NVIDIA DGX Spark and the GB10 Superchip — A Practical Guide for Malaysian Teams
A petaflop-class AI machine that runs off a normal wall socket has changed the maths for teams that want to run large language models without sending data to an overseas cloud. Here is what the GB10 Grace Blackwell platform actually is, where it fits, and where it does not.
For most of the last decade, running a large language model meant renting time on GPUs in a data centre — almost always outside Malaysia. NVIDIA's DGX Spark changes that calculation. It puts roughly one petaFLOP of AI compute and 128GB of unified memory into a box the size of a thick hardcover book, powered from an ordinary wall socket. For Malaysian teams weighing data residency, running API costs, and the practicalities of an on-premise AI setup, it is worth understanding what the underlying GB10 platform is before deciding whether it belongs on your rack.
What the GB10 Grace Blackwell Superchip actually is
DGX Spark is built around a single NVIDIA GB10 Grace Blackwell Superchip. On one package it combines:
- A 20-core Arm CPU — 10× Cortex-X925 performance cores plus 10× Cortex-A725 efficiency cores
- A Blackwell-generation GPU with fifth-generation Tensor Cores and the full CUDA software stack
- 128GB of LPDDR5x coherent unified memory shared between CPU and GPU, on a 256-bit interface with around 273 GB/s of bandwidth
- An NVLink-C2C link joining the CPU and GPU so both address the same memory pool without copying data back and forth
Rated tensor performance is up to 1 PFLOP at FP4 precision. The chip runs at roughly a 140W TDP with a 240W external power supply — no special electrical work, no dedicated cooling plant, standard office ambient conditions. Networking is handled by an on-board NVIDIA ConnectX-7 NIC with 200 Gbps connectivity. The system ships with DGX OS, an Ubuntu-based Linux build with NVIDIA's AI stack pre-installed.
Why the unified memory matters more than the FLOPs
The headline number people quote is the petaFLOP. In practice, the 128GB unified memory pool is what makes DGX Spark useful for work that a conventional workstation GPU cannot touch.
A 70-billion-parameter model quantised to 4-bit needs roughly 40GB just to hold its weights, before any context or working memory. On a consumer GPU with 24GB or 32GB of VRAM, you cannot load that model without splitting it across devices or offloading layers to system RAM — both of which cripple throughput. DGX Spark's 128GB coherent pool lets a single unit hold models up to around 200 billion parameters for inference, and comfortably fine-tune models in the 70-billion range using LoRA or QLoRA techniques.
Need more? Two DGX Spark units can be linked directly over their ConnectX-7 200 Gbps interfaces to run models up to around 405 billion parameters as a two-node system.
Where DGX Spark fits
It is a strong fit for:
- Local inference for LLM-backed products, retrieval-augmented generation (RAG) systems, internal copilots, and document-processing pipelines
- Prototyping and fine-tuning — iterate on a model locally, then rent a cloud cluster only for the final large training run
- Data that cannot leave the building — legal, medical, financial, and government workloads where sending prompts and documents to an overseas API is not acceptable
- Edge and robotics development where a full CUDA environment in a small power envelope is valuable
Where it does not fit
- Pretraining a foundation model from scratch — that still belongs on a multi-node data-centre cluster
- High-concurrency serving for hundreds of simultaneous users at maximum tokens per second — memory bandwidth here is 273 GB/s, not the multi-terabyte HBM of a data-centre accelerator
- Replacing an RTX workstation for 3D rendering, CAD, or video work — different tool for a different job
The data-residency angle for Malaysian organisations
Under the Personal Data Protection Act 2010 and its later amendments, and under sector-specific rules from regulators such as Bank Negara Malaysia and the Ministry of Health, where personal data is processed and stored is a live compliance question. Cloud LLM inference means user prompts, uploaded documents, and generated output all transit and are processed on infrastructure outside your control, usually outside the country. Running inference on hardware in your own rack keeps that data on premises and removes an entire category of cross-border data-transfer questions.
There is a cost dimension too. Cloud LLM APIs bill per token, and the bill scales with adoption — success makes it more expensive. An on-premise machine is a fixed capital cost with predictable running expense, which changes the conversation with finance once usage is steady.
A realistic first deployment
A common path looks like this. A team starts building on a hosted API because it is the fastest way to ship. Usage grows, the monthly bill climbs past a few thousand ringgit, and then an enterprise client asks where their data is being processed. The team buys a single DGX Spark, moves its RAG pipeline and its smaller inference workloads in-house, and keeps a cloud fallback only for occasional burst demand. Against the previous API spend, the hardware often pays for itself inside a few months — and the data-residency answer becomes "on our own hardware, in Malaysia".
The OEM question
NVIDIA designed the reference platform, and it is sold both as the NVIDIA Founders Edition and as OEM editions from ASUS, Dell, Gigabyte, Lenovo, MSI, and others. Every one of them ships the same GB10 board, the same 128GB of memory, and the same 1 PFLOP FP4 ceiling — no partner overclocks or unlocks the chip. What varies between editions is the chassis and cooling design, the SSD capacity on offer, the warranty length, and the support tier. We cover how those differences play out, including independent thermal testing, in a companion article: DGX Spark OEM Editions Compared.
Frequently asked questions
How large a model can one DGX Spark run?
Up to around 200 billion parameters for inference, thanks to the 128GB unified memory pool. Two units linked over their ConnectX-7 200 Gbps interfaces can run models up to around 405 billion parameters.
Can DGX Spark train models from scratch?
It is built for inference, fine-tuning, and prototyping — not for pretraining a foundation model from scratch, which still requires a data-centre cluster.
Does it need special power or cooling?
No. It runs from a 240W external power supply on a standard wall socket in normal office ambient conditions. Sustained heavy workloads do warm the unit; independent testing of the different OEM cooling designs is covered in our OEM comparison article.
How is DGX Spark different from an RTX 6000 workstation?
Different tools. DGX Spark's advantage is 128GB of unified memory for holding very large models. An RTX workstation card wins on raw memory bandwidth and on graphics, rendering, and CAD workloads.
Does running AI on-premise help with PDPA compliance?
Keeping inference on hardware you control avoids sending personal data overseas for processing, which removes a category of cross-border data-transfer questions under the PDPA and under sector-regulator guidance.
Which OEM edition should we buy?
The GB10 silicon is identical across every edition. Choose based on SSD capacity, warranty term, support tier, and cooling headroom for how hard you will run the machine. Full breakdown in our OEM comparison.
See more
Khalbyte supplies DGX Spark across the ASUS, Dell, Gigabyte, Lenovo, Leadtek, and MSI editions, and helps Malaysian teams match the configuration to the actual workload. Explore our IT supply service or talk to our team.