Chip Comparison

Procedure

1. Check comparisons/ for an existing document matching the requested chips. If found, summarize or quote it directly.

2. If no pre-written comparison exists, load the relevant entries from the spec files for each chip.

3. Normalize terminology: When comparing AMD vs NVIDIA chips, use terminology/amd-nvidia.md to map equivalent concepts (e.g. CU ↔ SM, LDS ↔ Shared Memory, XGMI ↔ NVLink).

4. Build a comparison table covering the dimensions most relevant to the query:

   Dimension	Fields to Use
   Memory capacity	gpu_memory_hbm_gb, total_unified_memory_gb
   Memory bandwidth	gpu_memory_bandwidth_tbs
   Compute	fp32_tflops, cuda_cores_fp32_per_gpu (or AMD CU count)
   Power	tdp_kw
   Cost	cost_per_chip_usd_range, platform_cost_usd_range
   Interconnect	cpu_gpu_interconnect, cpu_gpu_interconnect_bw_gbs
   Generation / Year	generation, year

5. Provide a qualitative summary of tradeoffs after the table: workload fit, scale (single chip vs rack), power budget, cost efficiency.

Output Format

• Comparison table first, followed by a "Tradeoffs & Recommendations" section.
• For cross-vendor comparisons, note any terminology differences inline.
• Cite source URLs for each chip.