Ubitium - ubiquitous compute

aka. a compute-abundant future and a $3.7M seed round to kick it off

If you have ever been stuck in an elevator for more than an hour, you probably had time to think about what goes on behind the metal plates. “Why did it stop!?” you might be asking yourself, stepwise going through your actions before it stopped, then going through the imagined machinery that normally makes the elevator go to the right floor. Odds are, if you’re reading this, you got out at some point. Maybe like me, with the help of a whole team of fire fighters and elevator service technicians, or maybe, with the help of the edge-AI and IoT if you were in an elevator conceptualized and built by Hyun Shin Cho.

Hyun Shin spent a good portion of his working life not just in the elevator business, but in building compute platforms for making smart buildings and cities a reality by connecting infrastructure with elevators and escalators while at ThyssenKrupp Elevators. But his first job was actually interning for Martin Vorbach - one of the main contributors to FPGA technology. Martin and Peter Weber, a veteran in semiconductors who once led Intel’s market expansion into Europe, had started a company together building on the FPGA IP developed by Martin. 20 odd years later the three of them are re-united in a singular mission, to enable a future with compute everywhere.

Why is my toaster so dumb?

Or, what makes the production and integration of advanced compute so hard?

Credit to Lumafield for scans of everyday things

The current compute industry relies heavily on heterogeneous architectures that combine multiple specialized processors. If you break open an Airpod, it contains highly specialized circuitry. Just from first principles, it likely has the following internal components:

1. Bluetooth chip: The Apple W1 or H2 chip

2. Audio codec: A Maxim audio codec for audio processing

3. Accelerometers: from manufacturers like Bosch and STMicroelectronics for motion sensing and gesture controls

4. Power management: Low-dropout regulators (LDOs) e.g., from STMicroelectronics 

5. Microphones: MEMS microphones are used for voice pickup and noise cancellation

6. Battery: A tiny lithium-ion battery (around 93 milliwatt capacity) per earbud

7. Optical sensors: Light sensors

8. Antennas: For connectivity

9. Speaker driver: A dynamic speaker unit (around 11mm diameter) produces audio output

This becomes quite a complex development, sourcing and integration cycle even if you have a massive market and a chokehold on the world's semiconductor industry. Luckily airpods are a huge success for Apple. But the core challenges are the same for any toaster/headphone/car/lamp/etc. company:

• Integration Complexity: Combining different processors requires complex hardware design and software development efforts. Each processor type has its own architecture, programming model, and toolchains

• Hardware Costs: Multiple processors increase the Bill of Materials (BOM), manufacturing and supply chain complexity, as well as power consumption. This is especially problematic for edge devices, which often have stringent cost and energy budgets

• Limited Scalability: As applications evolve, upgrading systems with specialized processors can be inflexible. Adapting to new workloads may require significant redesigns or completely new hardware

The above reasons are part of why the IoT revolution didn’t pan out as McKinsey forecasted in 2015, why we don’t have smart cities yet, and why firefighters had to come break open the elevator for me.

Hyun Shin had experienced the reasons for this himself, as he designed edge-AI enabled product platforms using existing silicon. By investing heavily in compute everywhere, he thought he could compensate for the underlying issue of the hardware heterogeneity. Each product was unique, with massive discrepancies in standards and composition. Uniquely tailored PCBs with specialized chips, which require specialized hardware teams. There was enough compute, but integration was slow, connectivity was difficult and most of all, it was too expensive. Both in the specialized teams that needed to be built up to deal with firmware and codecs, but also in the actual BOM cost. The business case just doesn’t make sense for most products today.

Rethinking the basics of computing

While Hyun Shin was innovating in the smart building space, Martin was brewing new ideas around how to remake the CPU. He was going as far back in his re-design as to the 1960’s when an IBM engineer named Robert Tomasulo came up with an algorithm for processing data in parallel. He needed to remove internal dependencies in the processing to speed up execution so that there was less idle time. He did this by introducing register renaming (using placeholder values when the real data isn’t available), reservation stations (all the data needed to perform one operation) and a common data bus (a central communication channel for connecting functional units). This worked great for many decades, in fact it, or variants of Tomasulo’s Algorithm, are the standard since the 90s in how the x86 and x64 architectures do so-called out-of-order execution in the CPU.

In the 20 years of designing coarse grain reconfigurable arrays (CGRAs), Martin had come up with a different way of doing exactly that, out-of-order execution, but better. Not everything in modern computing is out-of-order execution of course. There is a lot of data processing which should ideally be done as a dataflow loop, for example, most AI workloads. Luckily, the architecture Martin had in mind was a dual operation mode that would combine the flexibility of an FPGA with the simple, standardized programming mode of a CPU. This meant that not only could it replace the CPU, it could also remove a couple of the other components necessary for advanced capabilities!

Lowering TCO and scaling compute

Zooming back out, Ubitium is building for a future where we have advanced compute capacity everywhere. Software is pushing towards the same direction, with performant quantized, local models, toolchain abstractions, serverless and distributed compute orchestration. But hardware is growing more and more fragmented, with specialization as a second lever to improve performance and keep Moore’s Law alive as we’re approaching the limits of process innovation. While others are developing chiplets, 2D materials and more performant ways to do lithography with better mask coatings, etc., we can still do architectural innovation.

In Ubitium’s case, it’s not innovation for the sake of have a different type of chip, but rather a remake of the development and deployment of compute. Their design can be a drop-in replacement for a whole host of other, specialized chips. Most importantly, this lowers the total cost of ownership (TCO), from the sourcing, through software development, to over-the-air-updates. In building on top of the established RISC-V ISA, they can bootstrap from an ecosystem of software, while remaining independent of any vendor lock-in.

Ubitium GmbH

Inflection’s investment thesis

In summary, the current situation is:

• edge compute is 100-1000x more fragmented and heterogenous in components, use cases and development frameworks than the data center compute market

• this leads to prohibitively high cost and complexity with the device manufacturers in the supply chain, development cycles and maintenance

• this in turn, leads to less functionality and fewer products with advanced capabilities, longer time to market, more specialized development teams and less flexibility on the software side

With a re-programmable, universal processor built on RISC-V, we can replace a large number of the specialized components for the same end-device performance or better, at less cost and complexity. Moving the hardware configuration into software also keeps the light-cone of what the end-user should experience wide for longer. I.e., design decisions can happen later in the development cycle, with less lock-in!

We could not imagine a better team for the job, ranging technical excellence, industry experience, a vision to remake a core component of modern computer architecture, and a deep understanding for the customer. Martin started his first semiconductor company during his studies, and invented much of the modern FPGA, holding over 200 patents in the space. Hyun Shin served as a UN peacekeeper after studying in Korea, Germany and US, leading Global Manufacturing controlling and then digital transformation at ThyssenKrupp Elevator. Peter (chairman) was one of the first Intel employees in Europe and spent a whole career in global semiconductors.

They’ve already expanded the team with stellar first collaborators but still have a few vacancies.

Thanks to the team for trusting us to partner with you on the journey, and to our co-investors Dmitry and Francesco from Runa Capital and Rudi of KBC Focus Fund. We look forward to a compute-abundant future!