Two ways to get 128 GB of unified memory.
Every SparkHost machine is a real, physical unit dedicated to one customer. Specs below are the vendor platform specs; the exact unit you get (serial, storage, network) is shown in your portal.
Ready within 2 business days after payment. Rent it or own it: pay it off in 24 or 12 months and the hardware is yours. Clusters of 3–4 GB10 units include the networking (switch) at no extra cost.
NVIDIA DGX Spark
NVIDIA's GB10 Grace Blackwell Superchip with 128 GB of coherent unified memory and the full NVIDIA AI software stack. Built for running and fine-tuning large models locally, CUDA-native, aarch64.
Availability: if an NVIDIA DGX Spark isn't in stock, we'll provide an equivalent NVIDIA GB10 Grace Blackwell system from Dell (Pro Max with GB10), Lenovo (ThinkStation PGX) or ASUS (Ascent GX10). They have the same GB10 superchip, 128 GB unified memory and DGX OS software stack. Chassis and storage capacity can differ by model. Your portal always shows the exact model you're assigned.
Equivalent GB10 systems: Dell Pro Max with GB10 · Lenovo ThinkStation PGX · ASUS Ascent GX10
Good for
- Inference on models up to ~200B parameters at FP4
- Fine-tuning / LoRA on models up to ~70B parameters
- CUDA-native stacks: vLLM, TensorRT-LLM, NIM, PyTorch
- Prototyping for Grace Blackwell data-centre targets
| Superchip | NVIDIA GB10 Grace Blackwell |
|---|---|
| GPU | Blackwell architecture, 5th-gen Tensor Cores, 4th-gen RT Cores |
| AI performance | Up to 1 petaFLOP (FP4, with sparsity) |
| CPU | 20-core Arm: 10× Cortex-X925 + 10× Cortex-A725 |
| Memory | 128 GB LPDDR5x coherent unified system memory, 256-bit, 273 GB/s |
| Storage | 4 TB NVMe M.2, self-encrypting (DGX Spark; equivalent GB10 systems can differ, see note) |
| Networking | 10 GbE RJ-45 · ConnectX-7 SmartNIC (2× QSFP, up to 200 Gb/s) · Wi-Fi 7 |
| I/O | 4× USB Type-C · HDMI 2.1a |
| Software | NVIDIA DGX OS (Ubuntu-based), CUDA, NVIDIA AI software stack |
| Architecture | aarch64 (Arm64) |
| Power | 240 W external PSU (GB10 TDP ~140 W) |
| Size | 150 × 150 × 50.5 mm, 1.2 kg (DGX Spark chassis) |
Sold out Get notified when DGX Spark units are available
AMD Ryzen AI Max+ 395
AMD's Strix Halo APU pairs 16 Zen 5 cores with a Radeon 8060S (40 RDNA 3.5 CUs) and an XDNA 2 NPU on 128 GB of fast unified LPDDR5X. A strong x86_64 box for large-model inference with llama.cpp / Vulkan / ROCm, and for general compute.
Good for
- Quantised inference on 70B–120B-class models (llama.cpp, Vulkan/ROCm)
- x86_64 workloads that need lots of RAM next to a capable GPU
- Agents, RAG pipelines and dev boxes that run 24/7
- Cost-efficient always-on model serving
| Processor | AMD Ryzen AI Max+ 395 (Strix Halo) |
|---|---|
| CPU | 16 Zen 5 cores / 32 threads, 3.0 GHz base, up to 5.1 GHz boost |
| Cache | 16 MB L2 + 64 MB L3 |
| GPU | Radeon 8060S, 40 RDNA 3.5 compute units, up to 2.9 GHz |
| NPU | XDNA 2, up to 50 TOPS (up to 126 TOPS total platform) |
| Memory | 128 GB LPDDR5X-8000 unified, 256-bit (~256 GB/s); large share assignable to the GPU |
| Storage | NVMe SSD — capacity per unit, shown in your portal |
| Networking | Ethernet (2.5 GbE or better) — per unit |
| Software | Ubuntu LTS by default; ROCm, Vulkan, llama.cpp, PyTorch |
| Architecture | x86_64 |
| Power | Configurable 45–120 W TDP |
Sold out Get notified when Ryzen AI Max+ 395 units are available
Racked in the Iron Mountain data centre, Amsterdam
Our machines are hosted in the Iron Mountain data centre in Amsterdam, the Netherlands (EU). Every unit is a dedicated physical machine in our rack there: we handle power, cooling, networking, monitoring and hardware swaps; you work on it remotely over SSH or Tailscale. Your data stays in the Netherlands.
Quick comparison
| DGX Spark | Ryzen AI Max+ 395 | |
|---|---|---|
| Best at | CUDA-native training & inference, FP4/FP8 | x86 workloads, quantised inference via llama.cpp/Vulkan/ROCm |
| Memory | 128 GB LPDDR5x · 273 GB/s | 128 GB LPDDR5X-8000 · ~256 GB/s |
| CPU | 20 Arm cores | 16 Zen 5 cores / 32 threads |
| Arch | aarch64 | x86_64 |
| Software | DGX OS, CUDA | Ubuntu LTS, ROCm / Vulkan |