Novafora Post-Mortem

Novafora was one of the last of the dwindling chip design/mfg. startup companies in Silicon Valley.

I was on the engineering operations team there for 13 months (Devops/Release/DBA/webmaster.)

I haven’t seen much hard information about the shutdown, so I might as well make some general comments for posterity.

It’s universally true that top mgmt. is ultimately responsible for the success or failure of a company.

In this case, a slipping engineering schedule combined with inadequate funding and VC fixation on Web 2.0 doomed Novafora. Unlike most fragile startups, it took multiple factors.

Let’s be a little more specific though …

Novafora developed 2 projects:

  1. chip code-named “Spika” – a multi-cpu, low-power system that could be used for many things, including multiple stream simultaneous video encoding.
  2. video fingerprinting – a software/R&D project.
  3. received about $35 million in funding

The Spika chip is an approx. 400 million transistor design that incorporated 5 new technology processes, with a fully customized software toolchain and apps. All of this impressive work was done over a period of 3 years with an average of 30 employees (peak of 60). It was taped out twice in H1 2009, but not manufactured.

That’s pretty amazing … so what went right and wrong?

Right:

  • smart engineers – 25% ex-Transmeta, most of the rest from Terayon
  • good teamwork – like 1 big family
  • good infrastructure – IT, SGE grid, build, hardware, commercial Cadence tools, HW emulator
  • pretty frugal – half floor of office space was only $20,000/month, basic meals served evenings, no parties.

Wrong:

  • prolly late by a year overall (took 3 years instead of 2), needed a couple more senior engineering managers
  • 5 new technologies is 4 too many, design experts would say
  • software and apps should likely have started at same time as hardware – a year earlier
  • missing a few senior verification engineers, resulting in 2 tapeouts
  • too ambitious – chip too big for some partners to handle, cannot shuttle wafer
  • lacking laser-focus – chip design changed monthly, as did marketing plan. Effort split between Spika and video fingerprinting projects, some GPU experimentation.
  • only inefficiency I saw was design engineers waiting for 6-8 hour simulation runs instead of continuing work on other facets.

Lasting Impact

  • I doubt the chip got mfg. although it’s possible some samples were made on a “shuttle” wafer.
  • The founders were already wealthy, so the outcome didn’t matter much to them financially. They thought it was wasteful that the company was destroyed while looking for more capital.
  • Half the team leaders bailed a few weeks before the shutdown. Several of the engineers and managers ended up at Apple on their A10 ARM chip team.
  • I bought some of their servers at auction, including cs21. It’s a historic 4-way Opteron 64 GB server that was used to validate chips at both Transmeta and Novafora.
  • Intel bought the Transmeta source control server and IP so that nobody else could get it, and one of the founders bought the Novafora IP. Each paid in the area of $100,000 or $200,000.

Leave a Reply

Your email address will not be published.

This site uses Akismet to reduce spam. Learn how your comment data is processed.