Independent benchmarks will determine whether llama.cpp's SYCL TILE-kernel dispatch materially accelerates long-context quantized-KV decoding on Intel Battlemage GPUs across representative models and contexts.

state: expiredheat: lowuncertainty: highknownscott: lowlocal-inference llama-cpp intel-gpullama.cpp Intel

What is this?

A llama.cpp pull request changes Intel SYCL dispatch for quantized key-value-cache decoding from a VEC kernel to a TILE kernel, with its author reporting up to 169% faster decoding at 118K-token context on Intel Battlemage hardware. The supplied results establish that llama.cpp’s SYCL backend targets Intel GPUs and can outperform its OpenCL backend, but they also document uneven SYCL performance and call for reproduction across GPUs, models, quantizations, drivers, and context lengths. The claimed gain is therefore a promising single-source optimization result, not yet evidence of a broadly reliable acceleration.

Why it matters to Scott

This is another hardware-specific llama.cpp optimization awaiting representative validation, a pattern already tracked in “llama.cpp,” “BeeLlama.cpp KVarN and low-bit KV-cache validation,” and several backend-kernel cases. It touches Scott’s hardware-aware local-inference and capability-audit work, but the supplied hits show his active substrate is NVIDIA/CUDA rather than Intel SYCL, so this result is unlikely to change what he builds unless independent tests establish broader portability or compelling Intel economics.

dev:concept.hardware-aware-local-inferenceip:concept.capability-auditradar:concept.llama-cppradar:person.llama-cppradar:concept.kv-cacheradar:concept.long-context-inferenceradar:beellama-kvarn-kv-cache-validation

queries asked of Scott's wikis

local inference benchmark methodology
long-context KV-cache quantization tradeoffs
llama.cpp backend optimization strategy
Intel GPU local inference economics
hardware-specific kernel dispatch portability
representative benchmarking for local models

Measured heat

no measured readings yet — the hourly heat pass fills this in

How the heat travelled

no chain yet — the hourly chain pass fills this in

Evidence (2) — ⭐ canonical anchor

source	object	author	score	comments
🟠 reddit	llama.cpp PR reports up to 169% faster quantized-KV decode at 118K context on Intel Battlemage from one SYCL kernel switch LocalLLaMA	BTA_Labs	102	20
🟧 echo.github ⭐	The primary source is PR #26689, titled “SYCL: TILE for quantized KV decode.” Its author reports that quantized-KV decode was using VEC, whi	johnkarlhill	—	—

Interpretation history

2026-08-13T14:39:05Z
The expected controlled VEC-versus-TILE replication has not emerged, and repeated checks have added only anecdotal broader-SYCL reports. This narrow validation episode has faded without resolving the kernel-specific claim.
2026-08-11T13:52:47Z
No controlled VEC-versus-TILE benchmark has appeared; the independent reports remain anecdotal evidence of broader SYCL improvement rather than validation of this kernel-specific claim. With only stale engagement and no substantive delta, the case should cool while awaiting reproducible comparisons.
2026-08-09T13:29:28Z
A second independent user now reports exercising recent SYCL changes on a B580 with a Gemma workload at roughly 49K context, broadening anecdotal evidence beyond the B70. The visible comment still lacks a controlled VEC-versus-TILE comparison, so the specific acceleration claim remains uncorroborated.
2026-08-07T20:29:27Z
An independent B70 user now reports running the change in a daily SYCL build and seeing continued performance improvement, making this more than author-only evidence. However, the comment lacks a controlled before/after benchmark, so it does not yet validate the claimed long-context gains or justify corroborated status.
2026-08-07T17:33:08Z
grounded: known/low — This is another hardware-specific llama.cpp optimization awaiting representative validation, a pattern already tracked in “llama.cpp,” “BeeLlama.cpp KVarN and l
2026-08-07T17:30:45Z
origin walked (codex/luna, conf 0.99): anchor reddit.post.1vi6hmw -> echo.github.e74fcf74a8 by johnkarlhill
2026-08-07T17:29:41Z
case created — The reported upstream kernel change has large and technically specific long-context gains that merit replication.

