TcpNode separates CUDA development from general-purpose workloads. Check software architecture, OS dependencies, and data retention requirements before choosing a server. This guide covers DGX Spark development first, then planning workloads on AMD dedicated servers.
CUDA projects need compatible architecture, drivers, and toolchains—not just access to shared hardware. TcpNode offers dedicated NVIDIA DGX Spark systems with NVIDIA DGX OS, CUDA, and the NVIDIA AI software stack preinstalled, ready for development, debugging, and experiments on one machine.
Before compiling or debugging CUDA kernels, confirm that dependencies support ARM64. Use container images with ARM64 builds; x86 binaries and images won't run as-is. NVIDIA DGX OS is based on Ubuntu 24.04 and cannot be replaced with Windows or another distribution.
Run local inference, LoRA / QLoRA fine-tuning, and AI application prototypes on the same system. Its 128GB LPDDR5x is unified memory shared by the CPU and GPU, not dedicated VRAM. Model weights, runtime caches, the OS, and applications all use this memory.
NVIDIA's published figures put quantized inference at around 200B parameters and LoRA / QLoRA fine-tuning at around 70B as rough scale references. Whether a workload fits and meets throughput targets depends on quantization, context length, batch size, and software implementation—not parameter count alone.
Check ARM64 support for Python packages, native extensions, and container base images. Root access is available for installing other development tools. Models aren't preinstalled; prepare model weights and application dependencies yourself.
Fixed configuration: no disk, IP, bandwidth, or memory upgrades.
At 1 Gbps, a 40GB download takes about 5.5 minutes in theory—not a delivery-time guarantee. Actual transfer speeds depend on the source, network path, and protocol. Export model weights, code, and logs before the rental ends.
Data is erased after the rental ends. Request hard reboots and reinstalls through a console ticket; no BMC / IPMI.
$16.7/day. Total = days × daily rate; choose 1–365 days. No discounts, setup fees, or contracts. Before ordering, confirm that ARM64 won't run x86 programs or Windows, DGX OS is the only available OS, and data is erased at the end of the rental.
AMD EPYC 7543P dedicated servers offer 32 cores / 64 threads and a choice of 128GB or 256GB DDR4-3200 ECC memory. Both configurations include 2×1TB direct-attached NVMe drives, 250 Mbps unmetered bandwidth, and one dedicated IPv4 address. Plan around your workload, then check monthly pricing and options.
With 32 cores and 64 threads, applications, databases, and test environments can run in separate VMs. The CPU supports AMD-V and NPT for nested virtualization. Proxmox VE 8 or 9 is available at no extra cost.
Before assigning vCPUs, distinguish sustained usage from occasional peaks; don't allocate every thread to VMs by default. Reserve memory for the host OS, storage cache, and management tasks. Choose 128GB or 256GB based on the combined VM working set.
Nested virtualization also requires compatible host settings, VM CPU configuration, and guest OS support; it isn't fully configured by default.
DDR4-3200 ECC memory can support database caches, indexes, and middleware working sets. Both 1TB NVMe drives are direct-attached, with no hardware RAID. Plan data directories, logs, and backups yourself.
For disk redundancy, configure a ZFS mirror or software RAID. Mirroring reduces usable capacity and doesn't replace backups: accidental deletion, logical corruption, and application errors can still be replicated to the other drive.
Add up to two additional 2TB NVMe drives, each direct-attached. Keep recoverable off-site copies of important data as well.
Hosted in Tokyo, Japan. Each server includes one dedicated IPv4 address; add up to 255 more, for a maximum of 256. Suitable for multi-site deployments and workloads that need network separation or dedicated IPs.
Choose 250 Mbps, 1 Gbps, or 2 Gbps; all options are unmetered. Bandwidth is a transfer rate, not a response-time guarantee. Performance also depends on the client, network route, and application.
Use the first address in the assigned list for SSH / RDP; configure the remaining addresses in the OS.
Windows Server Standard / Datacenter is available in 2019, 2022, and 2025 editions for Windows-dependent services, management software, and business applications. Licenses are billed monthly; connect over RDP.
Choose English, Traditional Chinese, or Simplified Chinese as the system language when ordering. Check application licensing, component versions, and service account permissions separately; the OS license doesn't include business software licenses.
After payment and provisioning, check the address and credentials in the console. On first login, review access rules and manage account passwords securely.
Choose from 14 free operating systems for AMD dedicated servers. Select a version based on application dependencies, filesystem, and management needs—not familiarity alone. Windows Server Standard and Datacenter require separate licenses.
3 free versions for workloads that need enterprise Linux compatibility checks.
2 free versions. Check repository and component compatibility before deployment.
3 free versions. Choose based on application support.
3 free versions. Check runtime, driver, and container dependencies.
1 free version. Confirm application support for FreeBSD before deployment.
2 free versions for planning hosts, VMs, and network bridges.
Both paid editions are available in 2019 / 2022 / 2025 versions, with licensing added monthly to the configuration quote. Connect to Linux, FreeBSD, and Proxmox VE over SSH, and Windows over RDP. These OS options don't apply to DGX Spark.
Resource planning addresses capacity; a recovery checklist helps keep workloads running after mistakes or failures. TcpNode provides the hardware and network; the renter manages the OS, applications, data, and backups.
Record VM resources, service ports, and disk usage. Keep backups separate from source data—not just on the same machine.
Check that backups are readable, applications can be rebuilt, and configurations and keys are stored separately.
Include the order number, symptoms, steps already taken, and relevant logs in support tickets. Never send login passwords in email.
Memory cannot be upgraded on the original server after provisioning; changing tiers requires a new server. Select extra drives, bandwidth, IPv4 addresses, and Windows licenses separately on the configuration page. The monthly total is based on selected options; amount due = monthly total × number of months. Monthly, quarterly, semiannual, and annual billing have no discounts, setup fees, or contracts.
The catalog shows available options; live availability is confirmed in the console. All charges are in USD. Available payment methods are listed on the configuration page.
Check memory working sets, disk redundancy, and IP requirements, then review monthly pricing and options for both AMD dedicated server tiers.