Calxeda

» see calxeda.com as it lived, on the wayback machine

~ the obituary ~

The board had just been strengthened. Frank Frankovsky, who ran hardware supply at Facebook, had just joined it, which in December 2013 was roughly the strongest signal a server startup could send. Then on the afternoon of 19 December one of the company's own executives told The Register, in plain words, that Calxeda had simply run out of money. Employees were told the business was closed from that day. It was six days before Christmas and the website was already unreachable by the time the story went up.

The statement the company put out reads better than most: carrying the load of industry pioneer, it said, had exceeded its ability to continue operating as it had envisioned, and it remained committed to customers with ECX-2000 projects underway. Calxeda had been founded in Austin in 2008 under the name Smooth-Stone, by Barry Evans and Larry Wikelius among others, on a proposition that sounded eccentric at the time and obvious later: that the processors in phones had the right power profile for the kind of server work that is mostly waiting on a network.

The hardware was genuine. In March 2011 it described a 480-core server built from 120 quad-core Cortex-A9 chips. That November it shipped the EnergyCore ECX-1000: four cores at 1.1 to 1.4 GHz behind 4MB of shared L2, drawing one and a half watts a processor. A whole server node with four gigabytes of DDR3 came in around five watts, half a watt at idle, and each die carried five 10-gigabit Ethernet ports, four chips to an EnergyCard. Hewlett-Packard built it into Project Moonshot. Boston Limited in the UK sold the parts as Viridis appliances and demonstrated a five-watt-per-node ARM cluster at Supercomputing 2012. The Register counted ninety million dollars of funding on the day it died; later tallies put it at a hundred and three, including a fifty five million dollar round in October 2012, across roughly a hundred and thirty employees.

~ the case file ~

~ thirty two bits ~

Both shipping generations were 32-bit. The ECX-1000 was a Cortex-A9 part and the Midway ECX-2000 that followed was Cortex-A15 with 40-bit extended addressing bolted on, pitched openly as a bridge to the 64-bit chips that were coming. A bridge is a hard thing to sell to a datacenter buyer, because the memory ceiling is the first number anyone checks and the answer on an A9 node is a few gigabytes.

Against a Xeon that could address hundreds, the efficiency argument only worked for a narrow slice of web-tier workloads, and even there the software was not ready. The 64-bit ARMv8 part was scheduled for 2014. Analysts afterward said the same thing in different words, which was that the company had been too early for ARM servers, and that its first processors were not powerful enough even for the jobs they suited best.

The money ran out between the generation that could not win and the generation that might have. Karl Freund, speaking for the company that week, said the remaining staff would keep supporting the ECX-2000 deployments already in the field.

~ a week before christmas ~

The intellectual property surfaced again in December 2014 inside a company called Silver Lining Systems, which is the ordinary afterlife for silicon. The patents and the designs find a buyer even when the business does not. calxeda.com is a GoDaddy sale page now.

The argument won anyway. The big cloud vendors design their own ARM server parts today, and the boards on my desk in Ankara are ARM without anyone having to defend the choice. Calxeda had the architecture right and the calendar wrong, and from inside a funding round those two mistakes look identical.

@}-->--

              )
             (_)
            .-'-.
            |   |
            |   |
            |   |
            |   |
          __|   |__   .-.
       .-'  |   |  `-:   :
      :     `---'     :-'
       `-._       _.-'
           '""""""

« previous  ·  random  ·  next »