Pillar 2 of 3

Model Operational Metrics

This is the math you do need — and it's operational, not algorithmic. Four numbers drive almost every architectural decision: how much a model can read at once, how fast it responds, how much hardware it needs, and how good its answers actually are.

Context window — how much can it process at once?

An LLM can only "see" what fits in its context window, measured in tokens. Everything counts against it: system instructions, retrieved documents, conversation history, and the answer being generated. The window is your hard budget when designing prompts and RAG.

1 token ≈ ¾ word

1,000 tokens ≈ 750 English words ≈ 1½ pages. API pricing is per token — input and output both bill.

4K–8K window

A few pages. Enough for a chat turn plus a short document. Typical of small local models.

128K–200K window

A whole book or codebase slice. Frontier hosted models. Lets RAG pass in many retrieved chunks.

Bigger ≠ better recall

Models attend unevenly across huge contexts ("lost in the middle") — and every token costs money and latency. Retrieve selectively instead of dumping everything in.

Architect's question: "What is the largest input this workflow will ever need in one shot — and does it fit, with room for the answer?" If not, you need chunking, summarization stages, or retrieval — that's an architecture decision, not a model setting.

Latency vs throughput — which one does your use case buy?

Latency is how long one user waits for one answer. Throughput is how much total work the system completes per unit of time. GPUs love batching, so the two trade off: tuning a deployment for throughput (big batches) makes individual answers slower, and vice versa.

ScenarioOptimize forFeels right whenArchitecture lever
Chat assistantLatencyFirst token < 1 s, then 30–100 tokens/s streamedStreaming responses, small/fast model for first draft
Document pipeline (overnight batch)ThroughputDocuments/hour per GPU, not seconds/docBatching, queues, spot/off-peak compute
Image generationLatency (perceived)2–10 s with progress feedbackFewer diffusion steps, smaller resolution previews
Real-time vision (defect detection)Both — hard limitsEvery frame in < 30–50 msSmall CNN on edge hardware next to the camera
Rule of thumb: anything a human is actively waiting for is a latency problem; anything a queue is waiting for is a throughput (cost) problem. Mixing both workloads on one deployment usually means the interactive users lose.

Compute & VRAM — will it fit on the GPU?

The sizing rule of thumb: at FP16 precision a model needs about 2 GB of VRAM per 1 billion parameters just to load the weights. Quantization stores weights in fewer bits — INT8 halves the footprint, 4-bit quarters it — with a modest quality cost that is often acceptable in production. Then add ~10–20% headroom for the KV cache and activations.

VRAM needed to load model weights, by size and precision

GB of GPU memory · weights only — add ~10–20% for KV cache & activations

FP16 (~2 GB / 1B params) INT8 (~1 GB / 1B) 4-bit (~0.5 GB / 1B)
7B 7B · FP16 ≈ 14 GB 14 7B · INT8 ≈ 7 GB 7 7B · 4-bit ≈ 3.5 GB 3.5 13B 13B · FP16 ≈ 26 GB 26 13B · INT8 ≈ 13 GB 13 13B · 4-bit ≈ 6.5 GB 6.5 70B 70B · FP16 ≈ 140 GB — multi-GPU territory 140 70B · INT8 ≈ 70 GB 70 70B · 4-bit ≈ 35 GB 35 0 50 100 150 GB
Why quantization matters to you: a 70B model at FP16 is a multi-GPU cluster; at 4-bit it fits a single 48 GB card. That's a budget line item, not an ML detail.
Back-of-envelope sizing: 7B at FP16 ≈ 14–16 GB → fits one 24 GB card (NVIDIA L4 / A10). 13B quantized → still one 24 GB card. 70B → either 4-bit on a 48 GB card or a hosted API. This arithmetic — plus tokens-per-second and price-per-million-tokens — is most of the "10% math" your role needs.

Evaluation metrics — reading a Data Scientist's report

You won't compute these, but you must judge production readiness from them. For classifiers, everything starts from four possible outcomes of a prediction:

The confusion matrix behind precision & recall

model says YES model says NO actually YES actually NO TP ✓ caught it FN ✗ missed it FP ✗ false alarm TN ✓ correctly ignored precision reads the left column recall reads the top row
MetricPlain-language questionPrioritize it when…
Precision"When the model flags something, how often is it right?" (TP / TP+FP)False alarms are expensive — fraud alerts, content takedowns, anything that pages a human
Recall"Of all the real cases, how many did it find?" (TP / TP+FN)Missing a case is expensive — medical screening, safety defects, compliance checks
F1-score"Balanced single number combining both"You need one comparable score, and both error types matter roughly equally
FID"How close do generated images look to real ones?" (lower = better)Judging generative image models (GANs, diffusion) — pair it with human review
How to read the report like an architect: never accept "94% accuracy" alone. Ask (1) which error type hurts the business more — that picks precision vs recall; (2) what the baseline is (what does the current manual process score?); (3) whether the test data resembles production data. A model is production-ready when the expensive error rate is acceptable — not when the average looks good.
← Pillar 1: Model SelectionFamilies, traits, cheat sheet