context
4
ctx:discord/blah/hardware/4Source document
full texthardware-4
text/plain2 KB
doc:agent/hardware-4/6e16f9ff-009c-43a5-8570-f411f1917f0c[2026-04-28 13:06] xenonfun: https://github.com/parallelArchitect/sparkview [2026-04-29 01:26] xenonfun: (files: Screenshot_2026-04-28_at_9.25.54_PM.png) [2026-04-29 13:20] xenonfun: yeah very easy to OOM hard, TCP stack still works but not clever enough to figure out way to unfuck beyond the hit the physical power button. (files: Screenshot_2026-04-29_at_8.59.55_AM.png) [2026-04-29 14:03] xenonfun: Nemotron-3-Nano-Omni looks like 25tok/s having it now mess with stress tests at concurrency. claimed 41 sessions at 128K context will fit, will see how performance looks. (files: Screenshot_2026-04-29_at_10.00.22_AM.png) [2026-04-29 14:09] xenonfun: (files: Screenshot_2026-04-29_at_10.04.49_AM.png, Screenshot_2026-04-29_at_10.09.38_AM.png) [2026-04-29 14:24] xenonfun: (files: Screenshot_2026-04-29_at_10.15.30_AM.png, Screenshot_2026-04-29_at_10.22.54_AM.png, Screenshot_2026-04-29_at_10.23.49_AM.png, Screenshot_2026-04-29_at_10.24.31_AM.png) [2026-04-29 15:15] xenonfun: ~6s. tho the model itself cannot seem to figure out how to script that request/encoding properly. (files: Screenshot_2026-04-29_at_11.12.48_AM.png) [2026-04-29 15:42] xenonfun: Watching youtube (files: Screenshot_2026-04-29_at_11.41.20_AM.png) [2026-04-29 17:16] xenonfun: https://forums.developer.nvidia.com/t/dgx-spark-software-updates-04-2026/368114 [2026-04-29 22:03] ajaxdavis: lol so relevant i was trying to get the transcription for this video right now https://www.youtube.com/watch?v=NpPySD-3tl0 how long would it take for you to do for me [2026-04-29 22:04] ajaxdavis: lol nvm i will just cut download it and use traditional transcription [2026-04-29 22:04] ajaxdavis: how does the youtube watching work? just looking at frames at intervals? [2026-04-29 22:18] xenonfun: it pulls it to a file with standard youtube dl, and then base64s it in. I need to see what limits were, think audio was 1hr of mp3. obvious was >2m on that video i tested. tested 8bit on the M3 get like 80tok/s vs. 25 tho not sure on concurrency but 1 session blazing. was bf16 on the dgx which is suboptimal [2026-04-29 22:21] xenonfun: ``` Video: mp4, up to 2 minutes. For 1080p videos, sample up to 1 FPS / 128 frames. For lower-resolution videos such as 720p, higher temporal sampling such as 2 FPS / 256 frames may be used. ``` think you can pull transcription from youtube itself and request it be transcribed [2026-05-01 09:55] xenonfun: https://x.com/i/status/2050145704740577371
Facts in this context
Grouped by subject. Each subject links to its full article.
Log Entry 2026 04 29 22 1817 factsex:log-entry-2026-04-29-22-18
| assessesOptimization | Dgx System |
| comparesMetric | 25 |
| describesMechanism | Youtube Watching |
| hasAuthor | Xenonfun |
| hasTimestamp | 2026-04-29 22:18 |
| metricMode | bf16 |
| metricMode | 8bit |
| metricSubject | M3 Device |
| metricSubject | Dgx System |
| quotesSpec | Video: mp4, up to 2 minutes. For 1080p videos, sample up to 1 FPS / 128 frames. For lower-resolution videos such as 720p, higher temporal sampling such as 2 FPS / 256 frames may be used. |
| reportsConcurrency | 1 |
| reportsLimit | 1hr of mp3 |
| reportsMetric | 80 |