THE NOETON STACK · L2 / L5
L2 Tensor Dataflow Fabric
TENSOR DATAFLOW FABRIC
Dataflow does the heavy lifting; the CPU retires to housekeeping.
Everything clever about a general-purpose CPU — out-of-order execution, branch prediction, cache hierarchies — exists to tame unpredictable control flow. Transformer inference has almost none: it is a statically known graph of tensor operations, where every read, compute, and write can be scheduled at compile time. Running a deterministic workload on a speculation machine is using the wrong tool to do the right thing.
L2 hands the main computation to a deterministic dataflow fabric: operators map directly onto hardware pipelines (Groq TSP and SambaNova RDU have proven the route), and tensors flow between units on a compiled beat — no cache misses, no branch penalties. The classical CPU is still present, but only for initialization, monitoring, and error handling. It no longer sits in the middle of the datapath.
POSITION IN STACK
L5 Intent Interface L4 Network Fabric · Tiered Cognition L3 Model-OS Runtime L2 Tensor Dataflow Fabric ◀ you are here L1 Memory-Centric SubstrateTensor pipeline array
Matrix multiply and elementwise ops mapped to static pipelines
Attention engine
A datapath purpose-built for attention's access pattern
Deterministic interconnect
Communication scheduled at compile time, zero arbitration
Housekeeping core
The demoted general-purpose core: init, monitoring, exceptions
INTERACTIVE · LIVE
Determinism speaks through tail latency: same average speed — watch p99 collapse under dynamic scheduling.
Deterministic execution
Latency is predictable at compile time — real-time goes from "best effort" to "contract term".
Operators are the hardware
The bottom of the software stack is no longer an ISA but a tensor-operator library.
Control split from compute
CPU and fabric each mind their own world without blocking the other.