# strata 0.2.0 on my rtx 30 cards: +12 to +17% decode, but my agents don't finish any faster

a simple engine update takes solo decode from 91.7 to 107.6 tok/s on a 3090 + 3080 Ti, and yet the total time of my agents doesn't move, because they spend their life reading prompts

- date: 2026-10-19
- type: experiment (local ai)
- hardware: 3090 + 3080 Ti, 80gb
- source: measurements made by the author on his own hardware
- html page: https://pkillsleep.dev/en/labo/strata-020-ampere/

## what i wanted to know

Strata, the engine i use to run Qwen 3.8 Flash Next on my two cards, was releasing its 0.2.0 in beta, promising gains on small cards. i wanted to know what it gave on my exact setup, keeping everything else identical (two slots, 262K of context, both cards at 220 W).

## the numbers

on short requests (256 token answers, median of three rounds):

| requests at once | 0.1.45 | 0.2.0 | gain |
|---:|---:|---:|---:|
| 1 | 91.7 tok/s | 107.6 tok/s | +17% |
| 2 | 54.7 | 61.1 | +12% |
| 4 | 54.8 | 63.4 | +16% |

under real coding agent load (five sessions on a real repo, context up to 122K), the median decode of a single agent goes from 101 to 110 tok/s, and with four agents from 22.2 to 27.4 tok/s.

**generation, a single request**

| | tok/s |
|---|---:|
| Strata 0.1.45 | 91.7 |
| Strata 0.2.0 | 107.6 |

**time to finish the same 4-agent load**

| | s |
|---|---:|
| Strata 0.1.45 | 212 |
| Strata 0.2.0 | 215 |

## what surprised me

all the gain comes from a table of nine settings that 0.2.0 turns on by default for Ampere cards (the RTX 30 series): if you turn it off with `STRATA_SM86_PERF=0`, 0.2.0 drops back to 92.4 tok/s, which is exactly the old number.

but the most interesting part is that the total time to finish the four-agent load didn't move (212 s before, 215 s after). a coding agent spends most of its time reading prompts and tool results, not writing, and this update only speeds up the writing. faster tokens don't mean faster agents.

[meme: Cat Looks Inside]

## what i keep

0.2.0, obviously, because it's free and the answers stay the same according to Strata. but if i want agents that finish faster, prompt reading is what needs work.

## What these numbers don't tell you

- a single pass per version: on the agent test, a gap of less than 10% is within the noise
- the gain on short requests matches what Strata announces (+10 to 13% on RTX 3060 and A4000), but i only measured it on my setup