It looks like the makers of field programmable gate arrays (FPGAs) are not quite ready to cede the coprocessor market to GPUs just yet. Achronix Semiconductor, an FPGA maker based in San José, just down the highway from chip giant Intel, has announced that it will use Chipzilla's fabs to cook up a future generation of chips that will "extend the boundaries" of these exotic semiconductors.
In a blog posting, Intel spokesman Bill Kircos confirmed that Achronix has scored some space in Intel's chip fabrication facilities, and that FPGAs designed by Achronix will be baked using Intel's future 22 nanometer process. That will be Intel's third generation of hafnium-based high-k, metal gate semiconductors, and the company expects to roll out chips using this 22nm variant in the second half of 2011.
Intel is well known for being on the cutting edge of wafer-baking tech, and this is a big deal for Achronix, which is looking for an edge over Xilinx and Altera, the dominant producers of FPGAs today.
But Intel wants to be clear that they aren't necessarily going to transform into a hybrid wafer baker that does contract manufacturing for lots of other companies as well as making its own chips.
"Our manufacturing and factory network are our prize possessions," Kircos said in an email to El Reg. "For proper perspective, this agreement makes up significantly less than one percent of Intel's total capacity and is not financially material to our business." In his blog post, Kircos said that Achronix was getting selective access to its fabs and that even though it was a small amount of money, "it's still an important endeavor for us that we're committed to deliver on."
The reason that Achronix needs a wafer baker in the States has to do with national security. The US military and contractors that like to build weapons and spooky computer systems do not like using technologies developed outside of the United States. Achronix could go to Taiwan Semiconductor Manufacturing Corp, its current fab for the Speedster1 line of FPGAs, or it could have tried using IBM Microelectronics to get something that was redder, whiter, or bluer.
If AMD hadn't sold GlobalFoundries, you could paint a wafer baking deal as sort of a purple smear with some green thrown in for their Dresden, Germany foundries.
But this is about more than flag-waving and chasing government money. Achronix was founded in 2004 by researchers at Cornell University who wanted to push the performance limits of FPGAs and change this estimated $3bn market, challenging the market leaders Xilinx and Altera.
FPGAs are, as the name suggests, malleable and can be rejiggered to change their basic functions in ways that an ASIC cannot. It might take $30m or $40m to develop an ASIC to do a particular job — say, support the Ethernet or InfiniBand protocols — and if you make a mistake, you cannot erase and go back.
For very high volume products — with hundreds of thousands to millions of units where the cost per unit has to be low — you want an ASIC. But in places where you need a chip that might only require thousands to tens of thousands of units to satisfy an entire market, an FPGA, while more expensive to buy, is better because it is less expensive to make and is correctable in a way that an ASIC is not.
According to Greg Martin, a spokesman for the FPGA maker, Achronix can compete with Xilinx and Altera because it has, at 1.5GHz in its current Speedster1 line, the fastest such chips on the market. And by moving to Intel's 22nm technology, the company could have ramped up the clock speed to 3GHz.
But such a chip would have been a lot hotter, and thermal constraints in niche places where FPGAs are used (telecom and networking equipment as well as a smattering of supercomputers and server appliances such as Netezza's data warehousing appliances) are just as real as they are in the server racket these days.
And so the future Speedster22i FPGAs will run at the same 1.5GHz. They will come initially with 1 million look up tables (LUTs), which is akin to the gate count in an ASIC, and the Intel processes will allow Achronix to pump that up as high as 2.5 million LUTs, which is functionally equivalent to an ASIC with 20 million gates.
Martin says that none of Achronix' customers are quite sure how they would use such an FPGA yet. But you don't need a supercomputer to predict that if someone can build it, someone else can use it.
By the way, the goal is to bring the cost of that 1 million LUT FPGA down to around $400 a pop when they start shipping in the fourth quarter of 2011. FPGAs sell for $1,000 and higher today, depending on features.
It will be interesting to see if Intel will open up its fab to other players to try to cover some of the billions of dollars it spends each year on wafer bakers. ®