# telemetria-glm53-c1-20260917-2110

Public raw data package for this measurement. Every number on the measurements page citing this window comes from these files.

Subset of the telemetry window from 2026-09-17 (21:10 UTC): only the two single-stream points — unspeculated (tp8_c1) and native MTP k=3 (tp8_c1_mtp) —, 3 repetitions each. Higher concurrency points from the same window are not published.

Protocol: tensor parallel 8, one request at a time, a synthetic 256-token prompt (a marker word followed by one filler word repeated) sent to the completions endpoint, greedy decoding (temperature 0), 512-token output cap reached by all requests (finish_reason=length across all requests), thinking disabled, 60 s windows after discarded warmup (warmup_* fields in each JSON). The tok/s rate includes TTFT. The 2.04× MTP factor holds for this prompt; on real agent tasks MTP gives 1.52× per task (package agentico-glm53-mtp-20260903).

Counters come from nvidia-smi at 1 Hz across the 8 GPUs. utilization.gpu is the fraction of time with any kernel executing, not SM activity; utilization.memory is the fraction of time device memory was being read or written.

`resumo.json` is derived: recomputed from `*_rep*.json` and `*_rep*_gpu.csv` in this package (field `derivacao`). The `piso_de_banda` block is computed from model and hardware constants, not measured.

Outside package: server metrics logged every 2 s (unused in any figure) and plan log (internal operation).

| file | what it is | sha256 (first 16) |
|---|---|---|
| `resumo.json` | means, standard deviations, and window result (K repetitions per arm) | `a3acf396b78721e9` |
| `tp8_c1_mtp_rep1.json` | arm repetition: aggregates + 1 record per request | `b26f9ce74dc07438` |
| `tp8_c1_mtp_rep1_gpu.csv` | nvidia-smi counters at 1 Hz across the 8 GPUs during the repetition (utilization, memory, power, clocks) | `f2010ab1d9bfdaf2` |
| `tp8_c1_mtp_rep2.json` | arm repetition: aggregates + 1 record per request | `37e3810ba500102e` |
| `tp8_c1_mtp_rep2_gpu.csv` | nvidia-smi counters at 1 Hz across the 8 GPUs during the repetition (utilization, memory, power, clocks) | `87bc7d7ae27228df` |
| `tp8_c1_mtp_rep3.json` | arm repetition: aggregates + 1 record per request | `55e6d7310a4761e2` |
| `tp8_c1_mtp_rep3_gpu.csv` | nvidia-smi counters at 1 Hz across the 8 GPUs during the repetition (utilization, memory, power, clocks) | `d6f490255412229f` |
| `tp8_c1_rep1.json` | arm repetition: aggregates + 1 record per request | `5afe1fa8760364d4` |
| `tp8_c1_rep1_gpu.csv` | nvidia-smi counters at 1 Hz across the 8 GPUs during the repetition (utilization, memory, power, clocks) | `ed3115f6dbec8b48` |
| `tp8_c1_rep2.json` | arm repetition: aggregates + 1 record per request | `d1e7de692002ce1e` |
| `tp8_c1_rep2_gpu.csv` | nvidia-smi counters at 1 Hz across the 8 GPUs during the repetition (utilization, memory, power, clocks) | `495cdcb6f8d52519` |
| `tp8_c1_rep3.json` | arm repetition: aggregates + 1 record per request | `8387822e0d378688` |
| `tp8_c1_rep3_gpu.csv` | nvidia-smi counters at 1 Hz across the 8 GPUs during the repetition (utilization, memory, power, clocks) | `2ffa961a76f44006` |