Calculator
HBM Demand Calculator
Capacity is manufacturer-rated decimal GB. Parameter ceilings cover raw weights only; runtime memory, activations, KV cache, fragmentation, and distributed-system overhead are excluded.
AI infrastructure announcements often lead with accelerator counts while leaving the memory consequence abstract. This calculator applies manufacturer-published HBM capacity and peak-bandwidth specifications to a fleet size, then shows how much memory is installed, how much remains after a user-set reserve, and how many raw model parameters that usable capacity could hold at common precisions. It is a transparent scenario tool, not a shipment forecast.
How this hbm demand calculator works
The calculator multiplies a selected accelerator's manufacturer-rated HBM capacity and peak memory bandwidth by the fleet size. It applies your reserve only to the model-weight estimates; installed physical capacity and aggregate peak bandwidth remain unchanged.
Formula
Installed HBM = accelerators × GB per accelerator; raw parameters = usable GB ÷ bytes per parameter
This is a physical-capacity scenario, not an HBM shipment, supplier-share, pricing, or revenue forecast. Mixed fleets, yields, inventory, packaging, qualification, and delivery timing require separate evidence.
Before you use the result
Assumptions
- • Every accelerator in the scenario uses the selected model and published memory specification.
- • FP16 uses 2 bytes per parameter, INT8 uses 1 byte, and 4-bit weights use 0.5 byte.
- • Published peak bandwidth is aggregated arithmetically; real workloads do not sustain it perfectly.
Quick start
- 1. Choose the accelerator model closest to the deployment you want to test.
- 2. Enter the accelerator count from a disclosed or hypothetical fleet.
- 3. Adjust the reserve to reflect runtime memory and operational headroom.
Frequently asked questions
What does the HBM demand calculator measure?
It multiplies an accelerator's published HBM capacity and peak memory bandwidth by the number of accelerators in a scenario. It also applies your reserve assumption to estimate usable capacity and a raw model-weight ceiling.
Does installed HBM capacity equal annual HBM shipments?
No. Installed capacity is a point-in-time fleet scenario. Shipment demand depends on production timing, inventory, yields, replacements, accelerator mix, and whether systems are actually delivered.
How is model-weight capacity calculated?
The tool uses the simple weight-storage identity: parameters multiplied by bytes per parameter. FP16 uses two bytes per parameter, INT8 uses one, and 4-bit weights use one-half byte. Runtime overhead, activations, KV cache, fragmentation, and parallelism are not included.
Why include a memory reserve?
Real systems cannot devote every advertised gigabyte to model weights. Runtimes, activations, KV cache, communication buffers, and operational headroom all consume memory. The reserve is editable because the correct amount depends on the workload.
Can this calculator predict HBM supplier revenue?
No. Contract pricing, stack configuration, yields, qualification, packaging, and supplier mix are not public enough for a defensible revenue forecast. The calculator intentionally stops at physical capacity and bandwidth.
More calculators
The math is easy. The macro is hard.
Yield Theory turns macro data, filings, and capital flows into a source-linked monthly thesis with catalysts, risks, and written breakpoints. Founding price $15/mo.
$15/mo · cancel anytime · sources included