HighVRAM Hardware Watch: cited benchmark Snapshot 3
https://highvramhardwarewatch.com/benchmark-database/

43 externally reported metric observations from seven cited source publications.
Snapshot identifier: cited-2aca936bd67d79e8
Export/data-contract version: 0.2.0
New-record contract: highvram.metric-extension.v1

CONTENTS
benchmarks.json: structured records and source-specific metric definitions.
benchmarks.csv: flat columns and a lossless record_json column.
benchmarks.sqlite3: benchmark_results and dataset_metadata tables.
README.txt: this guide. This archive contains data only, with no importer,
software bundle, private notes or full copies of third-party articles.

WHAT IS COUNTED
13 prompt-throughput observations (four are serving-level prompt/TTFT proxies),
18 generation/decode observations, four streamed TTFT observations, four combined
input + output throughput observations, and four source-labeled GPU memory MB
observations. There are no populated unclassified-throughput or confirmed
peak-memory observations. Independent execution count is unknown. Two or more
metric observations can describe one request group or source profile; 43
observations do not mean 43 independent runs.

Snapshot 3 preserves every record object from Snapshot 2 and adds 13 observations
from two sources. The additions describe nine source profiles in three
configurations: one Qwen/H20 configuration and two separate two-RTX-3090 KV
regimes. These counts are not additional independent runs or hardware model
counts. hardware_label_count counts exact hardware.name strings, not an inferred
hardware-family taxonomy. added_profile_count and added_configuration_count
refer only to the 13 observations added in this snapshot.

KEY METRIC LIMITS
QwenLM/Qwen3 is the author of the H20 table; NVIDIA is the hardware vendor, not
the benchmark author. The four selected Qwen3-32B AWQ-INT4 profiles each have a
combined input + output throughput observation and a separate reported-memory
observation. Combined throughput is (prompt tokens + generation tokens) / time;
the precise timing boundary is unreported. It is not decode throughput or
phase-unspecified throughput. Its increase with input length does not establish
faster generation. The GPU memory column is labeled MB; the introduction says
GB. No conversion is made. Decimal/binary convention, measurement boundary and
peak status are all unknown. This is not confirmed peak memory.

The five chibop1 two-RTX-3090 observations use the source's Generation Speed
column, confirmed by the author as tokens/second. Exact timing boundaries are
unreported. Outputs vary. Earlier KV format is unspecified; the 20,234- and
32,170-prompt-token profiles use q8_0. These two KV regimes have separate
configuration IDs. Prompt counts are not configured context capacity. Do not
infer controlled context scaling, TTFT, prompt throughput, execution repetition
counts or a cross-hardware ranking from the selected generation observations.

Rapid-MLX's prompt/TTFT proxy includes serving/request overhead and is not isolated
engine prefill. Corrected decode is (completion tokens - 1) / (total time - TTFT).
Its prompt, decode and TTFT medians describe the same three requests per profile.
The historical llama-bench pp512 and tg128 rows describe separate microbenchmarks,
not a combined 512-token prompt and 128-token response. Read each source's full
record notes before comparing values.

PROVENANCE AND UNKNOWN VALUES
HighVRAM has not run or reproduced any of these hardware tests. Contract and
arithmetic checks do not verify execution. Different models, runtime builds,
workload shapes and timing definitions prevent an overall hardware ranking.
Source publication and edit/revision dates are not execution dates. The original
30 records retain their October 8, 2026 access date; the 13 additions have an
October 9, 2026 access date. Missing exact access times are not invented.

Qwen's source file was first published May 16, 2025; the pinned revision is dated
May 20, 2025. The Reddit post was published December 14, 2024 at 13:28:20 UTC and
edited December 16, 2024 at 04:47:20 UTC. Execution dates are unknown. Edit metadata
does not establish which text was present at initial publication.

JSON null, SQL NULL and empty numeric CSV cells mean unreported or not established,
never zero. Old JSON records retain their original five metric keys. Absence of
the two new keys means unknown/null; use .get(key) or an equivalent read adapter.
New records include all seven metric keys with exactly one populated metric.
The added fields are combined_input_output_tokens_per_second and
reported_gpu_memory_mb. Memory remains in the source's MB unit. No MB/GB/GiB
conversion or normalization is performed.

COMPATIBILITY AND PRESERVATION
All 30 old JSON record objects, IDs, citations, notes and absent keys are preserved.
All old SQLite column values, including original record_json, record_sha256 and
comparison_json, are preserved. Legacy SQLite record_json representations differ
from the corresponding JSON-export objects; neither is rewritten to match the
other. Old CSV cells in original columns are preserved. Expanded files have new
bytes, but existing record facts remain unchanged. CSV includes record_json;
strings that could be interpreted as spreadsheet formulas are escaped.
Source definitions and dataset-level metadata are included in JSON and SQLite.
The newly prepared standalone migration was checked locally; integration with
the original importer, which was unavailable, is not established by this export.

SOURCES
Each record links to its original source and includes a row/section locator.
Pinned GitHub URLs retain the cited revision; the Reddit source can change.
Read the original sources and retain their source-specific comparison limits.

PRIOR SNAPSHOT
Snapshot 2 (30 observations), identifier cited-4d82916d23e704c3:
https://highvramhardwarewatch.com/wp-content/uploads/2026/10/highvram-cited-benchmarks-20261008-v2.zip
SHA-256: 341ec57c913f269fa57fc683a56eced16f47cc6b3ea8015514601fbe706986d9
