#Benchmark Hardware, Method, And Coverage

These benchmarks are measured on a laptop using the Conway preset and Toroidal topology. The goal is to compare how max speed changes across grid sizes, bit-packing formats, and recording.

#Hardware:

CPU: Intel Core i7-12700H
GPU: NVIDIA RTX 3070 Ti Laptop GPU, 150 W, 8 GB GDDR6 VRAM
WebGPU maxBufferSize: 2 147 483 648 bytes
WebGPU maxStorageBufferBindingSize: 2 147 483 644 bytes
RAM: 64 GB DDR5 SODIMM Corsair 4 800 MHz, 2 × 32 GB
Storage: Samsung SSD 980 PRO 1 TB
Browser: Opera GX 131.0.5877.111, based on Chromium 147.0.7727.56
OS: Windows 10 22H2
GPU selection: dedicated GPU
Power: plugged in

#Method:

The laptop is plugged in and the browser is using the dedicated GPU.
Each sample is configured through the benchmark runner, then measured in max speed mode.
The simulation warms up for $5$ seconds so adaptive batching can settle.
After warm-up, the runner pauses, waits $2$ seconds for the generation counter to settle, reads the starting generation, waits another $3$ seconds, then measures a $t=60$ second max-speed run.
The ending generation is read $2$ seconds after the measured run stops, and reported speed is $\frac{\text{end}-\text{start}}{t}$ . Then, another $28$ seconds are waited to let the device cool down.
Each supported sample is repeated $5$ times, and the best observed run is reported.
Recording runs include OPFS write backpressure in the measured speed.

#Grid and packing coverage:

Grid	Tested bit packings
$128\times128$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$256\times256$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$512\times512$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$1\,024\times1\,024$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$2\,048\times2\,048$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$4\,096\times4\,096$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$8\,192\times8\,192$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$16\,384\times16\,384$	$1$ , $2$ , $4$ , $8$ , $16$ , $32$
$32\,768\times32\,768$	$1$ , $2$ , $4$ , $8$ , $16$
$65\,536\times65\,536$	$1$ , $2$ , $4$
$131\,072\times131\,072$	$1$

Most grids are tested with every bit-packing format. Wider packings make each frame larger, so the largest grids stop at the widest format still allowed by the device frame-size limits. A few maximum-size cases use one fewer row for the same reason: $32\,768\times32\,767$ at $16$ -bit packing, $65\,536\times65\,535$ at $4$ -bit packing, and $131\,072\times131\,071$ at $1$ -bit packing.

Recording is measured only where the engine supports it for that grid and packing.