Explore the public B70 benchmark catalog. Compare matching workloads and follow the proof link before choosing a stack.
P = prompt tokens · G = generated tokens · C = concurrent requests · n = samples. Different metrics and profiles are not a single speed ranking.
Scroll the table sideways to compare throughput and open proof links →
Loading the public benchmark catalog…
Loading approved community results…
Pending submissions never appear here.
The measurement rules behind every record on this page. Every speed comes with its hardware, profile and sample count.
Lane 1 Model Card: 5 anchor cells, n=5 samples, median ± range
Lane 0 smoke check, Lane 1 five-cell model card (p512/g128, p8192/g128, cold input anchors), Lane 2 speculation screens, Lane 3 full context and concurrency characterization.
≥64 chars unique prompt entropy mandatory on cold runs
Shared prompt templates let prefix caching accelerate repetitions 5×. Cold runs inject entropy and require prefix_cache_hits delta of zero.
Cold input rate = actual prompt tokens / client TTFT
C1 isolates single-user latency. Cn is total tokens over wall time. Per-request rates are never summed into aggregate throughput.
MoE 35B self-limits to ~140 W · 150 W cap saves 80 W of heat
Hardware energy-counter deltas over steady state. MoE sweet spot 150 W, dense 180 W. Cooldown to ≤55 °C before matched A/B sweeps.
The xe driver rejects per-message IPC exchange on hosts without P2P. The validated fix is spawn-time worker affinity plus oneCCL simple-threshold variables.
Prefix caching makes repeated prompts look 5× faster. Check the zero-delta rule and cold anchors; every record here lists its cache policy and sample count.
Rules on this page ↑Sparse MoE decode reads ~3 GB of experts per token. A 150 W cap matches 230 W throughput with 80 W less heat.
Measurements →AutoRound INT4 beat GPTQ conversationally on Qwen3.6; GPTQ stacks MoE errors. Measured per model in the table above.
Quality data →Compare coordinates: P/G/C, samples, engine version, driver. Records without them are listed as capabilities, not ranked.
Rules on this page ↑Run the official client or send a model output below. Reviewed community runs join the tables after review.
Submit ↓Private queue. Only the site owner can read what you send.
Requirements: Python 3.10+ and a local OpenAI-compatible endpoint. The runner measures vLLM, SGLang, or llama.cpp servers at C1, C2, C4, C8, C16, and C32 with streaming SSE and saves xecores-results.json locally first.
curl -fsSLo xecores-runner.py https://xecores.com/xecores-runner.py
python3 xecores-runner.py --engine vllm \
--endpoint http://127.0.0.1:8000/v1 \
--model Qwen3-32B-AutoRound-INT4 \
--quantization autoround-int4 \
--dtype int4 \
--concurrency c1,c2,c4,c8,c16,c32
Use the runner source or a pinned cookbook release. Check any published checksum before running a release download; this site does not publish one for the current source file.
Send one prompt and the exact response a model produced on your Intel GPU. You preview your submission before anything leaves your browser. It lands in a private queue only the site owner can see. Submitted code is rendered as text and never executed.
GET /api/v1/challenge (runner UA only) · POST /api/v1/results one concurrency profile, 256 KiB cap, HTTP 202 pending · GET /api/v1/results approved projections only, filter by engine, GPU, model, concurrency. Schema: result.schema.json.
POST /api/v1/model-outputs/challenge browser-issued one-time token · POST /api/v1/model-outputs consented artifact, 128 KiB cap, HTTP 202 pending · GET /api/v1/model-outputs owner-only listing, requires review token. Schema: model-output.schema.json.