Our robot base, Kachaka Pro, comes from Preferred Robotics — the robotics subsidiary of Preferred Networks (PFN), the Japanese deep learning company. A robotics company whose parent group designs its own chips and compilers is a rare thing in the global robotics industry. And the thread that runs from silicon to robot to application is a single goal: cheaper, more dependable robot applications.
What we want to cite here is PFN’s Preferred Computing Platform (PFCP) — described officially as “誰もが MN-Core を利用できる AI クラウドサービス,” an AI cloud service that lets anyone tap into PFN’s in-house MN-Core accelerator series.
What PFN Built
According to the PFCP website and PFN’s October 2024 press release, the stack has three layers.
The chip. MN-Core 2 is a processor co-developed by PFN and Kobe University, optimized for the core of deep learning — matrix operations. Its design philosophy is radical: functions traditionally handled by dedicated hardware circuits — network control, cache control, instruction scheduling — are moved wholesale into compiler software, so that as much of the die area as possible can be packed with compute units. The MN-3 supercomputer, powered by the previous-generation chip, took first place worldwide on the Green500 (the global energy-efficiency supercomputer ranking) three times.
The compiler. Once hardware functions move into software, “how to run an AI model as fast and as efficiently as possible” becomes the compiler’s job: it takes the computation graph defined by high-level frameworks like PyTorch or JAX and automatically generates the optimal instruction schedule and data movement, with no major rewrite of the user’s existing workloads required.
It is worth stating precisely: rather than PFN “using AI to find the optimal inference model,” the more accurate description is that they took the task of “finding the optimal way to execute each model” out of engineers’ hands and handed it to their own software to do automatically. The entity searching for the optimal solution is the compiler; the objects being optimized are inference and training workloads.
The cloud service. PFCP went live in October 2024, built and operated by PFN itself. Running on fully managed Kubernetes, it handles a range of workloads — interactive development, scheduled jobs, distributed training, inference servers — and is currently the only platform that can use the MN-Core series.
The next step goes deeper. According to PFN’s November 2024 announcement, the MN-Core L1000, purpose-built for generative AI inference, is already in development. It stacks 3D DRAM directly on top of the logic circuits to break through the memory bandwidth bottleneck — PFN expects up to 10× the compute speed of conventional processors such as GPUs, with substantially lower power consumption, and plans to bring it to market in 2026 (all of the above are PFN’s own forecast claims).
Why Sigma Cares
Because it is the same thing, just happening on a different floor.
The question PFN answers at the chip and compiler layer is: “For one and the same inference, how do you run it with the least electricity, the least silicon, the shortest time?” The answer is to let software automatically find the optimal way to execute it — not to pile on more general-purpose compute.
Sigma answers the same question at the application layer: “For one and the same patrol inspection, how do you run it with the least cloud cost?” Our answer is an edge-plus-cloud tiered architecture — continuous video stays at the edge, the cloud only interprets the single frames that matter, and once the problem is narrowed enough, a lightweight model is all you need.
In other words: upstream, someone is paving the road of “how cheap inference should be” deeper in silicon and compilers; downstream, we use architecture in the field to push the same curve lower. Every order of magnitude that inference cost falls makes a fresh batch of previously unviable robot applications viable — and that is how we move forward together, working to make robot applications worth far more than they cost.
Source: Preferred Computing Platform website, PFN press release: PFCP launch (2024-10-21), PFN press release: MN-Core L1000 in development (2024-11-15)