03 / THE CHIPS
Comparing GPU
specifications
Memory determines what fits. Bandwidth determines how quickly data moves. Compare the specification that matters to your workload.
A MEMORY BUDGET
A 70-billion-parameter model starts at about 140 GB.
At 16-bit precision, each weight uses 2 bytes. That is 70 billion × 2 bytes = 140 GB for weights alone.
COMPARE THE SPECIFICATIONS
Select a measure, then an accelerator.NVIDIA accelerators vs. H100
Vendor specifications. Larger is not always better.
Selected accelerator · specifications & buying context
What these numbers can—and can’t—tell you
- Use this for
- Comparing memory, bandwidth, power, and connection specifications.
- Keep in mind
- Vendor specifications are not real-world workload benchmarks.
Compute figures are not directly comparable across precision and sparsity. Hopper's displayed tensor figure uses FP8; Blackwell's and Rubin's use FP4. Benchmark the target model and serving or training stack before converting the displayed specification into a performance or price premium.
Roadmap cadence belongs in the contract model. A multi-year capacity term may overlap successor launches. Incumbent pricing still depends on supply, workload demand, provider financing, and successor availability, so address refresh, substitution, and rate-review rights explicitly.
Facility readiness becomes part of product fit at rack scale. Higher rack power density and liquid cooling requirements narrow the set of deployable facilities. Request delivered power, cooling, commissioning, and acceptance milestones alongside accelerator availability.
† Rubin memory and bandwidth are confirmed vendor specifications as of July 2026. The displayed per-GPU power remains a planning assumption; partner configuration, availability, and realized workload performance require current confirmation.
specifications verified 22 July 2026 · H100 · H200 · Blackwell · Rubin · methods