software-defined silicon, a new alternative

In general, publicity is good thing for a company – so it is nice to see XMOS Semiconductor finally emerging from stealth-mode and providing some information to the public on their company and what they claim to be a revolutionary new type of programmable semiconductor technology called Software-Defined Silicon (SDS). Now don’t get too excited, the web site is still relatively sparse on details, but a quick email to XMOS resulted in a nice presentation style response that provided more insight. Founded in 2005 and backed by Amadeus Capital Partners and Esprit Capital Partners, XMOS currently consist of a team of 25 people out of Bristol, England. The company is headed by James Foster, who has previously held a CEO position at Oxford Semiconductor and before that held various engineering positions at Lucent and Lattice Semiconductor. The CTO for the company is David May, who is well known for his invention of the Transputer and the OCCAM programming language

According to XMOS, SDS will provide designers with the advantages that are usually associated with System on a Chip (SOC) solutions while at the same time providing the flexibility that FPGAs offer. SDS is implemented as an array of XCore event driven multi-threaded processor elements which are linked via an XLink inter-core communication link, both of which can be easily programmed and configured through XC - an XMOS C-style proprietary language. The advantage of this approach is that the whole system can be specified using a high-level language and no more RTL coding is necessary. Further, the whole system can be reconfigured in about the same time that it takes to recompile the actual code. The chart below summarizes the advantages of SDS versus competing technologies:

Designing ASICs is very risky unless you have really high volumes. ASSPs are less costly, but since they are sold to many companies, differentiation becomes difficult. FPGAs have almost everything going for them - except that they are very silicon inefficient, which makes them expensive and utterly unusable for high-volume consumer products that have very low margins.  The fashion criterion displayed in the chart is utterly beyond me, but as you can see SDS is fashionable as well. On the performance front, the XCore is expected to yield a performance in the 500 MIPS range. Also, developers will be able to dynamically partition resources between control processing (MIPS), DSP processing (MSPS), and I/O processing (Mbps). The key to SDS success will be whether or not XMOS will be able to deliver a robust and intuitive programming environment; without this ability most developers will likely shy away since there is nothing more frustrating than dealing with buggy tools. More technical details of the architecture and tools are scheduled to be released in Q4 of 2007 and the launch of the first device family with the complete tools set and IP library is to follow in Q1 of 2008.