The book’s final two pages retell its whole argument — not as a summary of chapters, but as a design method: the ordered questions anyone building a system for efficient DNN processing should ask. This page walks that method, and — since this site hosts an excerpt — doubles as an honest map of which steps this site teaches in full and which live in the missing chapters of the complete book.
Why the field exists
DNNs saw explosive growth and now deliver better-than-human accuracy accuracy Quality of the result for a given task; units depend on the task (% correct for classification, mAP for detection) and must be read against task/dataset difficulty. defined in ch. 3 — open in glossary on vision, speech, and robotics tasks — at the cost of high computational complexity. With general-purpose compute scaling stagnating, the movement is toward domain-specific hardware. The mission statement is exactly Chapter 3’s opening: efficient DNN processing that improves energy efficiency and throughput without sacrificing accuracy and with cost-effective hardware, to expand deployment into existing and new domains.
The design method — and where this site covers each step
Click each stage: what it asks, and where to learn it (here, or in the full book).
The conclusion’s claims, checked against what you’ve seen
Every claim in the two-page conclusion is something this excerpt taught concretely. Click the evidence cells:
| The conclusion says… | …and you saw it at | |
|---|---|---|
| Metrics must be balanced, not maximized singly | accuracy, energy, throughput, cost evaluated together | §3.7 interplay + 6-step checklist |
| A workload = network architecture + input data | both components vary within and between applications | Table 3.1 (§3.2) |
| Data movement dominates energy | the primary focus of recent research is reducing it | Fig. 3.3 + the energy ladder (§3.3); all of ch. 10 |
| Performance depends on the mapping | optimally mapping workloads + throughput/energy models | §3.2 utilization; 10.2.1 loop nests |
| Representation & sparsity co-design must mind accuracy | precision and sparsity changes need network fine-tuning | Eq. 3.4 & 3.9 labs; 10.2.4 footnote 13 |
| Advanced technologies interact — combine with care | memristors, 3-D memory, sensor/in-memory compute, optics | the 10.2.4 cycle bill; 10.3-10.4 pattern box |
Where this leaves you
The book’s own last word: much has been done, and DNN acceleration remains an open, fast-moving field with innovation opportunities at every level of hardware design — it hopes to have provided a structured way of navigating the design space. For this excerpt’s reader, the structure you now own is:
- Chapter 3’s metric vector — the questions to ask of any design, and the reporting rules that keep answers honest.
- Chapter 10’s technology ladder — what happens when you move compute to the data, and the analog fine print (converters, non-idealities, utilization) that headline numbers hide.
The full book’s middle chapters (4–9: kernel computation, dataflows, mapping, precision, sparsity, and efficient-model design) fill the gap between those two — this page’s map shows where each would slot in.
Check yourself
Check yourself — the whole excerpt
1.The conclusion says a DNN workload has two major components. Which are they, and why does the split matter?
2.Why does the book's design method start from applications and their potential evolution, rather than from a hardware idea?
3.Which single fact, established in Chapter 3 and exploited throughout Chapter 10, does the conclusion call the primary focus of recent research?
4.Per the conclusion, what must hardware/algorithm co-design (precision, sparsity, model changes) always keep in the loop?
5.The conclusion urges considering advanced technologies 'in combination, while being careful to understand their interactions.' Give the excerpt's own example of such an interaction.