Your DGX Station Has 72 CPU Cores. Why Might Your Software Still Struggle?

Because those 72 cores are Arm, not x86. Every DGX Station GB300 pairs a Blackwell Ultra GPU with an NVIDIA Grace CPU of 72 Arm Neoverse V2 cores, so the operating system, drivers, Python packages and containers all have to be aarch64 builds. The mainstream AI stack is ready: CUDA, PyTorch, JAX, RAPIDS, vLLM, Ollama, Docker and NVIDIA's NGC containers all ship Arm versions. The problems sit in the tail: Python wheels published only for x86, older libraries that assume 4 KB memory pages, hand-written x86 SIMD code, and commercial applications such as MATLAB that have no Arm Linux release at all. Check your full dependency list before you order, not after delivery.

What runs natively on the Grace CPU today?

From NVIDIA's DGX Station Development Guide (software stack and dependencies pages, updated July to October 2026), the PyTorch blog (18 May 2026), the NVIDIA framework support matrix, NVIDIA AI Enterprise end-of-life notices, tensorflow.org, MathWorks R2026b requirements, Ansys' January 2025 Arm support announcement, Isaac Sim requirements (October 2026) and a secondary mirror of Siemens' STAR-CCM+ 2406 documentation, all accessed 4 Nov 2026. SourceIT has not tested these applications on a DGX Station.

Why does Arm matter if the GPU does the heavy lifting?

The GPU does the matrix maths, but the CPU still runs the operating system, the Python interpreter, data loaders, tokenisers, preprocessing and every library your code imports. If a single dependency in that chain has no Arm build, the whole job stops at install time. That is a packaging problem, not a performance one, and it is the most common reason a new Grace system sits idle in its first weeks.

Cornell's DGX Station team described exactly this in an early-2026 write-up: three students spent "many days" configuring software, noting that "even vLLM and Huggingface's TRL libraries are just shipping out compiled wheels as we write this post". Their advice was to start from NVIDIA's NGC containers. Since then, PyTorch 2.11 fixed one of the most confusing traps: on earlier versions, a plain pip install on Arm quietly gave you a CPU-only PyTorch, so code ran but never touched the GPU.

What is the 64 KB page-size issue?

This is the least-known catch. NVIDIA says DGX Station is "fine-tuned to support applications and workflows suited to a 64 KB page size", and that "legacy applications or models that require a 4 KB page size are not supported at this stage". Most x86 Linux systems use 4 KB pages, so some older software hardcodes that value. NVIDIA's advice to developers is to query the page size at run time with sysconf(_SC_PAGE_SIZE) rather than assume it. For your own code, that is a small fix. For a closed-source binary, you have to wait for the vendor.

Seventy-two fast cores do not help if one closed-source library in your pipeline only exists for x86. On Grace, compatibility is decided by your weakest dependency, not your strongest.

Can't you just emulate x86?

Partly, but not for GPU work. Box64, FEX and QEMU user mode can all run x86-64 Linux programs on Arm64. None of them carries the CUDA stack across for an x86 binary, so emulation may rescue a command-line utility but not a GPU application. QEMU user mode also cannot run container runtimes inside it. For source code with SSE or AVX intrinsics, NVIDIA points to SIMDe and SSE2NEON as drop-in translation layers, with Arm SVE added case by case.

The coming DGX Station for Windows does not remove this issue either. Windows on Arm's Prism emulation runs x64 apps in user mode only, and drivers must be native Arm64. We cover that side in NVIDIA DGX Station for Windows: Breakthrough or Another Compatibility Headache?.

How do you check your software before buying?

1. List every dependency, including Python packages, system libraries, licence servers and any commercial solver or plug-in.
2. Look for an aarch64 or "arm64 SBSA" build of each one on PyPI, conda-forge, the vendor site or NGC.
3. Prefer NGC containers for PyTorch, JAX, vLLM and TensorRT-LLM; NVIDIA publishes them for Arm.
4. Flag anything x86-only (MATLAB on Linux, x86-only wheels, closed binaries) and decide whether it stays on an x86 workstation.
5. Grep your own code for hardcoded 4096-byte page sizes and x86 intrinsics.
6. Trial the stack on Arm first. A GB10 system such as the NVIDIA DGX Spark also runs aarch64 Linux with CUDA, so it is a cheaper way to see whether your packages install. It is not identical to Grace, so treat it as a packaging test, not a performance preview.

Engineering teams face the hardest version of this question. If your simulation, CAD or MATLAB work is x86-only, a DGX Station will not replace those workstations; see Could DGX Station Replace a Room Full of Traditional Engineering Workstations?. For what the Station does run well, see What Can You Run on a DGX Station GB300?.

Is Grace slow for general CPU work?

There is no published evidence of that. Grace has 72 Neoverse V2 cores, each with four 128-bit SVE2 units, 1 MB of L2 per core and 114 MB of shared L3. It reaches its 496 GB of LPDDR5X at 396 GB/s and the GPU over 900 GB/s NVLink-C2C. NVIDIA's tuning guide recommends compiler settings and atomics that suit Arm, so performance-sensitive code may need recompiling with the right flags. We found no Grace-specific benchmark showing a single-thread or OpenMP scaling problem, so the realistic risk is software availability, not raw speed.

Buying in Singapore

All five DGX Station GB300 systems SourceIT supplies run the same 72-core Grace CPU: the HP ZGX Fury G1n, MSI XpertStation WS300, ASUS ExpertCenter Pro ET900N G3, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC, all price on request. To trial your software on Arm first, the NVIDIA DGX Spark GB10 128GB/4TB is S$9,300.00 (listed 4 Nov 2026). For x86-only applications, the HP Z2 Tower G1i with RTX PRO 2000 is S$5,788.00 (listed 4 Nov 2026). Browse the NVIDIA DGX Spark and DGX Station collection, and send us your dependency list with your enquiry so we can flag Arm gaps. For a quotation, use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Is the DGX Station CPU x86 or Arm? Arm. The GB300 superchip uses an NVIDIA Grace CPU with 72 Arm Neoverse V2 cores, so software must be built for aarch64 (arm64).

Does PyTorch run on DGX Station? Yes. NGC PyTorch containers are multi-architecture, and from PyTorch 2.11.0 a plain pip install on aarch64 gets the CUDA-enabled build. Earlier versions installed a CPU-only wheel by default.

Can I run MATLAB on a DGX Station? Not natively. MathWorks lists an Intel or AMD x86-64 processor as a requirement for MATLAB R2026b on Linux, so MATLAB work belongs on an x86 workstation.

What is the 64 KB page size problem? NVIDIA tunes DGX Station for 64 KB memory pages and says applications that require 4 KB pages are not supported at this stage. Software that hardcodes a 4 KB page size may fail and needs to query the page size instead.

Can x86 emulation run my GPU software? No. Box64, FEX and QEMU user mode can run some x86-64 Linux programs on Arm, but they do not bring CUDA support for x86 binaries, so GPU applications need native Arm builds.