I see a lot of product pitches, but I rarely see one that creates a new and useful product category.
I see a lot of product pitches, but I rarely see one that creates a new and useful product category. Last month's announcement by a startup called Stretch was one of those rare few.
One of Stretch's founders is Albert Wang, who created much of Tensilica's technology. Tensilica popularized the idea of a configurable CPU: Customers can design their own instruction set to achieve optimum performance on a particular application.
Tensilica and its customers have demonstrated that many algorithms can achieve a 10x improvement in performance by using optimized instructions. The drawback is that the approach requires the customer to create an ASIC.
With tapeout fees approaching $1 million, ASICs are only suitable for higher-volume applications. Customers with lower volumes are forced to make do with a general-purpose processor, sometimes adding an expensive FPGA programmed to accelerate specific functions.
Wang wanted to extend Tensilica's approach to customers that can't afford an ASIC. The Stretch chip combines a standard processor with a configurable data path. Whereas a traditional configurable processor must be carved into silicon, Stretch uses FPGA-like technology to configure its data path on the fly. As a result, customers can buy a standard Stretch chip and configure the data path to their liking, with little effort and no tapeout fees.
Like an FPGA, the data path can be programmed and reprogrammed to test various permutations.
The Stretch data path, however, has a much lower cost structure than an equivalent FPGA. Instead of a sea of gates, it is more like a sea of arithmetic logic units and registers that can be connected in various ways. Additional efficiency is gained from co-locating the configurable data path and the CPU on the same chip.
The best part of the approach is its programming model.
Everything is written in C/C++; a special compiler configures the data path directly from C routines, avoiding the need for gate-level FPGA design.
Stretch's biggest challenge is on the sales side: The company must rack up hundreds of design wins to build a revenue base.
Stretch's technology is appealing for anyone whose application is too specialized for a general-purpose processor but doesn't generate enough volume to attract standard accelerators. In other words, if you've strayed a bit from the well-trod path, this chip's for you.
Linley Gwennap is founder and principal analyst of The Linley Group (www.linleygroup.com/npu).