Editing Openai/6925f847-0600-800d-8a2e-57670d2a3feb (section)

==== 3. How they stack up in a 10 GW “AI factory” ====

If you hold total watts fixed:
* Throughput side: - MI450X gives ~3× the FP8 and ~3× the FP4 of Ironwood at the same 10 GW, and about 2× the FP8 and ~1.3× the FP4 of Rubin.
* Memory side: - MI450X cluster has >2.5× the HBM of a Rubin cluster at the same power, and ~1.6× Ironwood’s memory. - HBM bandwidth at 10 GW is also highest for MI450X (~140,000 PB/s).

Intuition:
* Ironwood wins on tight Google/Pathways integration and inference focus, but on raw flops & HBM per watt it’s behind the next-gen HBM4 GPUs.
* Rubin is the “max FP4” monster with extreme power draw; at equal facility power it gives a bit less FP8 and much less HBM than MI450X, but more FP4 than MI450X if you normalized per device rather than per watt.
* MI450X is the “dense memory + efficiency” option: at a fixed 10 GW budget you get the most HBM, the most bandwidth, and the most FP8/FP4 among the three, assuming these roadmap numbers hold.

If you want, I can redo the 10 GW table under different power assumptions (e.g., Rubin at 2.5 kW, MI450X at 1.5 kW, Ironwood air-cooled at 800 W) or switch the metric to tokens/s per watt for some target model size.