Validra performance under load

How long each feature takes, and how fast it runs, as more people use it at the same time — measured on the live system.

Measured 30 September 2026

The one thing to know

The server has a single graphics card, and it produces about the same number of tokens per second in total whether one person is waiting or twenty-five.

So waits do not hold steady as people join — they stretch. Ten people share the speed one person would have had alone, and each waits roughly ten times longer.

A token is roughly three quarters of a word, so 10 tokens/sec is about 7 or 8 words a second appearing on screen.

Sequential — each person waits their turn before starting.

Parallel — everyone starts together and shares the machine.

p50 (usual) — the middle result; half were faster, half slower.

p90 / p95 — 9 in 10, and 19 in 20, waited no longer than this.

Slowest — the worst single wait recorded.

Avg tok/s — tokens per second a single request generated at.

Validra Assistant

Answering questions about an attached document

People	How	Runs	Passed	p50 (usual)	p90	p95	Slowest	Avg tok/s
5	Sequential	1	5/5	35 sec	7m 35s	8m 28s	9m 21s	27
	Parallel	1	5/5	35 sec	3m 34s	3m 37s	3m 40s	22
10	Sequential	1	10/10	38 sec	10m 31s	10m 34s	10m 37s	40
	Parallel	2	20/20	44 sec	4m 14s	9m 2s	13m 50s	26
25	Sequential	1	25/25	50 sec	1m 30s	6m 7s	19m 56s	23
	Parallel	2	50/50	4m 28s	12m 22s	12m 30s	12m 43s	13

Its four test questions range from a one-line lookup to building a large table, so the usual and slowest figures sit far apart. Both are real.

AI Editor

Rewriting and generating text inside a document

People	How	Runs	Passed	p50 (usual)	p90	p95	Slowest	Avg tok/s
5	Sequential	1	5/5	34 sec	34 sec	34 sec	34 sec	29
	Parallel	1	5/5	37 sec	39 sec	39 sec	39 sec	26
10	Sequential	1	10/10	34 sec	36 sec	39 sec	42 sec	33
	Parallel	1	10/10	48 sec	54 sec	57 sec	61 sec	24
25	Sequential	not tested
	Parallel	not tested

Document Chunking

Splitting an uploaded document into sections

People	How	Runs	Passed	p50 (usual)	p90	p95	Slowest	Avg tok/s
5	Sequential	2	10/10	14 sec	25 sec	29 sec	32 sec	104
	Parallel	not tested
10	Sequential	1	10/10	14 sec	15 sec	18 sec	20 sec	104
	Parallel	not tested
25	Sequential	1	25/25	14 sec	16 sec	16 sec	28 sec	104
	Parallel	1	25/25	2m 28s	3m 58s	4m 9s	4m 22s	96

This one processes an uploaded document rather than answering a question, so its work is shaped differently from the others.

Data Analysis

Classifying an uploaded data file

People	How	Runs	Passed	p50 (usual)	p90	p95	Slowest	Avg tok/s
5	Sequential	2	10/10	16 sec	25 sec	34 sec	43 sec	25
	Parallel	1	5/5	38 sec	53 sec	57 sec	60 sec	21
10	Sequential	1	10/10	16 sec	18 sec	18 sec	18 sec	24
	Parallel	not tested
25	Sequential	1	25/25	16 sec	17 sec	18 sec	20 sec	22
	Parallel	not tested

Report Master

Writing a full validation report

People	How	Runs	Passed	p50 (usual)	p90	p95	Slowest	Avg tok/s
5	Sequential	2	10/10	5m 5s	5m 7s	5m 7s	5m 7s	31
	Parallel	1	5/5	5m 13s	5m 18s	5m 18s	5m 19s	27
10	Sequential	1	10/10	10m 6s	10m 8s	10m 9s	10m 9s	35
	Parallel	not tested
25	Sequential	not tested
	Parallel	not tested

Mixed load — every feature at once

People spread across all five features instead of queueing on one. This is closer to a normal working day than any single-feature run.

Date	People	How	Passed	Wall clock	Avg tok/s	Who was doing what
2026-09-30	25	Parallel	25/25	9m 34s	71	Validra Assistant 7, AI Editor 4, Document Chunking 6, Data Analysis 5, Report Master 3
2026-09-30	25	Sequential	25/25	11m 15s	13	Validra Assistant 8, AI Editor 3, Document Chunking 7, Data Analysis 4, Report Master 3
2026-09-30	10	Sequential	10/10	4m 46s	33	Validra Assistant 1, AI Editor 1, Document Chunking 2, Data Analysis 3, Report Master 3
2026-09-24	10	Parallel	10/10	5m 17s	65	Validra Assistant 3, AI Editor 2, Document Chunking 1, Data Analysis 2, Report Master 2

Percentiles are computed over every request in a cell. Where a cell holds only five requests, p90 and p95 sit close to the maximum and carry little more information than it does — they earn their keep at 25 people, not at 5.