In the last few years, there have been a slew of massively parallel chip vendors entering the embedded processor market. The massively parallel approach has become more accepted since Intel commercialized its multi-core PC architecture. It’s still a difficult area in which to build a successful business, however, because it requires not only creating a good architecture, but also developing a sound programming model and competent development tools. Even traditional processor start-ups fail at a prodigious rate, and massively parallel chips have the additional challenge of being perceived as hard to program. So perhaps it’s not surprising that several massively parallel chip vendors have recently fallen by the wayside: Ambric and MathStar, both of which have ceased operations, and Rapport, which is in the process of being acquired for some of its assets and whose processors will probably never be sold as such. Though their endpoints are similar, each of these companies had its own challenges and reasons for failure.
Ambric offered massively parallel chips with up to 344 heterogeneous processors. Of the three companies, Ambric was the one with the most compelling business story and the best shot at success. Recognizing that ease-of-use was the key stumbling block for massively parallel devices, Ambric focused on programmability from the outset of its design. Even though the designers chose Java as the architecture’s programming language (an odd choice for the embedded market) they were able to attract customers with their relatively straightforward programming model. The technology was solid and Ambric appeared to be gaining momentum—until last month, when Ambric abruptly announced that it was shutting down its operations and would be trying to sell the company.
Ambric’s story is particularly frustrating since the company was shipping silicon and says it had multiple design wins in four markets: video, defense, wireless base stations, and medical imaging. According to
MathStar and Rapport didn’t come as close to success as Ambric. MathStar, which was founded in 1997 and ceased operations earlier this year, was selling chips based on the company’s “FPOA” (or “field-programmable object array”) architecture. The chip contained an array of 400 functional units, which included ALUs, multiply-accumulate (“MAC”) units, and register files (rather than an array of coarser-grained processor cores). MathStar claimed that its medium-grained approach provided much of the flexibility of FPGAs while offering a simpler programming approach and higher clock speeds. In late 2006, MathStar announced that it would start shipping silicon, but apparently the company burned through over a hundred million dollars in capital without getting a firm commitment for a high-volume order.
Unlike most vendors in this space, MathStar was a public company. From transcripts of a shareholders meeting last June, it appears that the company hit some common pitfalls: its tools were unfamiliar to users and were not fully mature; it needed to provide more complete application solutions, but came to that realization too late; and it wasn’t able to meet customer requirements in the current silicon version and so needed to migrate to 90 nm. MathStar hoped to attract FPGA users who would otherwise go with chips from Xilinx or Altera, and was trying to do so with a fundamentally different approach in terms of architecture and design methodology. Even if its solution and tools had been top-notch, that’s a tough hill for a start-up to climb.
The third company, Rapport, created the ”Kilocore” architecture that used large arrays of homogeneous 8-bit processing elements and targeted mobile media and defense applications, among others. Rapport’s value proposition was straightforward – it claimed to offer chips with lower power and better performance than competitors. But the company didn’t appear to have convincing data to back up the story, and didn’t seem to have a clear strategy for getting its chip into target applications. Based on previous information on its website, Rapport was also quite top-heavy, with an unusually large number of executives for a company its size.
As of this writing, Rapport has not made a public announcement about its status, but its home page is no longer active, the executive management team of a year ago is gone (the company has a new CEO), and the company’s last press release was dated October 2007—in which it stated that it had raised $18 MM. BDTI contacted Rapport for an update on the company’s status, and was told that the company is in the process of being acquired by a company called Personal Web Systems, which will use some of Rapport’s programming tools and chip design building-blocks. The Rapport processor, however, appears to be dead in the water.
In BDTI’s opinion, it’s unlikely that MathStar or Rapport would have been successful, regardless of the state of the economy. These companies simply didn’t have strong, realistic strategies. Unfortunately, their demise—along with that of Ambric—may make it less likely that system designers will take a chance on other massively parallel embedded processors.
That would be unfortunate. We continue to believe that parallel processing is the way of the future for a broad swath of embedded applications. But it’s clearly going to be a bumpy road getting there. The best innovations often come from small companies, but even in ordinary times many start-ups with good technologies fail, while others soak up market opportunities that by rights should have been awarded elsewhere. Perhaps the best way to view these challenges is with a Darwinian perspective: in the coming years, we hope that a few truly fit vendors manage to survive.
Add new comment