PersonaPlex produces choppy/unusable audio on DGX Spark (GB10)

[roychri]

Environment

• Hardware: DGX Spark with GB10 (Blackwell, sm_121)
• Memory: 128GB unified
• PyTorch: 2.9.0+cu130
• PersonaPlex: latest from pip

Problem

Audio output is extremely choppy and unusable for real-time conversation. The model runs but cannot keep up with real-time audio streaming.

Steps to reproduce

1. Install PersonaPlex on DGX Spark following README
2. Run the server with python -m moshi.server --ssl <path>
3. Connect via web UI and attempt conversation
4. Audio is choppy both from host machine and remote clients

What I've tried

• Disabled torch.compile and CUDA graphs
• Disabled Flash Attention (SDPA fallback)
• Various quantization approaches (all failed or didn't help)

Expected behavior

Real-time conversational audio as demonstrated in the model card

[acatovic]

I actually get the same result on RTX5070 - everything is up and running but the sound is choppy and completely unusable.

Same procedure as OP.

[rajarshiroy-nvidia]

Could you try the suggested installation fix from #2 and let me know if that works. I am really sorry but I have to get a similar setup to debug what is happening exactly. Will try my best to debug. The only time I have personally experienced this choppiness is if GPU pytorch was not being used, or CUDA graph compilation was disabled so these functions:

graphed_main = CUDAGraphed(lm_model.forward_codes, disable=disable)
graphed_embeddings = CUDAGraphed(lm_model.forward_embeddings, disable=disable)
graphed_depth = CUDAGraphed(self.depformer_step, disable=disable)

at
https://github.com/NVIDIA/personaplex/blob/main/moshi/moshi/models/lm.py#L720
were super slow.

[jvrsgsty]

I'll be trying this out on a Jetson Thor soon. Is this expected to work realtime with this type of compute? Given the results on the DGX Spark which has similar specs seems like maybe not?

[darkduran]

I'll be trying this out on a Jetson Thor soon. Is this expected to work realtime with this type of compute? Given the results on the DGX Spark which has similar specs seems like maybe not?

I tried and failed because Cuda capability sm_110 is not supported, Please share your experience after. Thanks

[rajarshiroy-nvidia]

Working on reproducing and finding a fix. Thanks for raising this.

[rajarshiroy-nvidia]

Trying to reproduce and determine a fix. For now, the most likely fix is to do:
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

[wachawo]

Trying to reproduce and determine a fix. For now, the most likely fix is to do: pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

After installing on DGX Spark torch with https://download.pytorch.org/whl/cu130

Too same: Audio playback has frequent dropouts and the Latency is constantly increasing, GPU-Util 91%, and One CPU core is permanently loaded at 100%

Profiling shows that almost all CPU time is spent inside opus_loop in moshi/server.py:

%Own   %Total  OwnTime  TotalTime  Function (filename)
84.00%  96.00%   44.05s    49.71s   opus_loop (moshi/server.py)

Environment on DGX Spark:
torch == 2.10.0+cu130
torchaudio == 2.10.0+cu130
torchvision == 0.25.0+cu130
CUDA 13.0
Platform: DGX Spark (ARM / aarch64, NVIDIA GB10)

For comparison, on a different machine everything works perfectly:
torch == 2.10.0+cu130
torchaudio == 2.10.0+cu130
torchvision == 0.25.0+cu130
CUDA 13.0
GPU: RTX 4090
CPU: Intel i7 (x86_64)

Latency ~0.1s (stable), GPU utilization ~80%, No audio dropouts. Everything works PERFECTLY as in the Demo

[ordex]

Same problem here, even after having executed:

 pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

However, please note that at the end of the above command I got the following:

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behavio
ur is the source of the following dependency conflicts.
moshi-personaplex 0.1.0 requires torch<2.5,>=2.2.0, but you have torch 2.10.0+cu130 which is incompatible.

When launching moshi.server I get the following:

    Found GPU0 NVIDIA GB10 which is of cuda capability 12.1.
    Minimum and Maximum cuda capability supported by this version of PyTorch is
    (8.0) - (12.0)

Could those warnings be related?

[acatovic]

@ordex i actually had to change moshi/pyproject.toml from "torch >= 2.2.0, < 2.5", to "torch >= 2.2.0", to get it to work. I think that's a separate issue from choppy sound problem.

[s3world]

Following.. Hope someone has a fix to get Spark running this clean without drops.

[tomc98]

Yeah same problem, I watched the latency continue to climb, also it by default wants to use the efficiency cores, and is only limited to one. Choppy playback, the longer it goes on the worse the latency...

[rt974]

I had the same problem on vastai, was bad instance rtx 4090, cpu was limited i think. launched another instance.
connection is better, faster, and not choppy anymore.

`
Sat Jan 24 05:21:21 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 580.95.05 Driver Version: 580.95.05 CUDA Version: 13.0 |
+-----------------------------------------+------------------------+----------------------+
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 NVIDIA GeForce RTX 4090 On | 00000000:82:00.0 Off | Off |
| 30% 25C P8 14W / 430W | 20601MiB / 49140MiB | 0% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
+-----------------------------------------------------------------------------------------+
`

htop shows all the cores working fine.

That tech is so funny to use.

[emican]

Too same: Audio playback has frequent dropouts and the Latency is constantly increasing, GPU-Util 91%, and One CPU core is permanently loaded at 100%

Profiling shows that almost all CPU time is spent inside opus_loop in moshi/server.py:

%Own   %Total  OwnTime  TotalTime  Function (filename)
84.00%  96.00%   44.05s    49.71s   opus_loop (moshi/server.py)

Environment on DGX Spark: torch == 2.10.0+cu130 torchaudio == 2.10.0+cu130 torchvision == 0.25.0+cu130 CUDA 13.0 Platform: DGX Spark (ARM / aarch64, NVIDIA GB10)

On the DGX Spark, to address the CPU bottleneck, I built Opus with --enable-intrinsics, rebuilt sphn from source which bundled the optimized Opus and discovered GPU compute is the limiting factor. I observed frame time of 135ms where 80ms is required for realtime.

I believe the profiler is watching the event loop opus_loop with tight polling and busy-waiting and does not provide insight into the time spent on each task: receive audio, inference, and send audio.

[wachawo]

I observed frame time of 135ms where 80ms is required for realtime.

I absolutely agree.
I tested on DGX Spark and added stage-level profiling. With CUDA-event timing enabled, the GPU time is ~120-130 ms per 80 ms audio frame, while Opus is only ~0.5 ms. So the pipeline is GPU-bound and can’t keep up with real time, the backlog grows and latency steadily increases.

[ordex]

With CUDA-event timing enabled, the GPU time is ~120-130 ms per 80 ms audio frame, while Opus is only ~0.5 ms. So the pipeline is GPU-bound and can’t keep up with real time, the backlog grows and latency steadily increases.

would you conclude that the DGX Spark is just unable to keep up with the model/application? Is it really possible that a RTX 4090 can perform so much better?