ClearSpeed scales down dramatically, CEO resigns [UPDATED]Thursday 12th February 2009, 07:50:00 PM, written by Arun
- There seems to probably be a plan to still sell the current generations of products (latest was launched back in June) although I didn't get any details. Either way, if you are directly or indirectly a ClearSpeed customer, we do not believe any recent event should be affecting you in any way you weren't nearly certainly already aware of. We are deeply sorry if we made any third party panic for no good reason.
- ClearSpeed is not bankrupt, still has cash in the bank, and the company itself is not being closed down despite its intentions of no longer being quoted on the stock exchange.
- The company still wants to capitalize on its IP going forward, although not with the exact same business plan for obvious reasons.
- The company has made no formal announcement about how many employees remain or how many it wants to keep in the future.
For what it's worth, we still believe the news piece below remains well worth reading, especially as to the history of ClearSpeed's technology and architecture. We again apologize for any confusion it might have caused however, and wish the best of luck to what remains of ClearSpeed.
ClearSpeed, a parallel computing company based in the UK and a direct competitor to GPGPU, seems to be giving up on its current market: following revenue of only £0.5M in 2008 and losses of £10.4M, the CEO has resigned and our info tells us that nearly all remaining staff are being laid off (although much was already gone). In fact, the company is even talking about removing itself from the stock exchange and returning surplus cash to shareholders! It is definitely not bankrupt per-se, but that does seem to imply normal operations are being discontinued.
ClearSpeed has a very long history, and interestingly enough they started in the 3D Graphics workstation industry. They were named PixelFusion back then, and their Fuzion 150 GPU was, to say the least, a very exotic design. It was an evolution of the Pixel Planes architecture they (along with a few others) licensed from the University of North Carolina, and it was engineered in Bristol which remains to this day a center of innovation in parallel processing due to the heritage of the legendary Transputer.
It had 1536 8-bit Processing Elements (divided in 4 instruction streams and multiple threads), 24Mbits (3MBytes) of eDRAM, a Rambus RDRAM memory controller, 76 million transistors, and a 500mm² die size on a 250nm process... all that back in 2000. The company took longer than expected to execute on that first chip, and without enough momentum to keep going the company renamed itself to ClearSpeed, switched to the network processing industry, and tried to license IP instead of focusing on a fabless business model.
Without much more success, they then moved to the High Performance Computing industry and sampled a 130nm chip in Q1 2004 that only had 64 PEs, but each with FP32 support. It only had 41 million transistors, and most striking was its truly excellent power efficiency with a 3W TDP at 200MHz/25.6GFlops. In Q4 2005, they sampled a refresh with 40GFlops, 96 PEs, 128 million transistors, and a 10W TDP. Most importantly, the focus slightly changed again: it now had full-speed double precision processing! In June 2008, they announced a 90nm refresh that improved performance to 96GFlops (192 PEs @ 250MHz) and kept typical power consumption at 9W.
ClearSpeed's architecture is a relatively classic massively parallel one with a very wide vector size programmed in a variant of C. In terms of programming flexibility, it could be argued that it offered less than GPGPU, not more. It does have a scalar processor core on-chip for control logic to compensate for its awful branch granularity, which is very attractive in theory, but it didn't seem to make a huge impact on the programming model. Where it did win hands down was in terms of reliability (ECC, redundant PEs, fanless, etc.) and power efficiency. For workloads such as DGEMM, its efficiency also seems very good (75 GFlops), although it's not clear how it'd look for more complex workloads where not as much attention can be put into every line of code.
Also back in June, ClearSpeed indicated they'd skip 90nm and go straight to 45/40nm for their next-generation chip. They claimed to already be working on it. Sadly, the dream now seems to be over (in fact, one of our sources said hardware teams were likely shredded much earlier): with very little revenue (substantially less than Tesla's GPGPU group, even though they don't have all that much yet either!) and no clear path to profitability or further financing, they can't do anything but stop their operations completely.
We suspect the reason they likely didn't insist on their situation too much is that they didn't want to scare customers who were already programming on their chips but weren't yet purchasing their products in any quantity. We don't see why they should have switched either, unless they really care about having a multi-year roadmap. The company had already cut much of its staff in early 2008, but still kept support/sales and a relatively small R&D operation. ClearSpeed also had an IP license deal with BAE to use their processor architecture in a satellite-based system; because of the requirements for that kind of chip, such a deal appears normal. However it appears unlikely to us that ClearSpeed could build a long-term business on IP licensing, since that would require a number of further IP deals being signed...
We wish good luck to what remains of the company, as well as to their former employees. There's still plenty of exotic processor R&D going on in Bristol, so here's hoping the intellectual legacy of the company will live on somehow. And for those more interested in naked chips (coming mostly from this presentation - in order from Fuzion to CSX 700) than long paragraphs, enjoy: