GDS support for DGX Spark

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

Hi,

I am testing GPUDirect Storage on a DGX Spark system

(Grace-Blackwell, NVIDIA GB10 GPU, ARM64, Ubuntu 24.04, CUDA 13.0).

Gdscheck reports:

GPU index 0 NVIDIA GB10: Model Not Supported

NVMe / NVMeOF / BeeGFS / etc: Unsupported

properties.use_compat_mode : true

My versions:

- CUDA: 13.0

- Driver: 580.95.05 (open kernel)

- libcufile: 13-0, 1.15.1.6-1

- nvidia-gds-tools: 1.15.1.6-1

- Kernel: 6.11.0-1016-nvidia (aarch64)

- Platform: DGX Spark

Question:

Is GDS fast-path support for GB10 not yet enabled in the current libcufile/nvidia-gds?

If so, which future release (libcufile / GDS / driver) will add GB10 support?

I want to confirm the expected timeline for GDS enablement on DGX Spark.

Thanks!

1 Like
2

I’m also looking for GDS enablement on DGX Spark.

3

GDS is not supported on Spark, due to DGX Spark / GB10 FAQ

1 Like
4

Thanks @NVES , that entry in the FAQ helps. Two doubts though

  • The FAQ says “GPUDirect RDMA is not supported”. Seems “GPUDirect Storage” (GDS) has the same challenge (GPU memory is behind C2C and NVMe SSD is behind PCIe) and therefore not supported?
  • Is this constraint applicable for other NVlink C2C based systems (like GH200) where GPU is no longer behind the PCIe?
5

no. This limitation applies to Gb10 Spark only.

6

@NVES I see, interesting. But on other c2c system (GH or GB200) , GDS works without any hops? Notwithstanding architecture change, can NVMe SSD still do DMA into the GPU memory without any extra hop like bounce buffering of data into CPU side LPDDR before it moves to GPU memory via C2C?

1 Like