> ## Documentation Index
> Fetch the complete documentation index at: https://docs.thrindex.com/llms.txt
> Use this file to discover all available pages before exploring further.

# Energy & Latency

> Measured synaptic operation counts and modeled energy comparison — THRINDEX SNN vs dense MLP baselines on SHD keyword spotting.

<Warning>
  All energy values on this page are **modeled** from operation counts and published per-operation coefficients. They have not been measured on physical hardware. Silicon-measured power requires the M6 bring-up. See [METHODOLOGY.md](https://github.com/thrindex/thrindex/blob/main/benchmarks/METHODOLOGY.md) for full coefficient provenance.
</Warning>

## Task

SHD keyword spotting — the same model and samples used in the [SHD benchmark](/benchmarks/spiking-heidelberg-digits).

```
Dense(700 → 512) → LIF(τ_mem=20ms) → Dense(512 → 20) → LIF(τ_mem=20ms)
T = 100 timesteps  |  3 bundled SHD test samples  |  thrindex CPU simulator
```

Run date: 2026-07-16. Hardware: Apple M4 Pro (CPU, single sample, single thread).

## Energy results

### SNN via thrindex

| Metric                  | Value                                               |
| ----------------------- | --------------------------------------------------- |
| Avg synaptic operations | 3,664,877                                           |
| Avg firing rate         | 9.9% of dense maximum                               |
| **Modeled energy**      | **1,832 nJ**                                        |
| Coefficient             | 0.5 pJ/syn-op (7nm neuromorphic silicon reference)  |
| Sim wall time           | 14.2 ms (CPU simulator — not the deployment target) |

### Dense MLP baselines

| Baseline                                          | MACs       | GPU energy (2.56 pJ/MAC) | CPU energy (1.0 pJ/MAC) |
| ------------------------------------------------- | ---------- | ------------------------ | ----------------------- |
| A — Rate-collapsed (1 frame, loses temporal info) | 368,640    | 944 nJ                   | 369 nJ                  |
| B — Temporal (T=100 passes, same info as SNN)     | 36,864,000 | 94,372 nJ                | 36,864 nJ               |

### Energy ratios vs SNN

| Comparison                | GPU ratio | CPU ratio | Notes                                        |
| ------------------------- | --------- | --------- | -------------------------------------------- |
| **Temporal MLP vs SNN**   | **51.5×** | **20.1×** | Apples-to-apples — same temporal information |
| Rate-collapsed MLP vs SNN | 0.5×      | 0.2×      | Not apples-to-apples — MLP discards timing   |

**The apples-to-apples comparison is Baseline B (temporal MLP).** Both systems process 100 timesteps of temporal data from the same input. The SNN triggers only 3.7M syn-ops (9.9% firing) vs 36.9M MACs (100% dense) — a 10× reduction in operations at the same information budget.

At 0.5 pJ/syn-op (neuromorphic) vs 2.56 pJ/MAC (GPU), the combined advantage is **51.5×** on modeled energy.

## Why rate-collapsed MLP is not a fair baseline

Baseline A processes the **mean spike rate** across 100 timesteps — a single 700-dimensional vector. The timing information in the spike pattern is discarded. This approach:

* Cannot distinguish two inputs with the same rate but different temporal structure
* Would achieve significantly lower accuracy on SHD than the SNN (SHD specifically rewards temporal processing)
* Uses 100× fewer operations by throwing away 99% of the input data

A rate-collapsed MLP with equivalent accuracy would require a substantially larger architecture.

## Latency note

The SNN simulator wall time (14.2 ms) reflects the **thrindex CPU simulation tool** — a debug and validation artifact, not the deployment target. Neuromorphic silicon operates at hardware speed, typically 10–100× faster than an equivalent CPU simulation.

The MLP wall times (Baseline A: 0.012 ms, Baseline B: 1.18 ms) are measured on the same CPU for a consistent comparison point.

**No latency comparison between neuromorphic hardware and GPU is made here.** That comparison requires M6 silicon bring-up.

## Reproducing

```bash theme={null}
pip install thrindex
python benchmarks/run_benchmark.py
```

Writes machine-readable results to `benchmarks/results/shd.json`. No external dataset download required — the three SHD samples are bundled in the repository.

## Energy formula

```
E_snn = syn_ops × coeff_pJ
      = 3,664,877 × 0.5 pJ/syn-op
      = 1,832 nJ
```

where `syn_ops` = multiply-accumulates actually triggered by spikes (not the full dense count). See [the energy model page](/learn/energy-model) for formula derivation and the [Benchmarks methodology](https://github.com/thrindex/thrindex/blob/main/benchmarks/METHODOLOGY.md) for coefficient provenance.