TcpNode dedicates an entire NVIDIA DGX Spark to one customer, with no sharing. It is suited to CUDA development and debugging, local inference, fine-tuning experiments and AI prototyping.
SSH access · CUDA preinstalled on NVIDIA DGX OS
The TcpNode DGX Spark is not split into separate CPU, GPU or memory allocations. Storage, networking and an IP address are included with the system. Choose the number of rental days at checkout; hardware options are not available.
The chip, memory and operating system combine to provide an ARM64 development environment.
NVIDIA DGX OS includes CUDA and the NVIDIA AI software stack. TcpNode provides root access, so additional tools can be installed as required by the project. No models are preinstalled; model weights and datasets must be prepared separately.
Program compilation and debugging
Deep learning library
Inference optimization
ARM64 container resources
Model development framework
Inference serving experiments
Run models locally
Choose software and container versions compatible with ARM64, GB10 and the current CUDA environment. Check project dependencies and model support before deployment; x86 images cannot be used as drop-in replacements for ARM64 images.
The 128GB of unified memory is shared by the CPU and GPU; it is not dedicated video memory. Memory planning should account for KV cache, runtime, context length and concurrent requests in addition to model weights.
Around 200B parameters
According to NVIDIA's published guidance, DGX Spark can serve as a reference platform for quantized inference at this scale. Actual resource requirements depend on the model format, quantization level and inference framework.
Use this as a guide when evaluating an experiment, not as a guarantee that every 200B model will run with the same configuration.
Around 70B parameters
According to NVIDIA's published guidance, parameter-efficient fine-tuning experiments at this scale are possible. Training sequence length, batch size and optimizer settings still need to be validated for each project.
This reference does not indicate full-parameter training capability or guarantee throughput or completion time.
The TcpNode DGX Spark rate is $16.7 per day, with rental periods from 1–365 days. There are no discounts, setup fees or contracts, and no monthly, quarterly, semiannual or annual billing periods.
$16.7/day
Total = days × daily rate
$16.7
1 × $16.7 = $16.7
$116.9
7 × $16.7 = $116.9
$501.0
30 × $16.7 = $501.0
The 30-day example is a calculation by number of days, not a monthly plan. Set the rental period on the dedicated checkout page. Charges are settled in USD; data is erased when the rental period ends, so plan to transfer it beforehand.
Check instruction set, operating system and resource expansion requirements before evaluating model size. A dedicated system does not change software compatibility or the fixed hardware specifications.
No models are preinstalled on DGX Spark. The estimates below use the theoretical link rate of 1 Gbps to help plan file transfers. They are not guarantees of actual download speeds or delivery times.
| Total file size | Estimated time | Estimate based on |
|---|---|---|
| 40GB | About 5.5 minutes | 1 Gbps theoretical rate |
| 100GB | About 14 minutes | 1 Gbps theoretical rate |
| 140GB | About 19 minutes | 1 Gbps theoretical rate |
Actual transfer times depend on the download source, network path, protocol overhead and concurrent traffic. Allow additional time for extraction, verification and model loading. The 4TB NVMe drive stores the system, containers, model weights and experiment output, so leave sufficient working space before deployment.
Data is erased when the DGX Spark rental period ends. Any model weights, code, datasets, configuration files or logs that need to be kept must be transferred and independently backed up before expiry. A second copy on the machine is not a substitute for a separate backup.
Along with final models, check fine-tuning adapters, dependency lists, container build files and configuration needed to reproduce experiments. Verify file integrity after transfer to avoid keeping screenshots of a run while missing its actual outputs.
This system has no BMC / IPMI. Request a hard reboot or system reinstall through a console support ticket. Before a reinstall, state the intended operation and make an independent backup. Find the assigned IP address and login credentials in the console; connect over SSH for routine access.
These requirements are confirmed again on the TcpNode DGX checkout page. This page describes the product; it does not submit orders or process payments.
x86 programs and Windows cannot run on this system. Check ARM64 support for application code, binary dependencies and container images.
The system runs NVIDIA DGX OS only, based on Ubuntu 24.04 and including the CUDA stack. No other distributions are installed.
Data is erased when the rental period ends. Transfer any model weights, code and logs that need to be kept before expiry; customers are responsible for backups.
The TcpNode DGX Spark rate is $16.7 per day, for rental periods of 1–365 days. Choose the number of days, review the terms and place an order on the next page.