Upstream issue histories and Intel support articles reviewed September 10, 2026. All cases below are external reports; none was reproduced locally for this guide. Closed issues can have unverified outcomes.
Find the closest symptom
Start with the affected environment, then read the outcome. A similar symptom on a different build is a new observation until a reproduction connects them.
| Symptom | Reported scope | Status at September 10 |
|---|---|---|
| Crash after runtime upgrade | B70, Linux, compute-runtime 26.27 | Reporter-confirmed resolution |
| Long prompts suddenly slower | B70, Ubuntu SYCL, llama.cpp b10216 | Later build reported restored |
| MoE output corruption | B70, Windows 11, SYCL operation tests | Closed; fix unverified |
| Less memory than expected | Windows 11, Intel Graphics Software, ECC | Intel support explanation |
| DisplayPort has no image | Specific B70 port/display/Linux combination | Intel investigating |
These statuses describe source evidence. “Reporter-confirmed” is not a guarantee for every B70; “unverified” does not mean every card has the problem.
Crash after a runtime upgrade
Report: compute-runtime #972, opened August 15 and closed completed August 21, 2026. The original symptom followed a move from runtime 26.22.38646.6 to 26.27.39122.12.
Outcome: a maintainer identified a stale Level Zero loader. The reporter’s retest used loader 1.32.0, runtime 26.27.39122.14, Ubuntu 26.04/kernel 7.0.0-30, and llama.cpp b10546, with no crashes across 40 executions.
Useful next step: capture both loader and runtime versions, including the libraries the application actually loads. Compare them with your selected stack’s requirements. The original downgrade suggestion should not become permanent installation advice. The retest changed several versions, so it does not isolate one change as a universally sufficient fix.
Long-context prefill slows down
Report: llama.cpp #26396, August 1–4, 2026. Ubuntu SYCL FP16 binaries on B70, runtime 26.27.39122.11, Qwen3.6-35B-A3B GGUF, F16 KV and MTP. b10216 regressed against b10202 at roughly 200K prompt tokens; short microbenchmarks did not expose it.
Outcome: the maintainer corrected the workaround variable to GGML_SYCL_ENABLE_MKL_FA=0. The reporter later said b10256’s default matched the workaround. The issue was closed completed.
Useful next step: measure a real prompt at the length you need, retaining actual token counts and cache behavior. Do not carry this old toggle into every new environment. A successful 32K test would also not reproduce or clear a report about 200K inputs.
MoE prefill produces corrupt output
Report: llama.cpp #25455, opened July 8, 2026. The reporter described B70/Qwen3.6 output corruption and failing MUL_MAT_ID cases on Windows 11 with oneAPI 2026.1.0 and Level Zero driver 1.14.36605.
Outcome: closed by the stale bot on August 30 with reason not_planned. Closure supplies no verified fix. A related decode change is not proof that the prefill report was resolved.
Useful next step: preserve failing operation shapes and numeric comparisons, then test actual model responses. A Linux pass would be useful new evidence for that Linux configuration, not resolution of the original Windows report. Fluent text alone cannot establish that every backend operation is numerically correct.
Why does the card show less memory?
Intel’s June 8 support article describes 28 GB appearing instead of 32 GB on a B70 with ECC enabled, in Intel Graphics Software on Windows 11.
For your own observation, record the reporting tool, units, ECC state, and memory already used by other work. The ECC capacity tradeoff is a useful explanation for that documented case, not a reason to disable error protection by default or an explanation for every allocation failure. Model weights also need room for working buffers and cache; see the capacity guide.
DisplayPort 2.1 has no image
Intel’s August 7 support article describes a third-port problem with DP 2.1 that did not reproduce with DP 1.4. It remained under investigation at this review.
The description names Ubuntu 26.04 while the article’s environment metadata names older Ubuntu versions. Preserve that uncertainty. Record the exact board, port, cable, display mode, kernel and graphics stack when comparing your case; this report does not establish a universal B70 defect.
Capture evidence that helps someone else
Keep failures as well as passes. A useful report includes:
- A small input or operation that triggers the symptom, with expected and actual output.
- The software-stack record, including the application revision and local changes.
- Whether the problem appears on the first request, a later request, or only under concurrency or long input.
- Raw logs, actual prompt/output counts, and per-device memory where available.
- Exactly what changed between the failing and passing runs.
nvtop supports Intel Xe for monitoring. Available metrics depend on the driver and permissions; an absent reading is unknown, not zero. GPU telemetry does not measure whole-system wall power.
Use the relevant upstream issue tracker for software failures. Intel’s B70 support hub and board support routing help with hardware-specific cases. For a correction to this log, include the source comment or release that changes the outcome.