A GPU benchmark score alone cannot explain Strata performance. This article presents a v4 empirical formula built from 15 configurations across 14 independent machines. It predicts Prefill and Decode separately, then converts them into time for a fixed workload. This is a prediction model developed for this study, not an upgrade-performance guarantee from Strata.
Strata series: Model benchmarks · Performance formula · Upgrade costs and benefits
Data checked October 13, 2026. Scope: original ISTA IQ3_XXS, one NVIDIA desktop GPU in the 11–32GB VRAM class, without RAM or VRAM paging. Do not directly apply it to Swift, other quantizations, laptops, multiple GPUs or configurations with every expert resident in VRAM.
Six hardware inputs, two separate speed predictions
The fixed baseline is an i7-11700, RTX 5060 Ti 16GB, 64GB dual-channel DDR4-3200 and PCIe 4.0 x8. It retains the nine-run medians on Strata 0.1.41: 1,666.44 Prefill and 51.30 Decode tokens/s. The nominal 32K / 4K prompts contained 31,566 / 3,562 actual tokens. New-version measurements appear separately in the benchmark article and are not included in these frozen coefficients.
x = (G, Bv, V, C, R, B)
x0 = (11250, 448, 16, 20585, 51.2, 13.8)
P = 1666.44 × exp(Σ βP[i] × ln(x[i]/x0[i]))
D = 51.30 × exp(Σ βD[i] × ln(x[i]/x0[i]))
T = 51200/P + 4096/D
Time saved (%) = 100 × (1 − T/110.5682330691711)P and D are in tokens/s; T is in seconds. Positive time saved means faster, negative means slower. G is PassMark GPU Compute and C is multithreaded CPU Mark. Both are model averages checked on October 13, 2026, not scores measured on each individual machine. GPU Compute is not G3D Mark, TFLOPS or AI TOPS. PassMark methodology.
| Input and unit | Prefill exponent βP | Decode exponent βD |
|---|---|---|
| GPU Compute G，Ops/Sec | 0.7672002634 | 0.3270955375 |
| VRAM bandwidth Bv, GB/s | 0 | 0.1601723440 |
| VRAM capacity V, nominal class approximated in GiB | 1.0812454543 | 0.7791361997 |
| CPU Mark C | 0.0414423243 | 0 |
| Theoretical RAM bandwidth R, GB/s | 0 | 0.1868414154 |
| Effective RAM→GPU H2D bandwidth B, GB/s | 0.1244549227 | 0.1342949023 |
The formula uses nonnegative ridge fitting in log space with regularization λ = 0.3 and calibration constraints for the local and author machines. No separate version multiplier is applied. A zero coefficient does not mean the hardware has no effect: the limited, correlated samples did not identify an additional positive effect. The exponents are not fixed fractions of runtime, and cannot imply unlimited speed gains from unlimited VRAM.
How accurate is the formula?
Validation removes an entire machine, retunes and refits on the remaining machines, then predicts the held-out machine. When the author’s PC is removed, its calibration constraint is also removed. The local PC is the designated baseline and is excluded from holdout scoring, leaving 13 Prefill groups and 11 Decode/time groups. Runs with only 24 output tokens receive a fitting weight of 0.25.
| Method | Prefill | Decode | Synthetic time |
|---|---|---|---|
| GPU score + VRAM capacity only | 21.74% | 16.32% | 19.85% |
| v4 six-factor model | 19.81% | 8.47% | 10.88% |
After excluding the 24-token subset, v4 time MAPE is 10.28%, but some 40–121-token tests remain. Model selection used the same validation results; no new blind test set has been evaluated. Prefill was overestimated by 93.95% for a 2080 Ti and 57.49% for an OCuLink 5060 Ti, so average error is not an upper error bound for an individual upgrade.
The author’s measured RTX 5070 calibration point is 1,745.16 / 61.64 tokens/s: 95.79 synthetic seconds and 13.37% time saved relative to the baseline. With the author completely excluded, the formula predicts 1,811.02 / 57.51, with +3.87% time error. The former checks calibration consistency; only the latter is a holdout check. Author’s raw measurements.
