ASIC startup: hardware is hard (Google CEO Pichai)

Take it from Sundar Pichai, CEO of Google. Hardware is hard. And Google is a software giant that designs its own chips. Other trillion dollar companies, Amazon, Apple and Microsoft have hardware design teams too. Even Facebook has ASIC designers on the payroll in 2020. General purpose processing started the boom in hardware in the 20th century. Then multi-purpose chips served multiple application domains. A GPU takes over the graphics processing, off-loading the CPU. But in the AI market, the GPU speeds up training and inference. Which brings us to the age of Application Specific IC’s or ASIC’s. Hardware is the foundation that determines the speed and efficiency of any software algorithm. And this spawns HW startups with their own novel niche solution. But startups are hard. Most startups fail. The top-3 fails for startups are:
  • Market match (customers must want to spend money on it).
  • Cash flow (accounting): running out of cash.
  • Not finding the right expertise.
And hardware startups are harder than any other startup since they require significant capital investment. An ASIC is expensive to develop and to manufacture. It hides most of the complexity from the public eye. We only hear about the successful applications, not about the process. This is fine, it is intentional. But if you consider a startup, you need to open up pandora’s box. And we need to go back in time as well…
In the nineties (20th century), ASIC designers could do all front-end tasks. A graduate started as a verification engineer. To learn more about design and verification. With experience, they mastered other front-end design tasks. Synthesis, Static Timing Analysis (STA) and Design For Test (DFT). For the ASIC itself and for the emulator platform. The latter still exists but is very expensive. Startups use FPGA’s for prototyping an ASIC. And engineers sat down with the layout team to solve Clock Tree Synthesis (CTS), placement and routing congestion problems. And in the bring-up phase, silicon validation with embedded software (firmware) and application software required attention too. But few remained on the level of doing the work. Most climbed the ladder or changed their career to other fields. Companies realized it was easy for resources to switch companies. There goes your investment in your human resource. Basically, the path to a senior front-end designer role is long and expensive. Another factor that creeped up on us: the chips became more complex. In the 20th century, it involved a few people to design a chip. Let’s say maximum ten. Then it grew to 20 to 40 people. Today, we have 150 people and more working on a complex System-On-Chip (SoC) in the latest technology node. So, the front-end today has different sub-domains of expertise. They tuned ASIC down from all-in-one engineers to single skilled commodity. The reverse evolution of the ink-jet printers we all know so well. As with all commodity jobs, the new normal is that people are designers or verification people or silicon validation or synthesis and STA people or… It’s “OR”, not “AND”. Something digital designers recognise as boolean logic. Hence, today front-end experts are a rare find. Even if they could recognise all-in-one engineers, recruiters and agencies don’t like multi-skilled engineers.  because they are too expensive versus single skilled engineers. Big companies, prefer single skilled people (cheaper). They add up nicely in an excel sheet. And management today is all about reducing cost in spreadsheets. There are enough international semiconductor companies that stay afloat for decades because, like all big organizations, the efficiency is so low; it doesn’t make a big difference. But for startups in the semiconductor world, they need to be fast to market. And they need multi-skilled people to make it happen. So, when a hardware startup needs to choose between expensive EDA licenses or an expert IC design engineer, they should choose the latter. The few unicorns with decades of experience in ASIC design are invaluable to any HW startup. They understand the consequences of bad choices at the design phase. They ripple through verification, synthesis, STA and DFT. Essentially, they know what to do to avoid problems before they occur. And, big ASIC projects today cost so much money, it is crazy. Sure, there is a technology factor. However, they spend it on servers and EDA licenses instead of on smart people that save big on server/license cost. Project management detects issues too far in the pipeline. That creates feedback loops back to design and verification. They find issues weeks after they verified the design. Resources need to get up to speed again. HW startups with a successful chip that is selling on the market in about 2 years and that outperforms the competition all have silicon veterans in their top management. And that is no coincidence. You need experts and veterans to use commodity resources efficiently.
What makes a unicorn engineer valuable for a HW startup?
  • Simple rules for coding to avoid issues and to automate tasks further in the pipeline.
  • Running tools early and often with self-checking regressions.
  • Using methodology and tools proven by our friends from the open source software realm.
  • Automation of tasks, especially manual error-prone and tedious tasks. Including documentation, even better self-documenting code.
  • Structuring of the project and efficient communication flow. Lean meetings, avoid inefficient time consuming meetings to schedule meetings.
Having decades of experience has an enormous advantage. But not all with decades of experience have this advantage. Look for the people with a talent stack.
My stack includes a bit of recreational writing. Because I want to share my experience, especially things that were common knowledge but got lost in the sands of time. I have always observed to do things faster and better. To avoid problems, we had to deal with. Most just do lip service to quality and improvement. They don’t really care. I care deeply. Even though it requires unapologetic truths. Speak truth to power for the sake of efficiency. Even if it goes against the wallets of the powerful.
“A nation of sheep begets a government of wolves (E. Murrow)”.
This is no different in the semiconductor world.

Comments (0)

No comments yet.

Leave a comment