claude-dev-suite/claude-dev-suite/ai-hardware-selection

AI Hardware Selection

The metric that usually decides: memory, then bandwidth

For LLM inference, the binding constraint is typically VRAM/HBM capacity (weights + KV-cache must fit) and memory bandwidth (decode is memory-bound) — not raw FLOPS. Size first: weights ≈ params × bytes/param (e.g. 70B × 2B(FP16) ≈ 140GB → multi-GPU or quantize). Add KV-cache (grows with context × batch). Only then look at TOPS.

Accelerator families

Other levers

• Interconnect (NVLink, InfiniBand): decisive for multi-GPU training and

tensor parallelism — bandwidth between accelerators bounds scaling.

• Precision support: FP8/INT4 support multiplies effective throughput/capacity.
• Cost/Watt & TCO: cloud per-hour vs owned; power/cooling; utilization. The

honest metric is cost per token (or per request) at target latency.

• Training vs inference: training needs FLOPS + interconnect + memory;

inference needs memory capacity/bandwidth + latency.

When to recommend what

• Default / flexibility / training → NVIDIA GPUs (size by model VRAM).
• Edge/mobile/low-power inference → NPU.
• Hyperscale training on GCP → TPU pods.
• Fixed ultra-low-latency pipeline → FPGA (only if justified).