Questions and Answers : Getting started & support : Macos M1 - keraunos GPU task only consumes about 25% of GPU (as reported by SiliconScope app)
Message board moderation
| Author | Message |
|---|---|
rilianNew member Send message Joined: 28 Sep 26 Posts: 78 Credit: 5,244,405 RAC: 429,737 |
|
rilianNew member Send message Joined: 28 Sep 26 Posts: 78 Credit: 5,244,405 RAC: 429,737 |
sorry seems like it is a different issue. approximately at the same time i looked, this task just errored https://bitboinc.athena.org.tr/result.php?resultid=1268366 maybe that is why GPU consumption was low. New task started 15 mins ago and it consumes 100% GPU You were fixing such issue yesterday and it seems like issue is not fully fixed yet <core_client_version>8.2.10</core_client_version> <![CDATA[ <message> process exited with code 1 (0x1, -255)</message> <stderr_txt> Keraunos Metal (bounded discrete-log search, Apple GPU) v1 bitboinc -- https://bitboinc.athena.org.tr/ coded by Alperen Yavuz kangaroo_run_metal: using Metal device "Apple M1" kangaroo_run_metal: command buffer error: Internal Error (0000000e:Internal Error) kangaroo_run_metal failed (device 0) -- no usable Metal device? 2026-10-01 15:15:37 (84701): called boinc_finish(1) </stderr_txt> ]]> I crunch for Ukraine
|
rilianNew member Send message Joined: 28 Sep 26 Posts: 78 Credit: 5,244,405 RAC: 429,737 |
|
|
New member Send message Joined: 29 Sep 26 Posts: 22 Credit: 0 RAC: 0 |
This was actually fixed in source already -- the real cause was a scoping bug: every GPU round created a fresh MTLCommandBuffer/MTLComputeCommandEncoder, but the only autoreleasepool in the whole function wrapped the entire multi-hour run, so for a long task those objects (each holding real GPU driver resources) just piled up unreleased until Metal ran out and threw that "Internal Error". Fixed by scoping an autoreleasepool to each individual round instead. What wasn't settled was whether that fix had actually made it into the binary you were running, since Mac GPU builds were still labeled under the old "opencl_apple_gpu" plan class at the time even though they're genuinely Metal. Rebuilt and redeployed tonight under a properly separate "metal_apple_gpu" plan class, from the current source (fix included), so it's live now regardless of what was true before. Should be safe to turn GPU tasks back on if you want to test it. |
rilianNew member Send message Joined: 28 Sep 26 Posts: 78 Credit: 5,244,405 RAC: 429,737 |
|