Define what must remain acceptable.
Compound Pulse freezes model behavior, quality, context, concurrency, TTFT/throughput, reliability and the physical/economic boundary that matters.
OUTPUT → accepted-service contract + baseline machine stateLUMA · TECHNICAL PARTNER EXECUTION
Bring a technical mechanism.
Prove whether it moves the resource frontier.
LUMA is a hardware–software co-design system for useful AI. A partner capability—compression, memory, interconnect, accelerator IP, runtime control or another bounded mechanism—enters LUMA as a candidate intervention. PEX then tests whether it changes the minimum complete machine required for an accepted AI service or increases the accepted service obtainable from the same machine.
THE PARTNER EVALUATION LOOP
From partner claim to causal system evidence.
The useful question is not “does this technology make one benchmark faster?” It is “does this mechanism move the accepted-service resource frontier after its own latency, power, area, management overhead and economics are counted?”
Turn the partner technology into a measurable intervention.
The partner provides a software model, simulator, FPGA path, measured characterization or another defensible representation of the mechanism and its constraints.
OUTPUT → parameterized intervention modelFind minimum PASS and adjacent FAIL.
PEX establishes the smallest complete passing state, deliberately crosses the nearest resource boundary and identifies the actual binding resource.
OUTPUT → baseline causal resource boundarySubstitute or change the suspected resource.
The partner mechanism is introduced at its real architectural integration point—memory path, runtime, interconnect, accelerator or other bounded layer.
OUTPUT → intervention state + measured overheadsMeasure the frontier delta.
PEX repeats the accepted-service test and measures what changed: capacity, bandwidth, latency, accelerator count, concurrency, power, cost or the next newly binding resource.
OUTPUT → baseline vs intervention Resource Boundary ReportOne win is evidence. Recurrence is architecture knowledge.
LUMA repeats qualified results across workload classes and architectures. Only recurring, economically meaningful wins become candidates for deeper design-in or future silicon.
OUTPUT → retain / reject / expand / design-in decisionPARTNER INPUT PACKAGE
What a technical partner provides.
The smallest useful package is enough to represent the mechanism honestly inside a system model and then test it physically or against measured characterization.
01EVALUATION ROUTE
Software model, simulator, FPGA/emulation path, measured characterization, API or other defensible route to evaluate the capability.
02SUPPORTED DATA / WORKLOAD
Relevant data classes, precision/formats, access patterns, concurrency behavior and any workload-specific assumptions.
03PERFORMANCE EFFECT
Capacity, compression, bandwidth, throughput or other primary mechanism effect as a function of the workload—not only one headline benchmark.
04PHYSICAL COST
Added latency, power, silicon area, buffering, metadata, memory-management overhead and any throughput ceilings.
05INTEGRATION POINT
Where the mechanism sits in the real system—runtime, HBM path, memory controller, CXL tier, fabric, accelerator/SoC or another layer—and the constraints that creates.
06BOUNDARIES
Software transparency, coherence, supported platforms, toolchain maturity, evaluation/licensing limits and conditions needed for a valid result.
COMPOUND PULSE INPUT PACKAGE
What we freeze on our side.
Compound Pulse controls the experimental contract so the partner technology is compared against a stable accepted service instead of an arbitrary benchmark.
01WORKLOAD ENVELOPE
Model/model class, data or tensor classes, precision, context, concurrency and relevant access pattern.
02ACCEPTANCE
Quality, TTFT, throughput/TPOT, accepted goodput, reliability and any non-negotiable service constraints.
03CURRENT MACHINE
Compute, memory hierarchy, interconnect, runtime, power/facility and the current complete operating state.
04BOUNDARY TO MOVE
The purchasing, operating or architecture decision that matters: remove hardware, increase service, reduce energy, change memory tier or avoid future provision.
05BASELINE MATRIX
Candidate states around minimum PASS and adjacent FAIL, including the telemetry needed to identify the binding resource.
06INTERVENTION MATRIX
The same service tested with the partner mechanism and the same acceptance rule, so the resulting frontier delta is comparable.
WHAT COUNTS AS A RESULT
Measure the change in physical requirements, not just the benchmark percentage.
A strong partner result explains what hardware requirement disappeared, moved or became newly binding and whether the intervention beats the realistic alternatives.
Resource delta
HBM/host/CXL requirement, bandwidth, accelerator count, memory traffic, interconnect or another complete-machine requirement.
Service delta
Accepted concurrency, context, throughput, latency or another useful service dimension obtainable from the same machine.
Physical overhead
Latency, power, area, buffers, metadata and operational complexity introduced by the mechanism.
Economic delta
Cost/accepted output, energy/accepted output, deferred provision or lifetime value versus adding conventional hardware.
Next binding resource
What becomes limiting after the mechanism succeeds; removing one bottleneck is useful only if the new system state is understood.
Decision
Reject, retain, expand across workloads, integrate into control, or open a deeper hardware/IP design-in path.
RESEARCH + CONSORTIUM EXECUTION
Long-horizon research programmes plug into the same interface.
A 36-month consortium, bilateral research collaboration or HPC campaign can use the same evidence chain, but the programme duration is not the definition of LUMA. Work packages map into bounded inputs, owned technical artifacts, acceptance metrics and downstream handoffs around the same accepted-service/resource question.
Papers / formal work
Define the service contract, state space, equations, causality rules and falsifiable hypotheses.
Code / GitHub
Make each experiment executable and reproducible from service contract through telemetry and decision.
HPC / physical platforms
Test decision-changing states on heterogeneous physical systems rather than assuming one vendor's behavior generalizes.
Specialist partners
Own one bounded mechanism or scientific question and return the artifact/evidence needed by the next block.
CONTROLLED TECHNICAL ACCESS
Public context first. Role-scoped disclosure second.
Keep the first exchange non-confidential. Exact traces, checkpoints, private source, test vectors and customer-confidential material move only after the appropriate confidentiality/cooperation path is in place.
Public first
Capability summary, integration point, evaluation route and non-confidential system parameters.
Controlled second
Role-specific workload traces, vectors, source modules and interface records only when needed for the agreed evaluation.
Evidence return
Artifact identity, build/tool provenance, measured FOMs, qualification result and the decision the evidence supports.