# AVEVA System Platform and Historian AI Assistant Sizing and Value Calculator

> Source: https://netray.co/tools/aveva-system-platform-ai-assistant-sizing-value-calculator
> Publisher: Netray (netray.co), AI and integration engineering for ERP and manufacturing systems.
> Section: Free Tools

## Short answer

Estimate peak concurrency, GPU count, index size, hours saved and payback for a private on-prem AI assistant on AVEVA System Platform and Historian.

## Category

mes-shop-floor

## Erp

AVEVA System Platform / Wonderware Historian

## Industry

Manufacturing, Aerospace and Defense

## Audience

- IT manager
- CIO
- Controls engineer
- Plant manager

## Tool

**Kind:** calculator

**Currency:** $

**Inputs:**

#### Named users

**Key:** users

**Type:** number

**Default:** 150

**Min:** 1

**Step:** 10

**Unit:** users

**Help:** People with access to the assistant.

#### Daily active share

**Key:** active_share

**Type:** slider

**Default:** 35

**Min:** 5

**Max:** 100

**Step:** 5

**Unit:** %

**Help:** Share of named users who ask at least one question on a typical day.

#### Questions per active user per day

**Key:** qpd

**Type:** number

**Default:** 4

**Min:** 0

**Step:** 1

**Unit:** questions

**Help:** Typical range is 2 to 10 once people trust the tool.

#### Average prompt plus answer size

**Key:** tokens

**Type:** number

**Default:** 2000

**Min:** 100

**Step:** 100

**Unit:** tokens

**Help:** Includes retrieved context and tool results. Grounded answers are often 1,500 to 4,000 tokens.

#### Documents to index

**Key:** docs

**Type:** number

**Default:** 20000

**Min:** 0

**Step:** 1000

**Unit:** documents

**Help:** SOPs, work instructions, logs, reports.

#### Average pages per document

**Key:** pages

**Type:** number

**Default:** 8

**Min:** 1

**Step:** 1

**Unit:** pages

#### Minutes saved per question

**Key:** mins

**Type:** number

**Default:** 5

**Min:** 0

**Step:** 1

**Unit:** minutes

**Help:** Be conservative. Measure it in the pilot.

#### Loaded hourly cost

**Key:** rate

**Type:** number

**Default:** 50

**Min:** 0

**Step:** 5

**Unit:** $/hour

#### Throughput per GPU (batched)

**Key:** gpu_tps

**Type:** number

**Default:** 1500

**Min:** 100

**Step:** 100

**Unit:** tokens/s

**Help:** Aggregate generated plus processed tokens per second for your model and GPU under batching. Measure on your hardware; this default is a planning figure.

#### Monthly cost per GPU

**Key:** gpu_cost

**Type:** number

**Default:** 1800

**Min:** 0

**Step:** 100

**Unit:** $/month

**Help:** Amortised hardware, power and support, or private cloud rental.

**Outputs:**

#### Tokens per day

**Key:** tokens_day

**Expr:** users * active_share / 100 * qpd * tokens

**Format:** integer

**Description:** Daily active users times questions times tokens per question.

#### Peak concurrent requests

**Key:** peak_conc

**Expr:** users * active_share / 100 * qpd * 0.2 / 3600 * tokens / 30

**Format:** decimal

**Description:** Assumes 20% of daily questions fall in the busiest hour and a single stream generates about 30 tokens per second.

#### GPUs needed (before redundancy)

**Key:** gpus

**Expr:** max(1, ceil(users * active_share / 100 * qpd * tokens * 0.2 / 3600 / gpu_tps))

**Format:** integer

**Description:** Busy-hour tokens per second divided by throughput per GPU, rounded up, minimum one. Add a spare for availability.

#### Retrieval index size

**Key:** index_gb

**Expr:** docs * pages * 12288 / 1000000000

**Format:** decimal

**Description:** Assumes 2 chunks per page, 1024-dimension float32 vectors plus 50% overhead for text and metadata. Result in GB.

