September 2026 · 47-page PDF
AI Accelerator Field Guide
Your model needs a home. Work out which accelerator has enough memory, which connections matter, and what to measure before committing to a cluster.
This edition brings together 37 hardware profiles, 9 model references and 34 cloud configurations, with practical calculations and links back to the live reference.
Get the free PDF ↓Sign in to download. A free account is enough; no subscription required.
Reviewed 12 September 2026. A dated reference edition; availability and software support can change.
What you can work out with it
01
Read a spec sheet
Memory units, bandwidth, compute precision and the three different interconnects.
02
Compare the hardware
NVIDIA, AMD, Google TPU, AWS accelerators and a separate note on DPUs.
03
Fit the model
Weight memory, KV cache, quantization, MoE and worked capacity estimates.
04
Measure the service
Time to first token, token latency, goodput and cost under a stated workload.
05
Choose a cloud configuration
Provider mappings, topology caveats and a checklist before renting.
06
Check the evidence
Linked hardware profiles, model recipes and primary sources for follow-up.
A calculation you can use today
A hypothetical 70-billion-parameter model needs about 140 GB for BF16 weights: 70 billion × 2 bytes. At 8 bits, the raw weight estimate falls to 70 GB. Neither figure includes KV cache, quantization metadata or runtime memory.
An 80 GB card with a 10% reserve leaves 72 GB. If weights take 70 GB and cache plus runtime need 12 GB, the 82 GB budget exceeds one card. Two cards pass this arithmetic screen, but the runtime still needs a supported way to partition the model.
Try your own model in the memory calculator →How to read the evidence
Specification tables come from the site’s reviewed hardware catalog. Primary sources and model recipes are linked in the report. Worked examples are estimates with stated inputs. We have not independently benchmarked these accelerators, and memory fit does not establish tokens per second, latency or runtime compatibility.
The PDF is useful offline. For the current source notes and corrections, use the live hardware reference. For calculations, open the calculator collection.