Research
The Heterogeneous Compute Lab
An AI systems research laboratory studying a central scientific question.
"How does intelligence emerge when computation becomes physically heterogeneous?"
For most of modern computing history, computation could be treated as relatively uniform. Software, compilers, distributed systems, and AI infrastructure were built on the assumption that execution environments were predictable enough to abstract away most hardware differences. That assumption is now breaking.
The modern compute environment is fragmenting across GPUs, TPUs, domain-specific ASICs, photonic processors, neuromorphic systems, analog compute, robotics processors, scientific accelerators, edge devices, sovereign silicon initiatives, and model-specific hardware. In this environment, computation itself becomes an object of study — not models alone, not chips alone, but computation: how it behaves across heterogeneous hardware, how it should be structured across asymmetric environments, how it should adapt under changing conditions, and how hardware and intelligence increasingly shape one another over time.
Computational Physics for Compute Systems
The lab's core research program is Computational Physics for Compute Systems — a research program focused on studying computation under physical and architectural constraints. This is not physics simulation. It asks how computation behaves when hardware is diverse, communication is non-uniform, memory is hierarchical, performance is constrained by physical structure, and execution evolves over time.
Its objects of study include workloads, execution graphs, kernel behavior, hardware states, memory flows, communication flows, runtime policies, and infrastructure telemetry. Primary variables include latency, throughput, energy, utilization, memory pressure, communication cost, placement, contention, runtime instability, and scheduling efficiency. Methods include empirical measurement, telemetry collection, simulation, learned modeling, optimization, control theory, reinforcement learning, and systems prototyping.
The core premise: the structure of computation shapes the space of possible algorithms.
Research Groups
Hardware Intelligence Group
Studies how computation behaves across hardware. Focuses on performance behavior across processors, memory hierarchy dynamics, communication patterns across interconnects, telemetry-driven hardware modeling, learned performance estimation, and learned simulation of hardware environments. Core model: HP1 (Hardware Profiler Model) — profiles how workloads behave across heterogeneous processors, estimating performance surfaces, identifying bottlenecks, characterizing memory and communication effects, and generating hardware behavior profiles from telemetry.
Computational Structure Group
Studies how computation should be organized across heterogeneous environments. Research includes heterogeneous graph partitioning, execution graph synthesis, distributed workload decomposition, hardware-aware scheduling, cross-architecture pipeline design, and multi-stage workflow structuring. Core model: GP1 (Graph Partition Intelligence) — learns how to decompose workloads across heterogeneous systems to improve efficiency and capability. Defines CEP (Computational Execution Protocol) — the representation and protocol layer for execution structure, describing how computation should run across a heterogeneous environment including execution graph structure, placement decisions, data movement requirements, memory tier usage, runtime constraints, and scheduling objectives.
Adaptive Execution Systems Group
Studies systems that adapt execution in real time. Real compute environments are dynamic — workloads vary, hardware conditions change, cluster states shift, and performance targets evolve. This group builds systems that observe, model, and respond to these dynamics: adaptive scheduling, runtime policy learning, execution rerouting under failure, and feedback-driven performance optimization across heterogeneous infrastructure.
Research Philosophy
First-principles inquiry. The lab asks foundational questions about how computation behaves and how it should be structured, rather than focusing only on local optimization.
Scientific rigor. Research programs are defined by explicit questions, empirical measurement, reproducible evaluation, and clear boundaries between hypothesis and demonstrated result.
Systems realization. The lab does not stop at theory. It builds systems that embody its research and expose the theory to real constraints.
Continuous feedback. Operational systems produce traces, telemetry, failures, bottlenecks, and performance data that feed back into research. In this sense, the lab is both a scientific institution and an experimental infrastructure lab.
The lab is currently in stealth. Publications and further details will be shared as the research program matures. For collaboration or research inquiries, reach out at [email protected].