips outside its normal market is that it gives Intel a foot in the door to position itself as a supplier of more than just conventional x86 chips. We've recently focused on AMD's insistence on ISA independence and system-level integration, but both Intel and its erstwhile competitor are likely thinking in similar terms. Like AMD, Intel has no plans to move away from x86 or adopt ARM, but its work on digital radios and Near-Threshold Voltage are examples of technologies with applications well outside of the company's traditional interest.
The final benefit for Intel is that it can more easily negotiate access to any innovations or approaches that come its way. Offering fab space is an interesting way to see what companies are working on and evaluate licensing or outright purchase of the underlying technology. Intel doesn't want to become TSMC or GlobalFoundries, but working with more partners fits in with the company's SoC-centric approach to future CPUs.