SAOPS: the spiking net's raw synaptic-activation throughput
SAOPS is Synaptic Activations per second, the forward-pass compute of the binary spiking neural net behind Simargl. An evolved genome (5 type-vectors × 512 bits) is driven forward over a batch of encoded inputs, on the native XNOR + VPOPCNTQ kernel, fanned across every core. No floating point, no GPU. Measured live, end to end.
frontier inference
A GPU CLUSTER. Racks of accelerators, megawatts, floating-point matmul, and a metered API for every forward pass.
semurg, CPU-only
Trillions of synaptic ops/s. Binary spiking net, minus-one / plus-one synapses, XNOR + popcount, on the CPU you already own.
Target: 15 trillion synaptic activations per second, no GPU, no cloud.
Native VPOPCNTQ SNN forward over an evolved genome. Measured live on this box, across all cores.
driving the SNN forward at max concurrency, the live SAOPS render here…
One forward pass evaluates every synapse as a binary XNOR against the input activation,
summed with a single popcount, no floating point. SAOPS is that work fanned across every
core and timed end to end. If the native SIMD kernel is not engaged this runs on the pure
Elixir replica and we say so plainly: the number is real, but the 15T target needs the
native VPOPCNTQ SNN-forward path. This is the raw compute Simargl runs on.