Strata Token/s Estimator: Model Speed and Hardware Upgrade Savings

This estimator brings together four model configurations measured by this article’s author, Ethan Lab, on an i7-11700 / RTX 5060 Ti 16GB / 64 GB RAM PC, alongside official and community Strata benchmarks. It estimates tokens/s and fixed-workload time after GPU, DDR memory or PCIe changes. Ethan Lab’s local measurements anchor some model profiles; the public comparison uses the Strata project author’s Ryzen 5 7600 / RTX 5070 12GB PC. Results for the selected target configuration remain conditional formula estimates.

Local test PC memory: 32 GB × 2, dual-channel, 64 GB total (DDR4-3200).

How to use it

1. The default is IQ2_XS (Swift 1.5) on the complete author-PC configuration. IQ2_XS means the original model; its clearly marked Swift comparison proxy is not an author measurement of the original model.
2. Choose a desktop GPU and CPU, then enter DDR type, actual transfer rate, channels, capacity and PCIe generation. Changing only the GPU/CPU does not change memory or motherboard settings; Reset to author PC restores the complete configuration. Open hardware parameters to enter an unlisted component.
3. Compare Prefill/Decode tokens/s, synthetic total time and change versus the formula prediction for the author PC. The same model and hardware give 0%; negative time change means faster and positive means slower. Errors against reported benchmarks are separate, with the original IQ2_XS Swift proxy clearly marked.

Author reference and assumptions

The author PC has a Ryzen 5 7600, RTX 5070 12GB, 64 GB DDR5-5200 and PCIe 5.0 ×16. The formula assumes dual-channel RAM bandwidth of 83.2 GB/s and effective H2D of 50 GB/s; these are not measured bandwidths for that PC. Q2_0 uses v0.1.36 release data; the other measured models use the v0.1.26 single-run matrix with 256 output tokens per cell. Each model shows its source and version. No matching original IQ2_XS author measurement is available, so Swift’s 2,092.02 / 78.60 tokens/s is an explicitly marked comparison proxy.

What the numbers mean

Prefill follows a nominal 32K-input benchmark. Decode follows generation after a nominal 4K input—not a measured 4K-token output. Time is input tokens / Prefill + output tokens / Decode, defaulting to 51,200 + 4,096 tokens. It is a comparison index, excluding tools, builds, loading, queues, prompt caching and the full time needed to finish a task.

Custom models and AMD

Custom models require P0/D0 rates for the author’s Ryzen 5 7600 / RTX 5070 PC and an expert-RAM arena size, using the nominal 32K Prefill / 4K-input Decode protocol. Parameter count, bpw or expert count alone cannot determine tokens/s. Hardware extrapolation for custom architectures is not independently validated. AMD specifications are selectable, but numbers appear only after explicit exploratory opt-in. They are not AMD benchmarks or a validated NVIDIA/AMD ranking.

Scope, validation and sources

The v4 foundation uses 15 configurations from 14 independent machines, including two rig groups with OCuLink/eGPU or dock reports. Removing those groups and refitting increased cross-validation error on the same retained target rigs, so the historical data is retained with this limitation disclosed. This does not establish a reliable formula for arbitrary eGPU/dock connections, which this tool does not estimate. The interface targets a single NVIDIA desktop GPU with 11–32 GB-class VRAM, experts split between RAM and GPU, and no paging. Potential fully resident cases—including Coder on 24/32GB GPUs and Q2_0 on 32GB GPUs—are excluded. RAM capacity does not automatically increase speed; it is primarily a feasibility check. Automatic H2D values are disclosed assumptions, not measurements of your computer.

Historical v4 exploratory grouped holdout MAPE for original IQ3_XXS was 19.81% Prefill, 8.47% Decode and 10.88% synthetic time. These figures describe that dataset and formula. They are not universal accuracy claims for this revision, other models or AMD, nor per-prediction error bounds. Exact agreement with the author’s IQ3_XXS result is an existing calibration constraint, not independent validation.

Audit findings and evidence limits

The October 16, 2026 audit confirmed that author GPU Compute 14,333 and CPU Mark 26,960 match the saved calibration. Existing v4 hardware coefficients and verified local observations are retained. The equivalent author-centered formula is P̂ = A_P × ∏(x / x_author)^β_P, with the same form for Decode. A is the formula estimate on the author PC; B is its reported measurement or proxy. They need not match. Absolute target predictions are unchanged. Hardware total-time change uses input / A_P + output / A_D as its denominator, so identical configurations give 0%. Differences between A and B are separately labelled prediction errors; no prediction is forced to match a benchmark. Author-PC Swift Prefill remains about 18.80% below its reported measurement, as shown in the table above. Current official Q2_0 rates of 2,653 / 93.5 differ from the older calibration batch of 2,170.85 / 92.97; hardware alone cannot explain source-batch differences.

The audit also tested model-specific bias corrections, forcing baselines to the author results, and VRAM adjustments that align local and author measurements. Those candidates did not improve the corresponding held-out or external checks, and some errors increased, so they were not adopted. Matching every known measurement does not demonstrate better predictions on new PCs. More independent hardware/model pairs measured with comparable protocols are needed.

Retained local checks: original IQ2_XS has October 8 three-run medians of 1,628.57 / 67.25 tokens/s; IQ2_XS (Swift 1.5) has October 11 nine-run medians of 1,622.10 / 65.55. These are independent observations on an i7-11700 / RTX 5060 Ti PC, not the author-PC comparison rates above. Prompt files match, but dates and sample counts differ; they do not establish equal throughput or a fixed ratio on every machine.

Inputs are calculated in your browser. This tool does not transmit them to a server.