PCIe and DDR: theoretical bandwidth is not actual transfer speed
| Specification | Bandwidth |
|---|---|
| PCIe 4.0 x8 | 15.75 GB/s |
| PCIe 4.0 x16 | 31.51 GB/s |
| PCIe 5.0 x8 | 31.51 GB/s |
| PCIe 5.0 x16 | 63.02 GB/s |
| DDR4-3200 dual channel | 51.2 GB/s |
| DDR5-5600 dual channel | 89.6 GB/s |
| DDR5-6000 dual channel | 96 GB/s |
PCIe figures account for encoding but not packet or DMA overhead. H2D is also limited by available RAM bandwidth, controllers and concurrent load. The local 13.8 GB/s is measured. Dual-channel DDR5 and H2D values such as 50 GB/s filled into incomplete records are assumptions. In particular, the author’s memory-channel configuration is unconfirmed: neither 83.2 GB/s RAM bandwidth nor 50 GB/s H2D should be presented as measured. PCI-SIG bandwidth reference.
The 5060 Ti is natively x8, so a Gen5 platform provides at most PCIe 5.0 x8. Switching to an x16 GPU may widen the link at the same time: do not count the GPU and PCIe gains twice. Faster H2D may also change Strata’s CPU/GPU work distribution. 5060 Ti specifications; work-distribution source.
Separate VRAM capacity, cache hit rate and RAM capacity
Expert-cache capacity is total VRAM minus the desktop, other applications, shared weights, KV, workspaces and reserve. The local 16GB card had about 7.91 GiB of measured expert-cache space. Holding other allocations fixed, a 12GB card would have about 3.9 GiB. This is a capacity calculation, not a prediction of half the speed. Expert popularity is uneven, so capacity ratios cannot replace measured hit rates.
Some reported hit rates exclude PCIe offload. If H is that hit rate and x is the fraction of all routes offloaded, split the routes into GPU resident = (1−x)H, CPU = (1−x)(1−H), and PCIe dispatch = x. One RTX 5080 report had H = 0.981 and x = 0.401, making the resident-route share about 58.8%, not 98.1%. Raw RTX 5080 record.
RAM capacity is not a direct multiplier in the speed formula. The local experts are already resident in RAM. Without memory pressure, and with other conditions unchanged, increasing RAM from 64GB to 128GB may provide nearly zero direct acceleration. Extra capacity accommodates larger models and more applications and avoids SSD paging. Exceeding VRAM capacity can also cause severe slowdowns, so filling the cache as much as possible is not automatically optimal. RAM placement modes; VRAM-reserve measurements.
Limitations to understand before using it
- 51,200 input + 4,096 output tokens is a standardized comparison workload. It uses 32K Prefill / 4K Decode rates, not complete coding-agent task measurements. Loading, queuing, tools and compaction are excluded; actual Decode speed may change at longer contexts.
- Prompts affect Decode. MTP output per round depends on draft acceptance. The author obtained about 3.19 tokens/round versus about 2.13 in a local sample. Cross-machine differences are not purely hardware effects.
- Missing fields affect predictions. Changing the author’s assumed H2D shifts upgrade time-saving estimates by up to 3.60 percentage points; changing the RAM-channel assumption too shifts them by up to 7.42 points. This is sensitivity analysis, not a confidence interval.
- Average accuracy does not imply universal applicability. Out-of-scope time errors for a modified 48GB card with all experts resident and a laptop were approximately −34.47% / +51.09%. The desktop regression should not be applied to them.
Use this formula to prioritize upgrade hypotheses for testing. To establish actual gains, prefer same-machine A/B tests with fixed models, prompts, sampling and MTP. Measure H2D, cache routing, CPU/GPU overlap and actual tokens produced per round. Strata community benchmark methods and data.
Discussion