XIO

xWire

Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.

NVIDIA Developer Forums - Latest posts

Latest posts

GLM-5.3-Flash: 320B total parameters / 18B active

As I understand it, there’s no speed gain, or it’s small — the potential plus is only in parallel operation, but imho even 8 parallel instances, even working ones, in fact make real work impossible — so it’s really a downgrade (it seems) possibly the tool calling problem is solved, but I doubt it, because the problem is in the quant itself, not in the pipe

Is it worth buying a second DGX Spark now?

OllieO: And yes, answering title question: Yes, well worth it, despite pricing. Haha, that’s exactly what people ask me all the time, since I’m their go-to AI guy: “What should I get for private use?” Like anyone can afford this stuff :) 1-2 month ago, I would’ve said two Sparks. Now I’d definitely go for one Studio with 256GB. For a small IT team, I’d still go with 2–3 Sparks. That way, everyone can jump in and start getting things done right away (and again, big thanks to the Chineee labs for…

Intermittent EMBED_RUNAWAY with AR0235 TOP2 metadata on Jetson AGX Thor (R39.2.1)

Thanks for the detailed explanation. We checked your suggestions against saved callback buffers and collected a new streaming register snapshot. Configuration and streaming readbacks The baseline remains 1920×1200 RAW10, TOP2/bottom0, embedded stats off, with isEmbeddedDataTypeEnabled=true. VI descriptors show embed_x=3840, embed_y=2; metadata DMA stride is 3840 bytes. Device / registers Streaming readback AR0235 0x3064 / 0x327A / 0x306E 0x5940 / 0x0000 / 0x900F MAX96717 0x0313 / 0x031E /…

cusolverDnXsyevd returns info > 0 (no convergence) for an SPD matrix since cuSOLVER 12.2.6.9 (CUDA 13.3.1), regression vs 12.2.2.18

Hi all, After moving from CUDA 13.3.0 to 13.3.1 (cuSOLVER 12.2.2.18 → 12.2.6.9), cusolverDnXsyevd started returning info > 0 (info = 29, n = 256) for some symmetric positive definite matrices that it previously diagonalized without problems. For a given input matrix the failure is fully deterministic: the attached matrix_fail.bin fails on every call (1000/1000). However, it is extremely sensitive to the input. matrix_pass.bin differs from it in only 11 entries, by 1e-23 to 1e-17, and passes…

Prime k-tuplet sieve in CUDA on an RTX 5060 Ti: new record prime 18-tuplet

Hi, thank you! It’s my own software and still under active development, so I’d prefer not to go into the internals yet. In short: the GPU sieves to produce candidates, and the final primality checks are done separately. Most of the speed came from the mathematical side, i.e. from reducing which candidates need to be tested at all, rather than from exotic CUDA tricks. Maybe I’ll publish more details later. All test results on the my site

Senior Design Sponsorship Request – Jetson Orin Nano for Needle-Detecting Cleanup Rover (Drexel University)

Welcome to the Isaac ROS forum! Thank you for sharing your senior design project and your interest in NVIDIA Jetson. Unfortunately, we’re unable to fulfill hardware sponsorship requests. For purchasing options, please see the Jetson Orin Nano Super Developer Kit page. You and your faculty advisor can also review NVIDIA’s education-pricing information and ask an authorized distributor whether academic pricing is available for your institution and the requested kit. Good luck with your project!

Bit-reproducible MXFP4 GEMV on an RTX 4060 (sm_89): how to make it faster without changing the summation order?

arielstone883: after each CUDA update. If it is native data type and size it likely directly translates to SASS instructions 1-to-1 and won’t change with CUDA updates, which is, what you want. In addition to ensuring that the results are bit-identical, you can also simply check the produced SASS code of your sub-routine that it stayed the same.

GIGABYTE AI TOP ATOM (64GB DGX Spark) now available

Clustering 64Gb most probably wouldn’t make any sense. For single it would be enough for Qwen 3.8 27B and it will suck since it’s a dense model, but Strix Halo is around $3100 new and with gufo ( GitHub - gufo-org/gufo: The Open Source Strix Halo AI engine. Qwen 3.8 Flash Next Q4_K_XL: 1,801.01 pp, 73.17 tg single user, 194.37 tok/s 8 users; Qwen 3.8 27B Q4_K_XL: 656.33pp, 70.56tg tok/s single user with DFlash2 · GitHub ) it gives solid 60tps and >1000 prefill with Qwen 3.8 Flash Next 155B…

Is it worth buying a second DGX Spark now?

Side note. I used to be think that one needs full weights to run a model, GLM5.3F EXL3 changed my perspective. Starting with 1 spark running lower quant of 0731, it was dramatically worse than 0731 on 2x sparks I upgraded to after a week. But GLM5.3F EXL3 on 2x sparks (with 3M context) wasn’t any worse. I did not expect that, but I adapted to it well during last 4 weeks. This doesn’t change main discussion here though since 256Gb Mac would be probably same good (maybe will less parallelism)…

Guidance Required: Multi-Sensor Synchronization and Sensor Fusion on Jetson Orin NX (JetPack 7.2.1)

