One cache hierarchy
Primary GPU VRAM is the fast scarce tier. Secondary GPUs become warm capacity instead of idle hardware. System RAM is a reported degraded tier, never a silent fallback.
LUCERNA LABS / INDEPENDENT BUILD / GPU INFRASTRUCTUREACTIVE DEVELOPMENT
KEEP THE
FAST GPU
FED.
Aedes turns heterogeneous consumer GPUs into a coordinated memory hierarchy, moving upcoming blocks before the primary GPU asks for them.
LINEAGEINDEPENDENT · NOT ATOM
CORE SIGNALACQUIRE / RELEASE
FRESH MOCK SUITE46 TESTS PASS
SAFETYNO SILENT RAM FALLBACK
INTERACTIVE CACHE SEQUENCESee why release
See why release
changes everything.
Advance through one sequential workload. The primary computes while secondary GPUs stage what comes next; completed blocks become deterministic room for the future.
WORKLOAD-DRIVEN MEMORY MODELregister + load
PLACEMENT EXPLICIT
T0 / PRIMARY GPUFAST · SCARCE · COMPUTE
L0resident
BACKGROUND TRANSFER↑
T1 / SECONDARY GPUsLARGE · WARM · HETEROGENEOUS
L1residentL2residentL3resident
REPORTED LAST RESORT↓
T2 / SYSTEM RAMDEGRADED STATE · NEVER SILENT
EMPTY
STEP 01 / 04
Weights placed
The working set is larger than primary VRAM, so upcoming blocks live on the secondary tier instead of silently spilling into RAM.
HOT-PATH POLICYEXPLICIT SIGNALS → DETERMINISTIC ROUTING
Conceptual proof surface. It models the public API and state machine; it does not allocate browser GPU memory.
WHY IT IS USEFULMemory becomes
Memory becomes
a coordinated system.
Aedes is not a generic GPU scheduler and no ML decides the hot path. It is a placement, transfer, prefetch, custody, and telemetry layer driven by application signals.
Release is the signal
The application says when a block is finished. Released blocks become the first eviction candidates, so sequential workloads do not depend on an LRU guess.
Real background prefetch
Dedicated transfer workers start copying immediately. Work overlaps application compute instead of storing a lazy future that waits until acquire time.
Capability-advertising backends
Mock, CPU, CUDA, and Vulkan backends declare devices, memory, bandwidth, security, and P2P support. Placement routes on what a backend can prove.
Cross-client custody
Blocks carry an owner. Cross-client access is denied unless an explicit revocable capability grants it, keeping shared VRAM from becoming shared authority.
Integrity and reuse
Sensitive blocks are zeroed before reuse; verified blocks are BLAKE3-checked on acquire; shared tensors deduplicate through a content-addressed refcounted index.
OWNERSHIP BOUNDARIESThe Pool
The Pool
decides.
The dependency direction is deliberate: prediction suggests; the Pool authorizes; the transfer engine moves; backends prove their capabilities; telemetry reports the consequences.
POOL
register · acquire · release · state · stats · flush
Tier Manager
Topology, capacity, deterministic placement, released-first eviction.
Prefetch Engine
Returns candidate block IDs. It never owns or calls the Pool.
Transfer Engine
P2P when proven; pooled pinned-host staging when direct transfer is unavailable.
Diagnostics
Tier pressure, block placement, backend health, cache hits, stalls, and spillover.
CUDAVULKANMOCKCPU / DEGRADEDCAPABILITY-ADVERTISING BACKENDS
PUBLISHED BENCHMARK RECORD + FRESH TEST RUNUseful, with the
Useful, with the
proof separated.
The figures below are recorded by the public project benchmark. The fresh verification in this publication pass was the hardware-free Rust suite; real CUDA/Vulkan tests remain a separate ignored hardware gate.
PASS46 mock, integration, additions, and doc tests
OPENReal heterogeneous CUDA-primary + Vulkan-secondary validation
BOUNDARYBlueprint v1.0 is not declared shipped until every hardware and workload gate passes
Aedes shares Lucerna Labs’ preference for explicit signals, typed boundaries, capability advertising, and deterministic mechanisms. It does not depend on the Atom runtime or Spiderweb Bus.