Nvidia Spark GB10: MSI EdgeXpert running Steam Games!? Cyberpunk 2077, Doom Eternal and More?!? Quickie How-To

Wikis & How-to Guides
1

Background

20251112_150841
20251112_1508411920×1440 147 KB

Here’s the MSI EdgeXpert – it’s a GB10 device based on NVIDIA’s DGX Spark Platform with 128G of memory; MSI’s version of the Nvidia Spark. It has better thermal and power dissipation characteristics and as a result is 5-10% better for most tasks.

  • NVIDIA® Grace Blackwell Architecture:
  • NVIDIA Blackwell GPU & Arm 20-core CPU
  • NVIDIA® NVLink®-C2C CPU-GPU memory interconnect
  • 128 GB LPDDR5x coherent, unified system memory
  • 1000 AI TOPS (FP4) AI performance
  • Full stack solution, hardware & software, designed for AI developers
  • ConnectX-7 2x100G interface
  • 10G RJ-45 Ethernet Interface

One of the most demanding tasks for this platform, beyond pure benchmark compute, is multi-framegeneration gaming at 1080p and beyond, so it only makes sense to use that as a thermal load in order to test that it is able to sustain 200w+ power draw without overheating or throttling.

While one would not buy this system in lieu of a $3000-$4000 gaming PC, it is a lot of fun to see how it runs.

20251112_151255
20251112_1512551920×1440 137 KB

Step 1 - Box64

This is where I started. I don’t know that this step is really needed with the fex autoinstall (step 2) below. You should be aware of box64 and its ability to run 64 and 32 bit x86 code really well, though. For the sake of completeness, I’m including this here, but I think you only need the step 2 fex_autoinstall. Probably.

git clone this repo:


git clone https://github.com/ptitSeb/box64
cd box64
mkdir build; cd build; cmake .. -DBOX32=ON -DBOX32_BINFMT=ON
make -j16 
sudo make install

image
image795×126 32.8 KB

image

Next (re)start the binfmt systemd service:

image
image725×29 27.7 KB

Then run the steam_install.sh script:

./steam_install.sh

image
image817×531 139 KB

… but wait, there’s more!

Step 2 – Steam & Fex Installer for Ubuntu on arm64

forked from:

…which was last updated just a few days ago. I’m not sure which is the better install as of 2025-11-12 but leaning slightly toward esullivan’s fork.

esullivan at nvidia is awesome.

This is what I ran:
https://raw.githubusercontent.com/esullivan-nvidia/fex_autoinstall/refs/heads/main/fex_autoinstall_poc.sh

image

then from the cli when I ran steam, it picked up all the i386 stuff it needed.

Oddly, it didn’t find libgl1-mesa-glx but it still worked.

image
image821×519 95.8 KB

From there steam via the gui or /usr/local/bin/steam should work? and it should self-update:

image
image1023×603 90 KB

From there install games, and see how it goes? Here I’m using 4xMFG and playing at “170” fps.

20251112_151313
20251112_1513131920×1440 164 KB

3 Likes
2

hey Wendell, great talking to you at SC – I was the guy who George walked over to right after Jensen finished signing mugs :joy: I’m curious whether or not you’ve been able to try multi-Spark inference yet. Nvidia seems to Not Want To Discuss™ clustering GB10s with more than two nodes and I can’t decide if that’s because it will scale poorly when trying to run larger models than 128-256GB RAM can hold or because they don’t want people thinking too hard about how one might be able to use, say, an 8-node cluster of them to run very large models at something approaching reasonable performance. but I haven’t seen anyone actually post any benchmark results from testing inference on a GB10 cluster yet!

3

It works but It Is Not Fast ™

Qwen3-235B FP4

root@spark-2903:/app/tensorrt_llm# curl -s http://localhost:8355/v1/chat/completions   -H "Content-Type: application/json"   -d '{
    "model": "nvidia/Qwen3-235B-A22B-FP4",
    "messages": [{"role": "user", "content": "Seizing the means of computation might be necessary because"}],
    "max_tokens": 200
  }'


Okay, so the user is asking about \"seizing the means of computation\" 
and why that might be necessary. Hmm, first I need to understand what 
that phrase means. It sounds like a play on \"seizing the means of
 production,\" which is a Marxist concept referring to taking control of
 industries and resources from the capitalist class. So, applying that to 
computation, maybe it's about taking control of computational resources
 like servers, data centers, AI systems, etc.\n\nThe user probably wants to 
know the reasons why someone would advocate for seizing computational 
resources. Let me think about the possible motivations. Maybe it's about 
addressing inequality in access to technology, preventing monopolies, 
ensuring ethical use of AI, or democratizing control over data and
 algorithms. There's also the aspect of surveillance capitalism and how
 large tech companies collect massive amounts of data. Seizing 
computation could be a way to redistribute that power.\n\nI should
 also consider the potential negative consequences. If a government
 or group seizes computational resources

^ that response took more than a minute

this is TensorRT LLM over infiniband with docker swarm. the stack is two instances, one on each Spark.

4

ps. dont forget to take a peek at the benchmark notes thread:

1 Like