620 release feedback & discussion
Same vkCreateDevice hang on RTX 5060 (desktop) with 620.70.06 on Debian 13. Details and bug report here: Crash on vkCreateDevice
Posts from every feed QUEST has found across the web — blogs, news, podcasts, channels and more.
Same vkCreateDevice hang on RTX 5060 (desktop) with 620.70.06 on Debian 13. Details and bug report here: Crash on vkCreateDevice
Small sample curse, same issue in llama bench with defaults most use. Too small generation tokens, not precise measurements on small context and generation sample. All I am saying numbers in general are much better than ones you quoted, so take it as you want. I have numbers on my m5 max for qwen 3.8 fn from 60 t/s to 160 t/s depending on stack, plug-ins, version, parameters, int8 optimization, quant configuration, tensor padding method. I can’t speak for spark ring on 4 sparks but I feel like…
fdg270301: Title: vkCreateDevice Exception 0xc0000005 on RTX 3050 Laptop with driver 620.70.06 (Proton) Confirming the same vkCreateDevice issue on desktop RTX 5060 with driver 620.70.06. System: Distro: Debian 13 Driver: 620.70.06 (nvidia-open) GPU: NVIDIA GeForce RTX 5060 (desktop) Kernel: 6.12.117+deb13-amd64 (also tested 6.12.116, same failure) Problem: Teardown (Proton) hangs indefinitely at the loading screen. The process stays alive and allocates VRAM, but never reaches the main menu.…

I got the original Leap Motion Controller tracking hands in 3D on a Jetson AGX Xavier (JetPack 5.1, TensorRT 8.5). The controller is read as a plain UVC camera, and Monado’s Mercury hand models run as FP16 TensorRT engines. It produces 21 joints per hand at about 80 fps, 14.5 ms from camera frame to 3D. The biggest single gain on the Xavier was the GPU clock: the governor stayed at about 217 MHz for this bursty load, and raising min_freq to 1377 MHz took tracking from 27 to 67.5 fps. FP16…
Hello NVIDIA Team, I am trying to generate an API key on ://nvidia.com, but the system prompts me to contact support to verify my account. Could a moderator please help review and manually verify my account so I can start testing the APIs? Registered email: cauagraca51@gmail.com Country/Region: Cape Verd Intended Use: Personal learning. Please let me know if I should send my account email via Direct Message (DM) to complete the verification. Thank you in advance for your help!
We run 2; it’s night and day from running just one. Ideally, for dev you need 4.
Long term sure. I guess ccp has it suck it up attitude. I really don’t care if she’s ccp spy, if anything more likely she is alibaba agent of influence, not classic operative. But really it changes nothing. For us more important issue if the web of circular financing and price fixing untangle, alleged one of course. Nvidia stock p/e is extremely reasonable, one of the best in ai/tech sector. But it rides on high prices and rising demand. If prices are fixed and demand manufactured a more…
620.70.06 RTX 5080: Flip event timeout on two heads, display engine hang – regression from 595.112.07 System GPU: GeForce RTX 5080 (GB203), single GPU driving all displays; Intel iGPU present (i915 loaded) Driver: 620.70.06, open kernel modules (openSUSE package nvidia-open-driver-G07-signed-kmp-default 620.70.06_k7.3.3_1) Previous working driver: 595.112.07 (open) Distribution: openSUSE Tumbleweed 20261111 Kernel: 7.3.3-1-default Desktop/session: KDE Plasma 6.x, Wayland Displays: iiyama…
Dear @liuq1 , H265 encoding is supported using NVMedia APIs. Please check NvMedia IEP – Encode Processing (nvm_iep_sci) — NVIDIA DriveOS 7.2.5 Linux SDK Early Access Developer Guide. If you are looking to use DW, only H264 encoding is supported.
Have a look at this PR; also have a look at this MR which is about getting tearless Gnome/mutter working on tiled displays.
You summed it up perfectly. Thanks for showing me how many keystrokes I wasted :D
Different users have differing expextations of the kit. Some buy it, and because of the high cost, expect einstein on their desk. Others buy it, recognise its actual capabilities and work with the rest of the community to accomplish them. Which was Einstein?
Regression in 620.70.06: full screen video causes a black screen, then a heavy red tint on one monitor (worked fine on 615) System Driver: 620.70.06 (open kernel module). 615.71.10 does not have this problem. OS: Debian kernel 7.2.15 Compositor: Hyprland 0.57.0 built from branch v0.57.0 at commit 0aa6745f705f2c4b0a5004eb019ef76e5bbdb426 clean (VERSION: bump to 0.57.0), Wayland Monitors: 43" HDR-capable (affected) 27" rotated to portrait (not affected) 28" rotated to portrait (not affected)…
The maturity argument is a really strange one right now. imho, it is a mature system. The caveat of a mature, responsible and empirically sound qa, benchmark before production and update rollout schedule is up to the user. In an enterprise production environment, devops management isn’t likely going to update a day 1 patch without waiting for the first hot-fixes to be released (unless you’re php… alpine… ssh…). The risk doesn’t have to be enormous for downtime with enough lead time, but even…
Hello, I’m on a free NVIDIA build developer account. I think my organization is missing the “Public API Endpoints” permission, the same issue reported in the other cosmos3-nano 404 threads. What works: GET https://integrate.api.nvidia.com/v1/models with my API key returns 200 and 84 models. The cosmos3-nano Experience page works in the browser. What fails: Every POST to https://ai.api.nvidia.com/v1/cosmos/nvidia/cosmos3-nano returns a plain-text “404 page not found”. This happens even with the…
Me gustaría poder generar vídeo con su modelo cosmos3 Nano pero el enpoint está bloqueado da error 404
This is really for Nvidia. I am planning on using this RTX Spark as a daily driver, but Zoom does not detect it as a valid chipset that supports virtual backgrounds. Nvidia, I would love for you to connect with Zoom through your resources and get this resolved. Hard to use this device as an all in one when things like that don’t work and my ticket will have no movement.
Had the exact same problem yesterday. But instead of crashing during Steam launch, it happened after I played a game for 1+ hours, then tried to close Steam.
That’s easy to understand: apple produce lot more devices than nvidia does, have massive pricing muscle and long term fixed price contacts. Nvidia mostly produces chips for oem to put on boards and strap with own ram chips. Hence different pricing power than apple.
Why do you use the legacy Maxwell-Pascal driver for a Blackwell card? Have you tried v595 or v620 instead?
Hello! I’m an individual developer calling the NVIDIA NIM API (https://integrate.api.nvidia.com/v1). My API key is valid: GET /v1/models returns 200. But POST /v1/chat/completions returns 403 for all models: {“status”:403,“title”:“Forbidden”,“detail”:“Authorization failed”}. My organization is Personal, and I cannot select the Public API Endpoints permission when generating an API key. I would like to request enabling the Public API Endpoints permission for my Personal organization, for…
I’m == OS == PRETTY_NAME=“Ubuntu 24.04.5 LTS” VERSION_ID=“24.04” VERSION_CODENAME=noble DGX_NAME=“DGX Spark” DGX_PRETTY_NAME=“NVIDIA DGX Spark” DGX_SWBUILD_VERSION=“7.5.0” DGX_OTA_VERSION=“7.6.0” Linux 7.0.0-1024-nvidia aarch64 GNU/Linux Static hostname: spark2 Icon name: computer-server Chassis: server 🖳 Machine ID: 482288dd556b4387862a9cfd993bbe20 Boot ID: ecce7d8c24ad4b31ab1f39f4046574fc Operating System: Ubuntu 24.04.5 LTS == Board / firmware == NVIDIA_DGX_Spark Manufacturer: NVIDIA…

