Skip to playbook content
ClusterBidGPU Playbook
Browse all chapters
ClusterBid playbook · 11 chapters

Choose the system.
Not the sticker.

A practical, sourced guide to how GPUs work, which variant fits, how AI inference is served, when orchestration helps, and how to optimize cost without optimizing away the SLO.

Start the guided path Open the GPU catalog
01FitWhat must reside?
02MathWhich kernels and precision?
03ServiceWhat must users experience?
04TopologyWhere does data cross?
05EconomicsWhat does useful work cost?
Three entry points

Choose your path through the playbook

Read in order if you are new to accelerators. Experienced operators can enter at the decision they are making now.

Path 1

I need a shortlist

Selection gates → catalog → workload benchmark → cost

Start here Path 2

I need to run inference

Prefill/decode → KV → scaling → orchestration → operations

Inference Path 3

I need to operate a fleet

Layer ownership → scheduling → economics → optimization → runbook

Orchestration
The complete program

Eleven chapters, one decision system

Every chapter ends in a rule, worksheet, checklist, or operating artifact - so the playbook can be used during an actual purchase or deployment.

  1. 01

    Start here

    Turn a workload, service target, and operating envelope into a benchmarkable shortlist.

    8 min
  2. 02

    How GPUs work

    A practical model of compute, memory, bandwidth, precision, and the links between accelerators.

    12 min
  3. 03

    GPU catalog

    Compare representative NVIDIA, AMD, and Intel data-center accelerators without mixing incompatible metrics.

    15 min
  4. 04

    Workload selection

    Selection recipes for training, fine-tuning, inference, image and video, rendering, and HPC.

    12 min
  5. 05

    Inference

    Prefill, decode, KV cache, latency metrics, batching, caching, quantization, and capacity planning.

    18 min
  6. 06

    Scaling & topology

    Choose data, tensor, pipeline, expert, and context parallelism - and place them on the right fabric.

    12 min
  7. 07

    Orchestration

    Understand what the inference engine, Ray, Kubernetes, GPU Operator, Kueue, and Volcano each own.

    15 min
  8. 08

    Cost & procurement

    Model fully loaded cost, SLO goodput, purchase options, commitments, and interruption risk.

    14 min
  9. 09

    Optimization

    A measured sequence from idle cleanup and right-sizing through batching, precision, caching, and power.

    12 min
  10. 10

    Operations

    Observe service, engine, GPU, fabric, and scheduler signals - and rehearse failures before production.

    12 min
  11. 11

    Reference

    The reusable math, vocabulary, caveats, and dated primary sources behind this playbook.

    10 min
Evidence standard

Useful beyond the next GPU launch

Stable conceptsArchitecture, formulas, benchmarking, and operating principles are explained from primary sources.

Exact variantsMemory, links, power, and sharing are attached to a named package - not a family nickname.

Changing factsPrice, availability, lifecycle, and compatibility are dated or routed to live official sources.

Claims with conditionsVendor benchmarks stay vendor claims; production decisions require the target workload and SLO.