595.58.03 Certified Linux-aarch64 (ARM64) Display Driver and CUDA 13.2 - when for DGX Spark GB10

Accelerated Computing DGX Spark / GB10 User Forum DGX Spark / GB10
,
1

When for DGX Spark GB10, please?

Many thanks!

2

Appendix A. Supported NVIDIA GPU Products

image
image993Ă—540 62 KB

3

is this fully supported tho?
i don’t see gb10 under supported devices in op’s link

4

Driver 595 is not yet supported on DGX Spark

5

@aniculescu , then what was the purpose of the June Software release?

6

Just received notice there’s new firmware form my GB10. Any chance this allows the 595+ drivers?

7

why
why1034Ă—491 17.1 KB

8

13.2 had some issues, so maybe we will jumb strait to 13.3 … August?

9

Already running on Ubuntu 26.04 LTS — driver 595.71.05, CUDA 13.2, open kernel module, GB10 detected, with vLLM serving from a CUDA 13.2 container.

$ . /etc/os-release; echo "$PRETTY_NAME"
Ubuntu 26.04 LTS

$ uname -m -r
aarch64 7.0.0-1006-nvidia

$ cat /proc/driver/nvidia/version
NVRM version: NVIDIA UNIX Open Kernel Module for aarch64  595.71.05  Release Build
GCC version:  gcc version 15.2.0 (Ubuntu 15.2.0-16ubuntu1)

$ nvidia-smi
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 595.71.05              Driver Version: 595.71.05      CUDA Version: 13.2     |
+-----------------------------------------+------------------------+----------------------+
|   0  NVIDIA GB10                    On  |   0000000F:01:00.0 Off |                  N/A |
| N/A   47C    P0             12W /  N/A  | Not Supported          |      0%      Default |
+-----------------------------------------+------------------------+----------------------+
| Processes:                                                                              |
|    0   N/A  N/A         3005622      C   VLLM::EngineCore              95665MiB         |
+-----------------------------------------------------------------------------------------+

# inside the vLLM container (CUDA 13.2 toolkit):
$ nvcc --version | tail -2
Cuda compilation tools, release 13.2, V13.2.51
Build cuda_13.2.r13.2/compiler.37434383_0

Edit: I’m running this server headless, never tried display output

10

@eugr_nv @eugr maybe you can share with us some news, what we can expect in coming months?

11

Do you notice any difference after upgrading?

12

I haven’t upgraded. Very first thing I’ve done is installed from ubuntu server iso and wiped the nvidia OS. There were some minor quirks first (I think I had the stock kernel, not the nvidia version, and i was experiencing the PD bug) but resolved these. No issues eversince, Using it everyday for a bit of inference and as my home server.

13

We know - but WHY?
this is not “bleeding edge” anymore its been a long time and the customers are waiting.

14

If its listed in Supported GPU Products lists, then it should be deployed to Sparks. Why the massive lag in deployment times?

15

When? Due date, please!

Many thanks in advance!

CM

16

IMG_20260615_221646_228
IMG_20260615_221646_2281280Ă—550 96.4 KB

Ubuntu 26.04, fresh install on USB drive wth official external Nvidia cuda repository.
CUDA is not stable. This small model works, but PP should be around 4000. TG is OK.

17

I’m thinking to install ubuntu 26.04 too and convert to ZFS (to control file cache size through arc_max, also to use compression). You’ve installed nvidia kernel from the nvidia repository?

18

No, theres an nvidia flavored kernel in the official ubuntu repos.

19

Here are some of my recent benchmarks on the Ubuntu 26.04 system:

Gemma 4 31B on DGX Spark GB10 — NVFP4 vs QAT (fp8 KV)

Hardware/runtime (both runs): DGX Spark GB10 (Grace Blackwell, SM121, 122 GiB UMA),
Docker image vllm-node-tf5, single GPU (TP=1), dtype=bfloat16,
CUDA graphs FULL_AND_PIECEWISE, torch.compile (inductor).

Server configurations

NVFP4 QAT — fp8 KV
vLLM version 0.20.2rc1.dev234+gcf0d27914 (2026-05-11) 0.22.1rc1.dev124+gace95c9cf (2026-06-03)
Main model nvidia/Gemma-4-31B-IT-NVFP4 google/gemma-4-31B-it-qat-w4a16-ct
Quant method modelopt_fp4 (NVFP4) compressed-tensors w4a16 → Marlin kernel
Drafter (MTP) gemma-4-31B-it-assistant gemma-4-31B-it-qat-q4_0-unquantized-assistant
num_speculative_tokens 10 8
KV cache dtype fp8 (baked into NVFP4) fp8 (--kv-cache-dtype fp8)
max-model-len 262144 262144
gpu-memory-utilization 0.85 0.85
Load format (InstantTensor) fastsafetensors
Date 2026-05-12 2026-06-15

Shared serve flags: --max-num-seqs 4 --max-num-batched-tokens 16384 --enable-prefix-caching --enable-chunked-prefill --enable-auto-tool-choice --tool-call-parser gemma4 --reasoning-parser gemma4 --trust-remote-code --no-enable-flashinfer-autotune --default-chat-template-kwargs '{"enable_thinking": true}'

Results

llama-benchy: --pp 2048 --tg 128 --depth 0 16384 --runs 2 --latency-mode generation

Config pp2048 (t/s) tg128 (t/s) tg128 peak pp2048 @16k (t/s) tg128 @16k (t/s) tg128 @16k peak
NVFP4 1763.8 ± 119.4 14.8 ± 1.3 24.0 785.6 ± 1.2 17.9 ± 0.6 35.5
QAT — fp8 KV 816.3 ± 23.5 25.2 ± 0.2 38.5 630.8 ± 0.8 22.5 ± 3.5 33.5

pp = prompt processing (prefill), tg = token generation (decode),
@16k = at 16,384-token context depth, peak = best single-run decode rate (MTP acceptance varies).

Caveats

  1. Different vLLM versions (0.20.2 vs 0.22.1) — not a clean controlled comparison.
  2. Different n_spec (10 vs 8).

Takeaways: NVFP4 wins prefill (~2.2x at depth 0). QAT + matched QAT drafter wins decode
(~1.7x at depth 0); the matched drafter is the dominant decode lever.

20

This tests with 610 driver and cuda 13.3?
By default cuda repo install 595, and i force it to 610 just for check games performance.