@giles8 thanks, I’d love the help. Honest status first: TP2 isn’t possible yet. The piece that would make it possible is in PR #981: an expert cache that keeps part of each layer’s experts in memory and reads the rest from NVMe when the router asks for them, with bit-identical outputs in its tests. So far it has only run as a kernel and cache benchmark on one Spark. There’s no two-node profile yet, and nothing has served a request at TP2. Here’s the math as I see it. Measured numbers come from…
Hope the above suggestions are helpful to your issues. Feel free to open new topics if you still need help with the latest Isaac Sim release. I am closing this topic.
Greetings NVIDIA Deep Learning Institute Team, We are Team MechaDucks (FRC Team 12364), a team of 30 students from Delhi India participating in the FIRST Robotics Competition (FRC). We are preparing for our international FRC competition journey to Turkey. Alongside our competition activities, we are starting a student outreach initiative to make technology and practical learning accessible to more school students. Our initiative focuses on students from Classes 8–12 through interactive 4–5-hour…

Using the latest 620 driver with cachyOS kernel 7.3.3 and when trying to launch steam entire system refreezes needing forced shutdown. Reverted back to nvidia version 615 and steam works fine.
Jetson T4000 uses the same hardware resources as Jetson T5000. The Jetson Thor documents and design files in the Jetson Download Center apply to both Jetson T4000 and Jetson T5000. To check the differences between Jetson T4000 and Jetson T5000, please refer to the Jetson Thor Series Modules Data Sheet. For the detailed specifications of the Developer Kit carrier board, please refer to the Jetson Thor Module Carrier Board Product Specification.
I have an OpenClaw instance running on bare metal; how do I migrate to NemoClaw without losing anything?
If you update your repo, for TP=2, i would love to give it a validation phase. What context do you think you would get with 2 GB10s?
Adding my support here as well. This is another RTX 50-series report of severe black-screen behavior on 621.14, on different hardware from both the original poster and my own setup. I also want to emphasize that this does not appear to be a brand-new 621.14 issue. Similar display-related problems have been reported across multiple recent NVIDIA driver releases, and in my own case the regression began with 616.92 several months ago and has persisted across later drivers. There are also multiple…