RAM, VRAM, HBM—and the denominator problem

The size of a checkpoint on storage, the resident host memory of a process and the accelerator memory required during inference are related quantities, but they are not the same measurement. A change can improve one while leaving another unchanged.

Modern heterogeneous systems also move state across memory spaces. Device memory, system memory, managed allocations, caches and transfers can trade capacity against bandwidth or latency. A number without its location and access conditions cannot describe that physical system.

The operating footprint may include weights, activations, attention or recurrent state, key-value caches, routed experts, temporary workspaces, compilation products, allocator reservations and framework overhead. Which components dominate depends on architecture and workload.

A component-level improvement remains useful, but it should be named as such. The complete-system result should be measured after all candidate changes interact inside the target runtime.

• Persistent model state
• Workload-dependent state
• Temporary execution buffers
• Runtime and allocator overhead
• Replicated or distributed state
• Required operational headroom

Peak memory determines whether a workload fits. Steady-state residency influences serving density and sustained operation. Allocation volume and movement can reveal bandwidth or fragmentation costs that a single peak number hides.

The correct measurement depends on the constraint. A fit claim should report peak conditions. A serving-density claim may require steady-state and concurrency measurements. A system comparison may need both, together with throughput and latency.

Baseline and candidate must share the conditions that make their measurements comparable: hardware, software versions, batch regime, input lengths, concurrency, warm-up, measurement window and accounting rules.

When a necessary condition changes, the difference should be isolated or disclosed. The objective is not to prevent system co-design. It is to stop a changed workload or runtime from being silently counted as a memory reduction.

A ratio is only as physical as the baseline and denominator beneath it.

A VAMANIR result may ultimately include several linked measurements rather than one universal number: artifact size, host residency, accelerator residency, peak state, steady state and operational headroom.

That plurality is a strength. It makes clear which constraint changed, where secondary effects appeared and whether the candidate actually altered the physical economics of the target system.