Thanks for the dumps. Orin NX 16 GB is fine for a first DeepStream demo. The Ethernet/NVPPS picture is different from a Tegra MAC PTP grandmaster. enx9c69d36176ee / enx9c69d3618c27 (ax_usb_nic, /dev/ptp0 /dev/ptp1) are the only PHCs. enP8p1s0 (r8168) reports PTP Hardware Clock: none — do not use it for Ouster. NVPPS primary-emac must be a Tegra EQOS/MGBE DT node (reg-names = "mac"). A USB NIC cannot be that phandle, so /dev/nvpps0 is missing even though nvpps.ko is installed, and TSC cannot be…

Regression in v8.9.0 for UE5.8: DLSS-RR not working in MRQ with Start Console Commands

Hello, I would like to share some additional findings regarding the use of DLSS-RR on MRQ and MRG (I later checked the DLSS documentation, which states that DLSS-RR does not support MRG for the time being; nevertheless, I did some experimentation). I’m attempting to use this in MRG because I stumbled upon this feature. In MRQ, once I add the “DLAA/DLSS” option in the settings, the DLSS Denoiser gets forced to 0 — meaning DLSS-RR is disabled — regardless of the CVAR setting. The DLSS Denoise…

620 release feedback & discussion

Since I can’t uploaded, here is the difference:A = /home/fran/nvidia-hdmi-evidence/20261106-000008-26522 B = /home/fran/nvidia-hdmi-evidence/20261105-234355-39947 == loaded kernel module version == /home/fran/nvidia-hdmi-evidence/20261106-000008-26522 615.71.10 /home/fran/nvidia-hdmi-evidence/20261105-234355-39947 620.70.06 == on-screen notes == — /home/fran/nvidia-hdmi-evidence/20261106-000008-26522 /home/fran/nvidia-hdmi-evidence/20261106-000008-26522/00-meta.txt:note (on-screen observation):…

620 release feedback & discussion

# 620.70.06: onboard HDMI fails on Acer PHN16-73 / RTX 5070 Ti Laptop; 615.71.10 works ## System - Acer Predator Helios Neo 16 AI PHN16-73, BIOS V1.28 (2026-07-20). - NVIDIA RTX 5070 Ti Laptop / GB205M (`10de:2f58`), Intel Arrow Lake-S (`8086:7d67`), hybrid graphics. - Ubuntu 26.04.1, GNOME Shell 50.1, Wayland. - NVIDIA CUDA repository packages: `nvidia-driver-open` / `nvidia-dkms-open` 620.70.06-1ubuntu1; last known good 615.71.10-1ubuntu1. - Original driver A/B: Ubuntu `7.0.0-42-generic`.…

[Jetson AGX Thor][JP7.1/R38.4.0] RT5640 AUD_MCLK shutdown glitch at end of playback — request for official BPMP/kernel patch

Hello NVIDIA team, We would like to report an audio clock issue on Jetson AGX Thor and ask for an official fix. Environment Board: p3971-0080 + p3834-0008 (Jetson AGX Thor) JetPack 7.1 / L4T R38.4.0 Kernel: 6.8.12-rt-tegra Codec: onboard RT5640, connected to I2S1 DTS: clocks = <&bpmp TEGRA264_CLK_AUD_MCLK> clock-names = “mclk” mclk-fs = <256> aud_mclk parent = plla1_out1 Symptom At the end of playback, when the RT5640 driver drops BIAS_ON → PREPARE, rt5640_set_bias_level() calls…

[r36.4.4] vi5_fops.c: wrong address freed in tegra_channel_capture_setup() error path → Bad page state → kernel panic

hello LeeJ, LeeJ: Physically disconnect all 6 cameras (remove the camera modules from the SerDes links). honestly, we only test error recovery mechanism with the software-simulated stream loss, we’ve never test with singe camera or even six cameras being physically disconnected scenario. may I have more details for repeated concurrent STREAMOFF/STREAMON retries while all camera modules remain physically disconnected? what’s the actual use-case? as a temporary workaround, you should check…

Unaddressed issue in the Isaac Ros repo

Hello @olly4, Welcome to the Isaac ROS forum and thanks for the post! Both GitHub issues and the forum are valid reporting channels. Regarding your installation failure, Isaac ROS 5.0 currently supports RealSense cameras only in Docker mode. Virtual Environment and Bare Metal modes are unsupported, as documented in the release notes. Please follow the Docker tab of the RealSense setup instructions. The setup guide still includes Virtual Environment instructions, which can be confusing given…

Need to know at what speed camera i2c frequency Driver Thor SDK7.2.5 is running

Hi, We are running AR0341 sensor on our Driver platform SDK7.2.5 using nvsipl_camera , where we observed the first frame is coming at a time of 1.95sec, once we launched nvsipl_camera. We need to know where by default at what frequency camera i2c is getting configured and how to verify it?? If we need to change the frequency to high speed say 1MHz then what all changes are required and under which file? Is there any direct way of updating the i2c frequency to high speed so that nvsipl_camera…

NVKE on GB10s

I recently discovered GitHub - dsx-ai-factory/nvke: NVIDIA Kubernetes Engine (NVKE). A management plane for GPU Kubernetes on bare metal. · GitHub and was wondering if there’s a way to use it on multiple Sparks. The memory requirements in the guide at nvke/docs/get-started/local-site.md at main · dsx-ai-factory/nvke · GitHub are quite high, but given this would be a cluster of 2-4 GB10s, maybe it’s not that bad. Any thoughts because I dive in a rabbit whole for 2 days?