Decision trace

08-14 00:39expireThe expected controlled VEC-versus-TILE replication has not emerged, and repeated checks have added only anecdotal broader-SYCL reports. This narrow validation episode has faded without resolving the
08-14 00:39alert_silentThere is no new evidence or consequential event to surface; staleness alone does not justify an alert.
08-14 00:39alert_routeThere is no new evidence or consequential event to surface; staleness alone does not justify an alert.
08-11 23:52repriceNo controlled VEC-versus-TILE benchmark has appeared; the independent reports remain anecdotal evidence of broader SYCL improvement rather than validation of this kernel-specific claim. With only stal
08-11 23:52alert_silentNothing new establishes the claimed kernel-specific acceleration or changes Scott's decisions; engagement and elapsed time alone do not warrant an alert.
08-11 23:52alert_routeNothing new establishes the claimed kernel-specific acceleration or changes Scott's decisions; engagement and elapsed time alone do not warrant an alert.
08-10 01:21sensor_dirtyengagement_update
08-10 00:21sensor_dirtyengagement_update
08-09 23:29repriceA second independent user now reports exercising recent SYCL changes on a B580 with a Gemma workload at roughly 49K context, broadening anecdotal evidence beyond the B70. The visible comment still lac
08-09 23:29alert_silentThe refreshed discussion adds another real-world configuration but no controlled benchmark establishing the claimed kernel-specific gain; it can wait for a normal briefing or reproducible before/after
08-09 23:29alert_routeThe refreshed discussion adds another real-world configuration but no controlled benchmark establishing the claimed kernel-specific gain; it can wait for a normal briefing or reproducible before/after
08-09 23:21sensor_dirtycomment_update
08-09 05:21sensor_dirtyengagement_update
08-09 02:21sensor_dirtyengagement_update
08-08 22:21sensor_dirtyengagement_update
08-08 19:21sensor_dirtyengagement_update
08-08 15:21sensor_dirtyengagement_update
08-08 13:21sensor_dirtyengagement_update
08-08 09:21sensor_dirtyengagement_update
08-08 07:21sensor_dirtyengagement_update
08-08 06:29repriceAn independent B70 user now reports running the change in a daily SYCL build and seeing continued performance improvement, making this more than author-only evidence. However, the comment lacks a cont
08-08 06:29alert_silentThe new adopter testimony modestly strengthens the case but provides no detailed comparative results and does not change access, pricing, or Scott's near-term build decisions; it can wait for nor
08-08 06:29alert_routeThe new adopter testimony modestly strengthens the case but provides no detailed comparative results and does not change access, pricing, or Scott's near-term build decisions; it can wait for nor
08-08 06:21sensor_dirtyengagement_update
08-08 05:21sensor_dirtyengagement_update
08-08 04:21sensor_dirtycomment_update
08-08 03:40alert_silentAn open llama.cpp PR establishes a testable SYCL dispatch change and reports large Battlemage gains, but the results are author-reported, hardware-specific, unmerged, and outside Scott’s active NVIDIA
08-08 03:40surface_candidateAn open llama.cpp PR establishes a testable SYCL dispatch change and reports large Battlemage gains, but the results are author-reported, hardware-specific, unmerged, and outside Scott’s active NVIDIA
08-08 03:40alert_routeAn open llama.cpp PR establishes a testable SYCL dispatch change and reports large Battlemage gains, but the results are author-reported, hardware-specific, unmerged, and outside Scott’s active NVIDIA
08-08 03:33groundThis is another hardware-specific llama.cpp optimization awaiting representative validation, a pattern already tracked in “llama.cpp,” “BeeLlama.cpp KVarN and low-bit KV-cache validation,” and several
08-08 03:30promote_anchororigin walk conf 0.99
08-08 03:29createThe reported upstream kernel change has large and technically specific long-context gains that merit replication.