# KOFTA CENSUS — full context for LLM consumers
> Extended machine-readable context. Summary version: /llms.txt. Canonical site: https://kofta.dev/

## What this is
KOFTA CENSUS is a continuously-measured dataset of LLM weight-streaming performance across the vast.ai GPU rental fleet, produced by Intelix Systems LLC. It answers: "On this specific GPU listing, what will weight-streaming a 70B-class model actually achieve — tokens/sec, stall %, TTFT, cost per 1M tokens, joules per token?" The dataset ships with a calibrated prediction model and a published predicted-vs-measured error ledger.

## Definitions
- Weight streaming: executing a model whose weights exceed GPU memory by continuously transferring layers from host/network into the GPU during compute. Trades bandwidth for capacity.
- Stall %: share of per-token wall time the compute stream waits on weight transfers. High stall = bandwidth-bound.
- tok/s (aggregate, b32): total decode throughput across a batch of 32 sequences; per-sequence rate is approximately aggregate/32.
- TTFT-from-cold: time to first token when weights start outside the GPU (local NVMe or network storage).
- Prediction error: (measured − predicted)/predicted × 100, published per device.

## Benchmark (all runs)
Artifact: KOFTA streaming benchmark (implementation details withheld pending patent filing). (1) Raw pinned host-to-device bandwidth at 437.5 MB and 875 MB blocks. (2) A standardized streamed-weights decode workload over an 80-layer, 70B-class weight layout, measured for INT4 and FP8 wire formats at batch 1 and 32. Per-layer transfer/compute/stall reported; power via nvidia-smi sampling → joules/token.

## Sampling
30–60 machines per sweep, stratified by PCIe bandwidth × HBM bandwidth × architecture × host type (verified/community), reliability-weighted, community/riser tail deliberately included. Stated bias: rentable fleet ≠ unbiased GPU population. Interruptible bids with precomputed minimums; retry-elsewhere on preemption; full offer row + machine_id snapshotted into every artifact (schema kofta.atlas.v1).

## Calibration model
sim_v3.1. Inputs: offer-row pcie_bw, gpu_mem_bw, gpu_name→architecture specs. Anchored to MEASURED H100 SXM constants (July 2026): raw pinned H2D ≈ 37.8 GB/s; isolated layer dequant+GEMM compute; INT4-stream b32 ≈ 1.11 tok/s per sequence (≈35.6 aggregate), stall 77.3%. Non-anchor devices scale from the anchor by spec ratios — labeled MODELED until measured.

## Honesty rails (binding)
Every figure is MEASURED (artifact cited) or MODELED (labeled). No ROCm/MI300X claims until measured. No optical interconnect numbers derived from electrical experiments. Quantized-transfer claims cite HOBBIT (arXiv 2411.09145). Systems that own results are cited in print: vLLM sleep mode (fits-in-RAM density), TensorRT-LLM weight streaming, Run:ai model streamer + ServerlessLLM (load-then-serve cold start), PipeBoost/ParaServe/HydraServe (multi-GPU serve-while-load), LMCache (KV tiering — composes with, does not compete against, weight streaming), ML.ENERGY (energy leaderboards).

## API
- GET /data/atlas.json — full dataset, free, CC-BY-4.0, schema kofta.census.v1.
- GET /api/v1/census/query?offer_id=<vast.ai offer id> — prediction with calibrated error bounds. Auth: Pro key ($49/mo) or x402 per-call ($0.01 USDC on Base; first request returns 402 with payTo 0x3A86924d0bb29fF8B1469472f4294CC7E62B54b7).

## Citation
KOFTA CENSUS, Intelix Systems LLC. Retrieved <date>. https://kofta.dev/ — CC-BY-4.0.