DGX Spark & N1X
128 GB of unified memory in a desktop. The first system that makes 70B+ models a routine local workload.
What Is the DGX Spark?
The NVIDIA DGX Spark is a compact desktop AI computer — roughly the size of a Mac mini — built around the GB10 Grace Blackwell Superchip. It ships with 128 GB of unified LPDDR5X memory shared between its ARM-based CPU cores and its Blackwell GPU, connected at over 900 GB/s of internal bandwidth. It runs Linux natively and is purpose-built for local AI development and inference.
GB10 Grace Blackwell Superchip
The silicon at the heart of DGX Spark. Combines a 20-core ARM Neoverse CPU (Grace) with a Blackwell-generation GPU (the B10) on a single package using NVLink-C2C for chip-to-chip bandwidth that far exceeds what PCIe offers between a CPU and discrete GPU.
128 GB Unified Memory
A single flat pool of LPDDR5X memory — no VRAM ceiling, no system RAM vs GPU RAM distinction. The GPU can access all 128 GB at full bandwidth. Models, KV cache, OS, and active applications all share the same pool seamlessly.
FP4 Tensor Cores
Blackwell introduces FP4 precision — 4-bit floating point — doubling the throughput of FP8. For inference workloads where FP4 quantization is acceptable, this enables extraordinarily high token throughput even on desktop hardware.
NVLink for Two-Node Clusters
Two DGX Spark units can be connected via NVLink to form a 256 GB unified memory system — a local supercomputer capable of running 200B+ parameter models at full precision without quantization.
Why 128 GB Unified Memory Changes Everything
When running models on an RTX GPU, you constantly wrestle with the VRAM ceiling. A 70B model at FP16 needs about 140 GB — it simply doesn't fit in 24 or even 32 GB, so you either quantize aggressively (losing quality) or offload layers to slower CPU RAM (losing speed).
On DGX Spark, the question changes from "does it fit?" to "how fast does it run?" Consider what 128 GB unified memory enables:
- Llama 3.1 70B at BF16 (140 GB): fits with two nodes or barely with FP8
- Llama 3.1 70B at FP8 (~70 GB): fits with 58 GB to spare for KV cache and OS
- Llama 3.1 70B at Q4 (~40 GB): fits with 88 GB spare — enormous KV cache for very long contexts
- Multiple models loaded simultaneously: load a 13B chat model and a 70B reasoning model at once, switch between them instantly
What Is the N1X?
The N1X is the name for NVIDIA's GB10 chip when it appears inside computing products made by NVIDIA's OEM and system builder partners — not inside NVIDIA's own DGX Spark system. The underlying silicon is identical; the N1X designation simply signals that a partner is building a product around the same superchip under a different brand or system configuration. If you see an AI PC or workstation from a third-party manufacturer featuring the "N1X" chip, it contains the same Grace Blackwell architecture as DGX Spark.
Ideal Use Cases for DGX Spark
70B–200B Inference
Run Llama 3.1 70B, Mistral Large, or Command R+ at FP8 or Q4 — fully locally. The memory bandwidth and Tensor Cores deliver 30–50 tokens/sec on 70B models, comparable to a mid-tier cloud API for most tasks.
Fine-Tuning
LoRA and QLoRA fine-tuning on 7B–30B models is practical on DGX Spark. The 128 GB pool provides room for both model weights and optimizer states simultaneously, eliminating the gradient checkpointing tricks needed on constrained VRAM.
Long-Context Reasoning
Models with 128K context windows need enormous KV caches — often 10–30 GB for a 70B model at full context. DGX Spark has the memory to sustain this without degradation, enabling true long-document reasoning and large codebase understanding.
Production RAG Pipelines
Run a large embedding model, a full-size reranker, and a 70B LLM simultaneously in a single machine — a setup that would require 3+ discrete GPUs otherwise. Ideal for enterprise knowledge base deployments.
Two-Node NVLink Clusters
Link two Spark units via NVLink-C2C to create a 256 GB unified memory pool. Enables inference on the largest open-weight models — Llama 3.1 405B at Q4, Mixtral 8x22B at BF16 — from a desk.
Multimodal & Vision Models
90B vision models (Llama 3.2 Vision 90B, Qwen2-VL 72B) require 50+ GB — above RTX capacity but comfortably within Spark's pool. Analyze high-resolution images and complex multi-page documents with large VLMs locally.
DGX Spark vs RTX 5090: Head to Head
| Attribute | RTX 5090 | DGX Spark |
|---|---|---|
| Memory capacity | 32 GB GDDR7 | 128 GB Unified LPDDR5X |
| Memory bandwidth | ~1,792 GB/s | ~900 GB/s (CPU↔GPU unified) |
| Max model (unquantized) | ~17B (FP16) | ~65B (FP16) / 130B (FP8) |
| 70B inference speed (Q4) | Requires offloading (~5–10 t/s) | ~30–50 t/s native |
| Small model speed (8B Q4) | ~120–180 t/s (faster) | ~80–120 t/s |
| Fine-tuning (7B LoRA) | Possible with gradient checkpointing | Comfortable, no tricks needed |
| Multi-model loading | Limited by 32 GB ceiling | Load 3–4 large models simultaneously |
| NVLink multi-node | Not supported | 2-node: 256 GB unified pool |
| Price tier | ~$2,000 (GPU only) | ~$3,000–4,000 (complete system) |
| Best for | Fast inference on models < 30B; image/video gen | 70B+ inference; fine-tuning; long context; RAG |
RTX GPUs deliver the fastest tokens/sec for any model that fits in their VRAM — the raw memory bandwidth of GDDR7 gives discrete GPUs an edge in pure throughput at smaller model sizes. DGX Spark's advantage is not speed per se, but reach: it removes the "does it fit?" question entirely for models up to ~130B at FP8. Think of RTX as the sprinter and DGX Spark as the workhorse that never hits a ceiling — choose based on the model sizes and use cases that matter most to you.