The engineering side of the AI build-out, measured.
Spliced is a small research lab. We study what it costs to run a model, in watts and joules as well as seconds, from one laptop up to how inference clusters are designed.
Every result ships with its method, its code and its raw data, and every note says what it cannot show.
Inference engineering
Where decode time and energy go: memory bandwidth, kernels, quantization formats, batching, the KV cache, serving at cluster scale.
A2Data science
Measurement you can trust: repeated runs, held-out validation, and models whose coefficients have to make physical sense.
A3Software engineering
The tools that make the measurements repeatable, open source, small enough to read in one sitting.
what full precision really costs over 4-bit at the board. Measured at the chip alone, the two look the same: the memory energy is invisible there.
less energy per token from serving eight requests at once. Batch 2 saves 1%: the GPU reaches its power ceiling at the second sequence.
of the measured 266 GB/s memory bandwidth. Decode is memory bound, and each step down in bits uses less of the roof.
Published research.
Where the joules go: a roofline and an energy model for decode
A decode roofline from measured bandwidth, and a validated energy model whose residuals isolate the cost of unpacking each format. Two wrong first attempts, kept in.
Read note →N1 · 24 October 2026Fewer bits is not fewer joules
Eight quantization formats and batch sizes 1 to 8, energy at the chip and at the board. Chip-only measurement misranks F16 by 2x; batch 2 saves 1% per token and batch 8 saves 72%.
Read note →On the bench now.
- A 7B model and MLX. The format study on a model five times larger, and on Apple's own runtime.
- A second machine. The same study on a Windows laptop, to see which findings are Apple Silicon and which are physics.
- Per-kernel energy. Each format's matrix-vector kernel timed and power-sampled alone, to explain what the energy model cannot.
- Cluster design. Tokens per megawatt-hour across GPU generations, from public specs and a roofline, checked against published benchmarks.
- Speculative decoding. Does a draft model save joules, or only seconds?