A64 × MXU
HBM 8.0 TB/s · Scale-up mesh
Evidence-driven hardware architecture exploration
Design the hardware before it exists.
Blueprinting turns candidate compute, memory, interconnect, and system organizations into inspectable, simulatable, and comparable architecture blueprints—grounded in workload truth, versioned evidence, and explicit claim boundaries.
- Shared truth
- One workload semantics for every candidate
- Late binding
- GPU, LPU, and future targets stay comparable
- Traceable
- Every claim leads back to evidence and assumptions
EXPERIMENT / BP-042
REPRODUCIBLE
Architecture question
Which memory and fabric choice actually moves the bottleneck?
B96 × MXU
HBM 5.2 TB/s · Ring fabric
01 / Exploration loop
From an architecture question to evidence that can carry a decision.
Not parameters poured into a formula—every comparison shares semantics, constraints, evidence, and verification boundaries.
-
01
Frame
Fix the workload suite, objectives, constraints, and uncertainty.
-
02
Blueprint
Describe compute, memory, fabric, and system candidates.
-
03
Map
Derive legal parallelism, placement, schedules, and buffer plans.
-
04
Simulate
Resolve analytical, network, hardware, and measured evidence as needed.
-
05
Decide
Compare bottlenecks, sensitivity, uncertainty, and Pareto frontiers.
02 / System capabilities
Organized around hardware decisions—not one simulator.
Early models may be coarse; semantics and provenance may not. Increase fidelity as the design matures without rewriting upstream facts.
Candidate blueprints
Represent engines, memory hierarchy, NoC, package, node, and cluster organization.
Explore the design space →Workload and mapping
Conserve operations, bytes, messages, and dependencies while comparing sharding, recomputation, and scheduling.
Understand the workload model →Multi-fidelity simulation
Unify analytical estimates, network simulation, hardware simulation, and future runtime replay.
See the simulation architecture →Evidence and calibration
Tie every cost claim to a versioned provider, applicability domain, confidence, and observations.
Read the evidence model →Explainable decisions
Produce feasibility, bottleneck, utilization, sensitivity, uncertainty, and Pareto comparisons.
Follow the exploration workflow →Agent-ready analysis surface
Stable identity, structured diagnostics, and reproducible experiments let agents propose, run, and audit explorations.
Inspect module boundaries →03 / Formal spine
One workload truth, progressively resolving new design decisions.
Blueprinting borrows typed IRs, staged lowering, and verifiers to serve architecture modeling, derivation, and validation—not to redefine the product as a compiler.
Enter the formal analysis foundations →- 01ModelIRmodel semantics and dataflowlive
- 02DistributedTaskIRlogical mesh, shards, and collectiveslive
- 03PortablePlanIRexact work and target-neutral planlive
- 04ConcretePlanIRarchitecture binding, resources, and schedulecontract
- 05MachineIRoptional target program realizationcontract
04 / Evidence ladder
As the design sharpens, evidence gets richer—and old conclusions remain reproducible.
Analytical model
Roofline, capacity, and explainable envelopes for early pruning.
Network simulation
Topology, routes, collectives, contention, and scale-out effects.
Hardware simulation
Queues, resource occupancy, memory lifetimes, and cycle-level detail.
Measurement and silicon
Profiler, prototype, GPU, LPU, and silicon observations create new revisions.
05 / Honest status
The target architecture and implemented capability are stated separately.
This site treats Implemented, Experimental Contract, and Planned as auditable states—not marketing language.
Read the capability matrix →Transformer → PortablePlanIR
Typed workload accounting, distributed mapping, checkpoints, analytical estimates, and Calculon calibration.
Two-blueprint exploration
Compare two parameterized virtual architectures with resource simulation, bottleneck, and utilization results.
06 / Reading paths
Start with the decision you are making.
01
Hardware architects
Begin with the product thesis, design space, and exploration loop.
Start here → 02Modeling engineers
Enter the hardware model, workload mapping, evidence, and simulator protocols.
Inspect the models → 03Systems and compiler teams
Understand formal representations, derivations, verifiers, and target-binding contracts.
Enter the foundations →BLUEPRINT / SIMULATE / EXPLAIN