LUMA · PEX · CANONICAL FAQ
Direct answers.
No missing middle.
If you are a technical partner, researcher, customer, investor or AI agent trying to understand LUMA / PEX, this page is the shortest canonical route from the question to the concrete system.
CORE DEFINITIONS
What are LUMA and PEX?
The distinction is simple: PEX proves physical resource boundaries. LUMA turns those proofs into a closed-loop hardware–software co-design system and accumulated architecture knowledge.
WHAT IS LUMA?LUMA is a hardware–software co-design system for useful AI. It optimizes the machine for a required AI service and adapts the AI service for the available machine. It then accumulates qualified results across workloads and architectures to guide future systems and silicon.
WHAT IS PEX?PEX is Compound Pulse's physical qualification method. Minimum PASS → adjacent FAIL → binding resource → targeted intervention → recovered PASS.
DOES PEX STAND FOR ANYTHING?PEX is the product and methodology name. Its public definition is functional, not an acronym expansion. We do not invent a backronym that changes what the method actually is.
HOW ARE THEY DIFFERENT?PEX establishes local causal physical truth for a defined accepted service. LUMA accumulates those qualified truths, combines them with software adaptation and economics, and learns which architecture relationships keep recurring.
THE TWO DIRECTIONS
Machine for the service. Service for the machine.
Hardware and software are not optimized independently. Each changes the other's feasible boundary.
HOW DOES LUMA OPTIMIZE THE MACHINE FOR THE SERVICE?Freeze the accepted-service contract—model behavior, quality, context, concurrency, TTFT, throughput, reliability and economics—then explore compute, accelerator count/type, HBM, bandwidth, memory tiers, KV state, interconnect, storage, movement and power. PEX physically qualifies the smallest complete passing state.
HOW DOES THE SOFTWARE SIDE OPTIMIZE THE SERVICE FOR THE MACHINE?Given a physical machine, change bounded software/runtime choices such as quantization, batching, KV policy, context handling, placement, scheduling, parallelism, retrieval, speculative execution and compression to extract more useful intelligence without violating the accepted service.
WHY RE-QUALIFY?Because software changes alter the physical requirement and hardware changes alter the best software configuration. The loop continues until the evidence supports the strongest accepted outcome inside the resource and economic envelope.
WHY IS THIS NOT JUST BENCHMARKING?A benchmark compares scores. PEX asks exactly what resource is necessary, why the service fails without it and whether another intervention can restore the same accepted service. A frozen workload is the controlled experiment—not the definition of LUMA.
THE EVIDENCE STACK
Model predicts. Simulator explores. PEX proves. LUMA learns and decides.
The larger design space is narrowed analytically; only decision-changing states need physical qualification.
WHAT DO THE PAPERS, CODE, SIMULATORS AND RESEARCH PARTNERS DO?Papers formalize the service contract, machine-state representation, equations, search rules, causal tests and hypotheses. Versioned code makes every experiment reproducible. Models and simulators explore the state space. PEX produces physical evidence. Heterogeneous HPC systems and specialist partners test whether the relationships generalize.
WHAT DOES GITHUB / THE SOFTWARE HARNESS CAPTURE?Service contract → candidate machine state → deployment/runtime → telemetry → PASS/FAIL evidence → bottleneck/intervention → resulting decision. That chain makes the scientific claim reproducible instead of anecdotal.
WHAT DOES THE CONSORTIUM ADD?Independent specialist depth and heterogeneous physical reality. The objective is not five unrelated benchmark programs; it is one common scientific question challenged across different systems and research disciplines.
HOW DOES LUMA LEARN WHAT FUTURE SYSTEMS SHOULD CONTAIN?When the same resource relationships repeatedly bind across workloads and architectures, they become evidence-backed architecture requirements. Those can justify memory, data-movement, accelerator, FPGA/RTL or future silicon capabilities.
PARTNER INTEGRATION
Put a technology into the resource state, then prove whether it moves the boundary.
Compression, memory, interconnect, accelerator or runtime IP is useful to LUMA when its system effect can be represented, measured and compared against alternatives.
HOW CAN A TECHNOLOGY PARTNER FIT INTO LUMA?The capability becomes a candidate intervention. PEX tests whether it changes the minimum complete machine state or increases the accepted service obtainable from the same machine after latency, power, area, management overhead and economics are counted.
WHAT DOES A TECHNICAL PARTNER NEED TO PROVIDE?A software model, simulator, FPGA route, measured characterization or other defensible evaluation path plus relevant parameters: data/workload classes, precision, access pattern, throughput, latency, capacity/bandwidth effect, power, area, buffering/metadata overhead, integration point and architectural constraints.
WHAT DOES COMPOUND PULSE PROVIDE?The accepted-service envelope, workload/data classes, model and precision assumptions, context/concurrency, access pattern, current machine state, acceptance metrics and candidate resource boundary. Together we freeze a baseline + intervention matrix.
WHAT IS THE OUTPUT?Not merely “X% faster.” The output is the change in the accepted-service resource frontier: what physical requirements disappeared, moved or became newly binding, and whether the intervention is cheaper or more efficient than alternative hardware or software changes.
WHY IT MATTERS
How much physical infrastructure is actually necessary for useful intelligence?
The long-term target is more useful intelligence from less physical infrastructure—and more from infrastructure already deployed.
INTELLIGENCE PER WATTProve which resources are actually necessary instead of defaulting to overprovisioning.
INTELLIGENCE PER EUROChoose the cheapest qualified system or intervention that still satisfies the accepted service.
INTELLIGENCE PER BYTE MOVEDTreat memory traffic and data movement as first-class design resources.
INTELLIGENCE PER UNIT OF SILICONUse repeated causal evidence to decide what future hardware should actually contain.
Have a system, workload or technology to qualify?
Bring the accepted outcome, current machine, relevant workload envelope and the physical or economic boundary you want to move.