AI Development Machine,
128GB Unified Memory
Shared by CPU and GPU
A dedicated NVIDIA DGX Spark for one customer, with no resource sharing.
SSH access · CUDA preinstalled on DGX OS
One DGX Spark per customer, with a 128GB memory pool that is never shared.
NVIDIA DGX OS (based on Ubuntu 24.04) is the only available OS. The NVIDIA AI software stack below is preinstalled. Root access allows additional tools to be installed; no models are preinstalled.
One dedicated machine. Three facts, clearly laid out.
Model, capacity, and billing formula are all listed upfront. No shared-GPU queues or resource contention on a dedicated machine.
128GB LPDDR5x unified memory
The CPU and GPU share one memory pool. Model weights, KV cache, and CUDA programs are scheduled in that pool, without copying data back and forth between host and device memory.
Dedicated GB10 Grace Blackwell system
Each NVIDIA DGX Spark is assigned to one customer, with no sharing. Jobs do not queue or get preempted by other tenants.
No setup fee, no contract, no discounts
Total = days × daily rate. The listed rate is the price charged: no one-time fees and no minimum rental period.
Weights, KV cache, and CUDA programs in one pool
NVIDIA GB10 Grace Blackwell gives both the Grace CPU and Blackwell GPU access to 128GB of LPDDR5x unified memory. According to NVIDIA, the pool can support quantized inference with around 200B parameters or LoRA/QLoRA fine-tuning with around 70B parameters.
See TcpNode model size guideOne GB10.One customer.
Drivers and tools come ready with DGX OS. Root access is included, so additional versions can be installed as needed.
Dedicated 20-core Arm and Blackwell GPU system
The CPU combines 10×Cortex-X925 and 10×Cortex-A725 on ARM64. Connect over SSH with root access. The machine is not shared, so builds, debugging, and long-running inference run without a queue.
View TcpNode DGX Spark specsPay per day. One simple billing formula.
DGX Spark is rented by the day, for 1–365 days. No discounts, setup fee, or contract. All data on the machine is erased at the end of the rental; transfer any weights, code, and logs to another location before then.
Explore TcpNode DGX SparkTwo types of work, one machine
NVIDIA DGX OS (based on Ubuntu 24.04) is provided, with CUDA and NGC containers ready to use.
Build and debug CUDA kernels on ARM64
The CUDA development environment is ready with the system, without having to install each layer starting with the drivers. Use NGC containers for deployment; ARM64 images can be pulled and run directly.
Local LLM inference and fine-tuning experiments
Well suited to local inference, LoRA fine-tuning, and AI application prototyping. No models are preinstalled; model weights are downloaded separately. At the theoretical 1 Gbps rate, 40GB takes about 5.5 minutes.
Is DGX Spark a good fit?
The hardware and OS are fixed. Compare these two lists before ordering.
Good fit
- CUDA development and debugging
- ARM64 container deployment and NGC image testing
- Local LLM inference and throughput testing
- LoRA fine-tuning experiments
- Getting started with DGX OS and AI app prototyping
Not a fit
- x86 software or Windows; ARM64 cannot run them
- Multi-machine clusters or cross-machine training
- Workloads requiring bandwidth, IP, or storage changes
- Workloads requiring more memory or a different OS
Frequently asked before ordering
What happens to data when the rental ends?
All data on the machine is erased at the end of the rental. Transfer any model weights, code, and logs to another storage location before then.
Can it run x86 software or Windows?
No. DGX Spark uses ARM64 and NVIDIA DGX OS is the only available OS. x86 images need to be replaced with ARM64 versions.
Are any models preinstalled?
No models are preinstalled. The NVIDIA AI software stack, including CUDA, cuDNN, and TensorRT, is ready with the system; model weights must be downloaded separately.
Where are login details, and how are issues handled?
After delivery, find the IP address and login credentials in the machine details in the control panel, then connect over SSH. Submit support tickets in the control panel for faults, hard reboots, or OS reinstallation.
DGX Spark, from $16.7 per day
NVIDIA GB10 Grace Blackwell · 128GB LPDDR5x unified memory · Data center in Tokyo, Japan
Validate CUDA stream dependencies and buffer lifetimes on DGX Spark
A complete two-stream example links producer and consumer with an event, checks every output element and defines when asynchronous buffers can be reused.
AI AutomationAuditing CUDA build artifacts on DGX Spark: targets, SASS and PTX
Build for the detected GPU target, inspect embedded device code, and verify native execution and PTX fallback through separate test runs.
AI AutomationFind CUDA out-of-bounds accesses on DGX Spark
A controlled invalid write demonstrates source-level diagnosis, asynchronous error checks, failure exit codes, and verification of the corrected kernel.
New articles are in the works. In the meantime, browse the blog.