BENCHMARKS

Less memory. More work.

Memory under load, processing speed at equal Page concurrency and network acquisition with Optimize. Controlled workloads with explicit correctness checks.

October 10, 2026 · Chrome 152 · Windows 11 · i7-14700KF · 32 GiB RAMMethodology
ACTIVE MEMORY5.64×

Less memory under load.

727.78 vs. 4,102.04 MiB process-tree RSS with 50 active static Pages.

PROCESSING SPEED7.45×

More Pages per second.

164.62 vs. 22.11 Pages/s at the same 50-Page concurrency, including create and teardown.

OPTIMIZE ACQUISITION98.1%

Acquire fewer bytes.

284,591 → 5,276 HTTP body bytes for Books extraction with Optimize on versus off. Separate workload; 5/5 passes.

SAME NUMBER OF PAGES

50 Pages. Compare the work.

250/250 measured static batch attempts passed for each runtime. These runs are separate observations, with a small instrumentation-boundary difference recorded below.

Active memory

Lower is better
Mimic727.78 MiB
Chrome 1524,102.04 MiB

Batch processing

Higher is better
Mimic164.62 Pages/s
Chrome 15222.11 Pages/s

CPU cost: 69.94 vs. 133.38 ms per Page. Mimic batch latency p50/p95: 161.33 / 197.39 ms; Chrome: 1,853.23 / 2,382.80 ms. Batch speed does not imply faster single-Page completion.

Concurrency, memory and CPU

Measured active RSS, batch throughput, CPU per Page and RSS 250 ms after teardown. A failed series is excluded from successful comparisons. An absent Chrome value means no successful matching reference is available.

WorkloadPagesMimic / Chrome RSS MiBMimic / Chrome Pages/sMimic / Chrome CPU ms/PageMimic recovery RSS MiBValid / attempts
static1106.44 / 1210.8116.87 / 8.6338.28 / 222.6688.0520 / 20
static5148.94 / 1373.7480.55 / 22.1551.88 / 130.6385.4825 / 25
static10213.13 / 1682.5379.79 / 24.1961.88 / 126.56101.8950 / 50
static25418.23 / 2571.97164.27 / 25.0359.13 / 123.25117.70125 / 125
static50727.78 / 4102.04164.62 / 22.1169.94 / 133.38133.59250 / 250
static1001351.21 / —170.15 / 21.6775.63 / 136.41169.20500 / 500
cpu1124.77 / —9.19 / 8.6681.25 / 234.3896.7720 / 20
cpu5206.46 / —43.22 / 20.8987.50 / 161.8892.1625 / 25
cpu10316.48 / —74.82 / 26.8597.19 / 170.31112.4450 / 50
cpu25671.74 / —95.82 / 29.18142.88 / 159.50122.14125 / 125
cpu501225.69 / —111.25 / 29.50147.38 / 154.50141.21250 / 250
cpu1002333.55 / —108.20 / 19.14157.28 / 194.75184.64500 / 500
react1119.95 / —9.22 / 6.87103.91 / 319.5397.6920 / 20
react5180.53 / —43.64 / 5.89121.25 / 230.6394.6325 / 25
react10288.28 / —56.41 / 8.65143.13 / 186.88123.6150 / 50
react25569.34 / —75.96 / 24.76174.50 / 157.25159.75125 / 125
react501013.87 / —83.26 / 17.08180.81 / 194.63208.83250 / 250
react100— / —— / 16.57— / 212.44—400 / 500 · failed

RSS sums process working sets and can count shared pages more than once. Recovery is a short observation, not proof of leak absence or an absolute capacity limit.

Single-Page latency and CPU

20 measured warm attempts per workload and runtime, one warmup excluded. These tables include workloads where Mimic completes more slowly. CPU covers Page creation through teardown; completion spans navigation through the exact workload result.

WorkloadMimic p50 / p95 msChrome p50 / p95 msMimic / Chrome CPU ms/Page
static32.05 / 35.0218.44 / 22.0731.25 / 140.63
cpu59.47 / 65.8345.59 / 52.1462.50 / 234.38
dom155.73 / 163.3048.28 / 66.57187.50 / 210.94
async71.39 / 82.7450.98 / 55.1585.94 / 203.13
react66.33 / 69.2344.95 / 51.9093.75 / 250.00
wasm30.96 / 37.6123.13 / 30.5946.88 / 179.69

Optimize on versus off

Books SSR extraction: 98.1% fewer acquired HTTP body bytes, 284,591 → 5,276; transport responses 29 → 1. Both Mimic Default and installed Auto passed 5/5 matched local trials. Manual tied Auto. Training and validation took 27.1 seconds.

Separate recorded workload from October 5. Encoded acquired bodies are not physical wire traffic. One captured state does not establish savings on other sites or future states.

Optimize results and limits

Measurement boundaries and failures

The final Mimic executable passed 120/120 warm single-Page attempts. Static and CPU completed 100 Pages; React completed 50 Pages. React-100 failed in its final wave with a native access violation and is excluded from successful comparisons: 2,810/2,910 density attempts valid.

Active-memory comparisons reuse the frozen Chrome reference; timing/CPU use the saved October 10 Chrome capture. Both runners use the same fixture, frozen execute, navigation barrier and 50 ms completion polling. Candidate wave timing includes the initial accounting snapshot and excludes the final one; reference timing does the reverse. Guards differ (120 vs. 180 seconds) and were not reached. Runs are unpaired; background load can vary. CPU is process-tree user+kernel time; latency quantiles use linear interpolation.

Mimic does not render pixels. These fixtures do not establish general website compatibility or a universal speedup. Exact binary, fixture, adapter and capture hashes are available in the methodology and data.