#### Hours saved per month

**Key:** hours_month

**Expr:** users * active_share / 100 * qpd * 21 * mins / 60

**Format:** hours

**Description:** Assumes 21 working days per month.

#### Monthly value of time saved

**Key:** value_month

**Expr:** hours_month * rate

**Format:** currency

#### Net monthly value after GPU cost

**Key:** net

**Expr:** value_month - gpus * gpu_cost

**Format:** currency

**Primary:** true

#### Simple hardware payback

**Key:** payback

**Expr:** if(gt(net, 0), gpus * gpu_cost * 36 / net, 0)

**Format:** months

**Primary:** true

**Description:** Assumes the monthly GPU cost represents a 36-month amortisation, so purchase cost is about 36 times the monthly figure. Hardware only; excludes implementation. Zero means no payback at these inputs.

**Chart:**

#### Monthly value and net value vs named users

**Type:** line

**X:** users

**From:** 20

**To:** 1000

**Steps:** 25

**Series:**

- value_month
- net

**Disclaimer:** Planning estimate only. Throughput per GPU depends on model, quantisation, context length and serving software, so benchmark your own stack. Time saved is the weakest input, so measure it in a pilot before committing budget.

## Content

**Intro:** This calculator sizes a private, on-prem AI assistant for AVEVA System Platform and Historian and estimates its value. Enter users, usage, document volume and costs to see peak concurrency, the number of GPUs, the index size, hours saved and a simple payback.

**Sections:**

#### How to use it

**Body:** Start with the people who will actually use the assistant, not everyone with a login. Set the daily active share and questions per user from a pilot or a similar rollout. Keep the tokens figure realistic: a grounded answer includes retrieved context and tool results, which is usually much larger than the visible question and answer. Then enter your documents, minutes saved per question and cost assumptions. The chart shows how value and net value change as the user count grows.

**Bullets:**

- Use conservative minutes saved
- Benchmark throughput per GPU on your hardware
- Add a spare GPU for availability
- Re-run after the pilot with measured data

#### The formulas and assumptions

**Body:** Tokens per day equal active users times questions times tokens per question. The busy hour is assumed to carry 20 percent of daily questions, so peak tokens per second is daily tokens times 0.2 divided by 3600. GPUs equal that peak divided by the throughput you enter, rounded up. Peak concurrency divides peak requests by a 30 tokens per second stream rate. The index assumes two chunks per page at 1024 dimensions. Value assumes 21 working days a month.

#### Reading the result for AVEVA System Platform and Historian

**Body:** For AVEVA System Platform and Historian, most of the cost is usually fixed: one or two GPUs, the integration work and the care of the data pipeline. That is why net value rises quickly with users up to a point. If net value is negative at your real user count, the honest conclusion may be that a smaller pilot, a shared model server, or your vendor's own tools are the better first step.

#### Common mistakes

**Body:** People overestimate daily active share, ignore the retrieval context in token counts, assume vendor benchmark throughput without batching and long contexts, and count every saved minute as cash. Another frequent error is forgetting redundancy. Treat the output as a range, test it on a pilot, and update the inputs with measured numbers.

**Cta:** If you want these numbers checked against your actual environment and a pilot designed to measure them, Netray can size and build the private AI layer with your team.

**Faq:**

#### Why is the GPU count so sensitive to tokens per question?

**Answer:** Because most tokens come from retrieved context and tool results, not from the question. Doubling the context roughly doubles the load, so trimming context with good retrieval and summaries is the cheapest way to cut GPUs.

#### What throughput per GPU should I enter?

**Answer:** Benchmark your chosen model and serving stack with realistic prompts under concurrent load. The default is a planning figure, and real results vary widely with model size, quantisation and context length.

#### Is the payback figure reliable?

**Answer:** It is a rough hardware-only measure under stated assumptions. It excludes implementation and data preparation, and the time-saved input is the least certain. Use it to compare scenarios, not to promise a return.

#### Does the index size include the model?

**Answer:** No. It covers only the vector index and metadata for your documents. Model weights need separate disk and GPU memory, typically tens of GB depending on the model and quantisation.

## Search terms

System Platform AI assistant sizing, how many GPUs for on-prem LLM, on-prem LLM sizing calculator, private AI ROI calculator manufacturing, System Platform AI business case, tokens per day GPU sizing, RAG index size calculator, local LLM GPU count estimate

---
Talk to an engineer: https://netray.co/contact?intent=erp_ai | Machine catalog: https://netray.co/agent/catalog.json