Short answer: choose a used RTX 3090 when your exact workload fits 24GB, you already have a suitable x86 PC, and the price/condition advantage survives a real inspection. Choose DGX Spark 128GB when you need substantially more one-system capacity and your software works on Arm64/DGX OS. Two 3090s do not automatically become one 48GB GPU.
The decision in one minute
- RTX 3090: 24GB dedicated GDDR6X, 936GB/s published bandwidth, 350W reference graphics-card power, active desktop cooling and optional NVLink for supported two-card workloads.
- DGX Spark 128GB: 128GB coherent shared LPDDR5x, 273GB/s published bandwidth, 140W GB10 SoC TDP and a supplied 240W PSU in a complete compact Arm system.
- Used-market split: a low card price is not the whole 3090 system cost. Verify board health, memory thermals, connector/PSU, chassis, host RAM and the workload’s actual multi-GPU behavior.
RTX 3090 vs DGX Spark specifications
| Boundary | RTX 3090 | DGX Spark 128GB |
|---|---|---|
| Product role | Replaceable Ampere desktop graphics card, commonly bought used | Complete compact Grace Blackwell personal-AI system |
| Advertised memory | 24GB dedicated GDDR6X | 128GB coherent unified system memory |
| Published bandwidth | 936GB/s | 273GB/s |
| Published power boundary | 350W reference card power; verify partner board and PSU | 140W GB10 SoC TDP; 240W supplied system PSU |
| Host / software | Separate x86 PC; Ampere CUDA path | 20-core Arm CPU; DGX OS; verify Arm64 packages and containers |
| Multi-GPU | NVLink supported, but device memory remains separate unless software explicitly distributes/places the workload | ConnectX-7 can connect systems, but capacity is not automatically pooled |
| Upgrade / condition | Replaceable card; used condition and remaining life matter | Integrated, non-user-upgradeable memory; buy the intended configuration |
Published bandwidth is not measured local-AI speed. RTX 3090 board power and Spark’s GB10 SoC TDP are also unlike boundaries, not equal wall-power measurements. Compare complete systems under the same workload.
Exact Qwen3.5 file-size gates
This is file arithmetic, not a successful-load guarantee. RTX margins are against dedicated device memory; Spark margins are against shared system memory that must also serve the CPU, OS and runtime.
| Exact artifact | Listed files | RTX 3090 24GB | DGX Spark 128GB |
|---|---|---|---|
| Qwen3.5-27B community Q4_K_M + F16 projector | 17.668GB | 6.332GB file-only device margin | 110.332GB nominal system margin |
| Qwen3.5-27B community Q8_0 + projector | 29.523GB | No file-size fit | 98.477GB nominal system margin |
| Qwen3.5-27B official BF16 | 55.563GB | No one-card fit; two-card execution requires explicit partitioning and per-device validation | 72.437GB nominal system margin |
| Qwen3.5-35B-A3B community Q4_K_M + projector | 22.915GB | 1.085GB unsafe boundary | 105.085GB nominal system margin |
| Qwen3.5-35B-A3B official GPTQ-Int4 | 24.420GB | No file-size fit | 103.580GB nominal system margin |
| Qwen3.5-122B-A10B official GPTQ | 78.864GB | No one- or two-card file-size fit | 49.136GB nominal system margin |
| Qwen3.5-122B-A10B FP8 | 127.163GB | No one- or two-card file-size fit | 0.837GB unsafe shared-memory boundary |
Use the exact 27B, 35B-A3B and 122B-A10B guides to preserve artifact provenance and runtime/cache caveats.
Two RTX 3090s are not one 48GB device
Two cards provide 48GB aggregate device-local capacity, but each still has a 24GB physical pool. NVLink can improve peer-to-peer transfer for supported software; it does not make every runtime see one automatic 48GB allocation. Record tensor/pipeline/expert placement, layers per device, peer access, CPU offload, interconnect use and peak memory on each card. A model that needs a single allocation above 24GB can still fail.
Used RTX 3090 qualification checklist
- Confirm exact board, 24GB visible memory, serial/firmware and seller return policy.
- Inspect connector condition, fans, cooler mounting, corrosion and modification history.
- Run sustained memory tests and record temperatures, clocks, errors and throttling.
- Verify PSU capacity, transient tolerance, cabling, case clearance and airflow.
- For two cards, verify slot spacing, motherboard lanes, bridge compatibility and per-card cooling.
- Price host CPU, RAM, storage, second card, bridge, PSU and cooling—not only the cheapest listing.
Choose the platform, then measure
Compare the RTX 3090 product page, DGX Spark guide, 3090 vs 4090 vs 5090 guide, DGX Spark vs RTX 4090 and DGX Spark alternatives. Use the workload test and hardware recommender.
Primary sources
- NVIDIA GeForce RTX 3090 specifications
- NVIDIA RTX 3090 family specifications
- NVIDIA DGX Spark specifications
- NVIDIA DGX Spark porting guide
Specification and artifact evidence checked October 25, 2026. Used hardware condition is listing-specific. HighVRAM has not converted peak specifications into performance claims.