GPU VRAM vs Model Size Ladder

Inspectable records

Ledger state

The current ledger marks this instrument verified clean within the source-bound verification corpus.

Area

Computing & Information

Tranche

tranche-4

Evidence ID

t4-052

Ledger rows

1

No ledger note is attached.

Review findings

The weight-storage portion is standard under the stated byte and overhead assumptions. The instrument cannot receive a PASS until the KV-cache equation, architecture inputs, KV dtype behavior, overhead semantics, output units, and GPU comparison ladder are made explicit. The honest-label warning is appropriate but should also state that activations and temporary runtime memory may be excluded.

Formula and scope findings

• The KV-cache term is not defined as an actual formula, so its value is not uniquely computable.
• The declared inputs omit architecture values required for KV-cache estimation, including layer count, head count or KV-head count, and head dimension.
• The formula does not specify whether heads means attention heads or KV heads; this materially changes cache size for GQA or MQA models.
• The formula does not specify whether KV-cache dtype is the same as weight precision. These are often different.
• The formula does not define the overhead factor, its allowed range, or whether overhead applies only to weights or also to KV cache and runtime buffers.
• The claimed fits-on-GPU comparison ladder requires a GPU-capacity table and a comparison rule, neither of which is specified in the formula.

Unit assumptions

• Parameter count is a raw count, not millions or billions; values such as 7000000000 mean 7 billion parameters.
• Bytes per parameter are assumed to be fp32=4, fp16=2, int8=1, and int4=0.5.
• VRAM outputs are reported in bytes and binary GiB, where 1 GiB=1073741824 bytes.
• The standard KV-cache formula is assumed to be 2 × batch_size × context_length × layer_count × KV_head_count × head_dimension × KV_bytes_per_element; the factor 2 accounts for keys and values.
• The test vectors explicitly provide layer count, attention/KV head count, head dimension, and KV-cache dtype because the supplied input list does not.
• The overhead factor is assumed to multiply weight storage only, as suggested by params × bytes_per_param × overhead_factor + KV cache.

Boundary hardening

BREAK_CLAIMED · SELF_FLAGGED

An extreme but numerically representable parameter_count causes the JavaScript arithmetic to overflow to Infinity, so the instrument must reject or explicitly flag non-finite estimates. The specification also leaves int4 packing, metadata overhead, and GB-versus-GiB display conventions ambiguous.

Errata and corrections

No applied erratum is attached to this unique-name record.

Inspectable example

4 review vectors are recorded. One source-bound sample is shown below.

{
  "inputs": {
    "parameters": 7000000000,
    "precision": "fp16",
    "bytes_per_param": 2,
    "overhead_factor": 1.2,
    "context_length": 4096,
    "batch_size": 1,
    "layers": 32,
    "kv_heads": 32,
    "head_dimension": 128,
    "kv_dtype": "fp16",
    "kv_bytes_per_element": 2
  },
  "expected_outputs": {
    "weight_bytes": 16800000000,
    "kv_cache_bytes": 2147483648,
    "total_vram_bytes": 18947483648,
    "total_vram_gib": 17.646
  },
  "working": "Weights = 7000000000 × 2 × 1.2 = 16800000000 bytes. KV cache = 2 × 1 × 4096 × 32 × 32 × 128 × 2 = 2147483648 bytes. Total = 16800000000 + 2147483648 = 18947483648 bytes. GiB = 18947483648 ÷ 1073741824 = 17.646 GiB."
}

Related workflows

No source-defined multi-instrument workflow membership is recorded.

Sources and evidence

These pointers bind the page to evidence. They do not grant authority or collapse verification into deployment.

• VERIFIED_LEDGER.json · VERIFIED_CLEAN
• evidence-tranche4/tranche4-results.json
• BOUNDARY_HAZARD_CATALOG.json