Business, Hardware

The Bristol Curse: How the UK lost its CPU technology edge

Many of our elder readers, those in their 30’s and 40’s, may remember a really lovely microprocessor family – a kind of wet dream from our days of youth. Transputer was something truly special for a CPU some 25 years ago: they went 32-bit at the same time as Motorola 68020 and Intel 386, had comparable performance despite a unique 3-register & on-chip memory model, and… Every CPU had four full speed links to other CPUs & beyond, something you only see in Intel CPUs in 2007 [even the ever-ahead Alpha only had such stuff in Year of the Millennium Bug 2000]. The 32-bit Transputers of the mid-’80s, the T414 and T800 were followed by a delayed T9000, and then, supposedly after certain funding cutbacks, there was nothing more.

Inmos Ltd, the creator of Transputer, was based in Bristol, a hot bed of the British IC industry and computer design research for quite a long time. At their peak, they were comparable in every aspect to the cream of the crop of Silicon Valley’s initial 32-bit CPUs, if not more advanced in the system architecture outlook. Even their specific highly parallel programming language, Occam, had its devoted following at the time.

Even though the company is long gone, and the memory of Inmos has all but faded away, there were other interesting Bristol companies that at least partially sprang out of the Inmos roots. ClearSpeed with their controversial PC Floating-Point accelerators for supercomputing apps was famous for its attempt at dedicated – and expensive & proprietary – PC acceleration well ahead of GPGPU time, while Quadrics, an clear offshoot with some key Inmos people setting it up, had a decade long history of providing the world’s most advanced supercomputing interconnects that enabled a kind of shared virtual memory space even among thousands of machines with ultralow latency yet high bandwidth.

Both of these companies went downhill – while ClearSpeed still exists in theory, Quadrics ceased to exist as a going concern exactly a month ago. The Curse of Inmos – or the Bristol high tech is again at work? Knowing the company well, and hoping that ole Blighty stops losing its high tech edge, here’s a story that I hope never repeats again in this business…

After the Inmos downfall, a little Japanese-sounding company Meiko was formed in Bristol, with shared memory SPARCStation interconnect as its first product. It worked well – too well in fact, as Sun didn’t exactly like the idea or efficiently ganging up dozens of cheap workstations instead of coughing up millions for their large servers.

QsNet in 3rd Gen: 64-bit 7-core NPU e.g. PlayStation 3 Cell CPU of networking
QsNet in 3rd Gen: monster powered by a 64-bit 7-core NPU e.g. Cell CPU or a PlayStation 3 of networking interconnects.

The troubled offshoots of Meiko were acquired by Italian conglomerate Finmeccanica, and Quadrics Supercomputer World – a very humble name – was born. Now, you know what might happen when Italians and British work together on something – a lot drama-wise, and not much work-wise. It seems that’s what may have happened to an extent. The impressive QsNet I from the year 1999 Alpha SC cluster times was followed by slightly delayed QsNet II in 2004, and the QsNet III, supposed to appear in 2006 or so, was about to be finally completed this year, when the big bosses in Rome, after several unsuccessful sale or revival attempts, decided to pull the plug. In between QsNet II and QsNet III, the competing, originally inferior, Infiniband went through three generations! There are many reasons why this happened, some not for [even The INQ sting!] press.

A real pity: QsNet specs provided for way more than Ethernet or Infiniband, and more than just much faster, hardware accelerated MPI message passing. Full comms offload by the 64-bit NPU, Cray-style shared virtual memory application level approach across whole cluster while still having independent OS on every node, hardware broadcasting to thousands of nodes from a single node, very good actual usable bandwidth versus theoretical peak, full ECC protection all the way, and reliably running like Energizer bunnies on some of the world’s largest multi-thousand node supercomputers for nearly a decade at a time … the list goes on and on.

Some of the large machines were affected. One of Europe’s largest systems this year, the 3000 node JUROPA at Julich University in Stuttgart, was the first supercomputer supposed to have the superb spec QsNet III, and had to settle for Infiniband instead – no happy shared memories there, virtual or otherwise. Certain UK government departments – in particular, those with expected Big Brother needs, would be somewhat upset too. Imagine, after all, tens of thousands of cheap DP server nodes put together running a large database within a single unified virtual memory space, rather than laborious Infiniband-style passing messages between themselves. Searching on anything about anyone within such database would be oh so fast…

As Intel and AMD start focusing on tighter interconnect control – both QPI and HyperTransport will have external equivalents, one way or another, as you saw from Optical QPI spec as well as HyperTransport over IB/10GE physical layers – there will be less and less space for independent interconnect vendors, and in fact even for custom supercomputing brands like Cray. So, even a buyout would have been too late for Quadrics.

Again, a pity: the UK could have had its matching CPU with Transputer, matching interconnects from those CPUs a la Quadrics, and of course matching Tier 1 vendors like what Apricot, ICL or BBC Micro were at the time. Who controls the CPU controls ultimately everything around it – look at nVidia’s headaches to know it. Or, in recent past, how HP was degraded from a true vendor to a big SI by self-murdering their Alpha and PA-RISC processors. After all, what’s a vendor without their own CPU too? Just a reseller cum assembler…

Morale of the story: if a big government conglomerate acquires a small innovative high-tech co, make sure the team has a strong incentive for performance and delivery, rather than civil servant-type salary lock-in. And, have a clear management and deli
very control system if the places are apart, or else, thing run amok with no one knowing what happens, and the whole business becomes like a skewed distance relationship – you supposedly love someone, but share the bed with someone else. Inmos failed, so did Quadrics and ClearSpeed… let’s see Bristol recover from this and regain its deserved place in the computer high-tech that it had two decades ago.

Can they do that?