NVIDIA DGX Spark and partner systems such as ASUS Ascent GX10 combine a GB10 Grace Blackwell Superchip with up to 128GB of coherent unified system memory. That large shared pool can hold workloads that exceed consumer-GPU VRAM, but it is not a 128GB discrete graphics card and capacity alone does not predict speed.
Check current screened systems · Estimate your model memory
DGX Spark specifications that matter for local AI
| Specification | Published DGX Spark configuration | Buying implication |
|---|---|---|
| Processor | NVIDIA GB10 Grace Blackwell Superchip | Integrated Arm CPU and Blackwell GPU; confirm application and container compatibility. |
| System memory | 128GB LPDDR5x coherent unified memory | Shared by CPU, GPU, operating system and applications—not 128GB of dedicated VRAM. |
| Memory bandwidth | 273GB/s | Large capacity does not imply the bandwidth of a high-end discrete GPU. |
| Published tensor figure | Up to 1 PFLOP FP4 | A precision-specific peak specification, not measured tokens per second. |
| GB10 TDP | 140W | SoC thermal figure; not measured wall power for a workload. |
| Operating system | NVIDIA DGX OS | Arm Linux environment; verify your dependencies rather than assuming x86 desktop compatibility. |
| Networking | ConnectX-7 at up to 200Gbps | Enables supported multi-system configurations; it does not automatically pool memory for every application. |
What does 128GB unified memory mean?
The CPU and integrated GPU can access one coherent pool without maintaining separate system-RAM and graphics-card-VRAM copies. That can make much larger model artifacts practical on one compact system. The operating system, runtime, cache, buffers and other applications also use that same pool, so the advertised 128GB is not all available to one model.
For a concrete scale check, raw weights for exactly 70 billion parameters at exactly 8 bits equal 70GB before format overhead, cache and buffers. A 128GB system has more nominal room than a 24GB, 32GB or 48GB GPU, but actual fit still depends on the artifact and runtime.
Capacity is not speed
NVIDIA publishes 273GB/s of memory bandwidth for DGX Spark. By comparison, NVIDIA publishes 768GB/s for RTX A6000, 1,792GB/s for RTX 5090 and 1,792GB/s for RTX PRO 6000 Blackwell Workstation Edition. These are memory-subsystem specifications, not complete application benchmarks, but they show why “more memory” must not be read as “faster generation.”
Measure prompt processing and token generation separately at the same model, quantization, context, batch and runtime settings. Software maturity and kernel support can matter as much as the headline hardware figures.
Arm compatibility comes first
DGX Spark uses an Arm CPU and DGX OS. NVIDIA provides its AI software stack and documents common frameworks, but an x86-only binary, extension, container image or desktop application may not run unchanged. Before buying, list the exact runtime, Python packages, CUDA extensions, containers, device integrations and management tools you require, then verify Arm64 support.
DGX Spark vs RTX 5090
- DGX Spark: favor it when the workload needs substantially more than 32GB, supports the Arm/DGX environment, and values a compact integrated development system.
- RTX 5090: favor it when the workload fits 32GB, depends on a conventional x86 CUDA workstation, and benefits from the discrete card’s much higher published memory bandwidth.
- Neither conclusion is a speed ranking. Compare representative measurements and complete-system cost.
DGX Spark vs Mac Studio
Both are compact systems with unified memory, but they occupy different software lanes. DGX Spark provides NVIDIA’s CUDA-oriented DGX stack on Arm Linux. Mac Studio uses Apple silicon with Metal, MLX and supported macOS applications. Current Mac Studio configurations can scale beyond 128GB, while DGX Spark emphasizes NVIDIA’s development and deployment environment. Choose the software path first, then capacity and measured performance.
Can you connect multiple systems?
NVIDIA documents connecting up to four DGX Spark systems with ConnectX-7 networking and appropriate cabling or a switch. Aggregate installed memory is not proof that an application can use it as one transparent pool. Distributed inference or training requires supported software, model partitioning and network-aware configuration.
ASUS Ascent GX10 differences to verify
ASUS publishes 64GB and 128GB Ascent GX10 configurations based on GB10. Storage and regional configurations vary. Verify the exact memory capacity, SSD, included power adapter, warranty region and DGX OS support rather than assuming every GX10 listing matches NVIDIA’s reference configuration.
Before buying
- Confirm the exact vendor, memory capacity, storage and warranty region.
- Verify Arm64 and DGX OS support for every critical application or container.
- Estimate weights, context, cache, buffers and concurrent sessions.
- Find a representative benchmark with the same model and runtime; do not substitute FP4 peak compute.
- Compare delivered complete-system cost with a CUDA workstation and a compatible Mac Studio.
Current screened DGX Spark and GX10 systems
Browse the screened hardware board for DGX Spark and ASUS Ascent GX10 observations. Listings can be stale, incomplete or misidentified; confirm the current configuration, seller evidence, delivered price and return terms. A marketplace title is not proof of memory capacity or availability.
Related comparisons
For models around 70B, start with the 70B VRAM requirements guide. Compare dedicated tiers in 24GB vs 48GB VRAM, or compare Apple and NVIDIA desktop paths in Mac Studio vs RTX 5090. Submit a reproducible result to the benchmark database if you test a GB10 system.
Reviewed October 24, 2026. Specifications come from NVIDIA and ASUS documentation linked above. HighVRAM did not independently benchmark DGX Spark or Ascent GX10 for this guide. Some marketplace links are affiliate links.
Choosing between 32GB discrete VRAM and 128GB unified system memory? Read RTX 5090 vs DGX Spark for local AI for exact Qwen3.5 file margins, bandwidth and Arm compatibility.
Choosing between compact unified-memory systems? Compare DGX Spark with Mac Studio across 96GB–512GB tiers, CUDA versus MLX and exact Qwen3.5 boundaries.
Comparing the two Blackwell professional paths? Read DGX Spark vs RTX PRO 6000 for 128GB shared system memory versus 96GB dedicated ECC VRAM.
DGX Spark alternatives: Choose a DGX Spark alternative from the actual capacity, software and deployment constraint in the constraint-first comparison hub.
DGX Spark vs H100: Compare this compact 128GB system directly with H100 80GB and H100 NVL 94GB in the exact capacity and platform comparison.
DGX Spark vs H200: Compare this 128GB compact system directly with H200 141GB in the exact 128GB-versus-141GB comparison.
Comparing the compact 128GB system with an older data-center accelerator? Use DGX Spark vs A100 for exact 40GB/80GB/128GB file-fit, bandwidth, Arm64, server, cooling and MIG boundaries.
Comparing Spark with a common 24GB desktop GPU? Read DGX Spark vs RTX 4090 for dedicated-versus-shared memory, x86/Arm64 and exact artifact boundaries.
Comparing Spark with the common used 24GB path? Use DGX Spark vs RTX 3090 for NVLink, condition, x86/Arm64 and exact artifact boundaries.
Evaluating this system for gpt-oss-120b? Use the gpt-oss-120b hardware guide to compare DGX Spark with the official H100 class, 96GB workstation GPUs and multi-GPU paths.