Name: Mastering AI Compute
SKU: 69150776
Price: 28.22 EUR
Availability: InStock
Author: Neil Han
ISBN: 9798178418642

Type

Physical Book

Publisher

Independently published

Author

Neil Han

Language

English

Pages

166

ISBN13

9798178418642

Mastering AI Compute

Neil Han (Author) · Independently published · Physical Book

New Book Imported to Netherlands

Delivery: 02 Nov - 04 Nov

€ 28,22

Import costs and 9% BTW included in the price ✅

€ 28,22

Add to Wishlist

Synopsis "Mastering AI Compute"

A faster GPU does not automatically produce a faster training job or a cheaper inference service. The result depends on the whole path: workload shape, memory traffic, communication, recovery, scheduling and the definition of useful output.Mastering AI Compute teaches AI infrastructure engineers and experienced systems engineers how to reason about that path. Start with transformer shapes, a training step and the generation loop. Build workload budgets, read GPU profiles, map parallelism and communication, account for checkpoint durability, and diagnose admission, placement and data stalls. Then follow serving from prefill and decode through KV-cache allocation, batching, quantization and speculative decoding to latency, quality and cost per useful output.The book works through named systems including FlashAttention-3, NCCL, FSDP2, Megatron, Kubernetes, Kueue, vLLM and MLPerf. Published GPU profiles and logs are embedded as annotated tables and original data figures. CPU worksheets and a tested small training/restart example accompany the chapters. Evidence labels distinguish published results, calculations and book-authored measurements. GPU and cluster experiments are recipes for the reader to run on their own hardware.For engineers who design, operate and pay for AI compute, this is a systems reference for asking better questions, reading the evidence and choosing the next test.