Only on models that fit in 32 GB. There, a desktop GeForce RTX 5090 generates text about four to six and a half times faster than an NVIDIA DGX Spark, because its GDDR7 memory moves data about 6.6 times faster (1,792 GB/s against 273 GB/s). Once a model outgrows 32 GB, the tables turn. The DGX Spark's 128 GB of unified memory runs gpt-oss-120b, Llama 70B and 70B-class fine-tuning, which the 5090 either can't load or can only run by spilling into system RAM. It does that from a 240 W power supply, against 575 W for the 5090 card alone. Each one embarrasses the other, just on different halves of the job.
How do the RTX 5090 and DGX Spark compare?
| GeForce RTX 5090 (desktop card) | NVIDIA DGX Spark (GB10) | |
|---|---|---|
| Memory | 32 GB GDDR7, on the card | 128 GB LPDDR5X, unified (CPU and GPU share it) |
| Memory bandwidth | 1,792 GB/s | 273 GB/s |
| Peak AI compute (FP4, with sparsity) | 3,352 AI TOPS | Up to 1 PFLOP |
| Power | 575 W card; NVIDIA requires a 1,000 W system power supply | 240 W power supply |
| gpt-oss-20b, decode / prefill, tok/s (LMSYS, Ollama, batch 1) | 205 / 8,519 | 49.7 / 2,053 |
| Qwen3 32B 4-bit, decode, tok/s | 61.4 (llama.cpp, 4K context) | 9.4 (Ollama, q4_K_M) |
| gpt-oss-120b, decode, tok/s | Roughly 10–30, with MoE layers offloaded to system RAM | 58.7, fully in memory (llama.cpp) |
| Llama 70B 4-bit, decode, tok/s | Doesn't fit on one card | 4.4 (Ollama, q4_K_M) |
| US price | US$1,999 MSRP; street about US$6,500–8,500 (Oct–Nov 2026) | US$4,699 Founders Edition (since Feb 2026) |
Source: LMSYS DGX Spark review (13 Oct 2025, same test rig for both gpt-oss-20b rows); hardware-corner RTX 5090 benchmarks (Oct 2025, llama.cpp Q4_K_XL) and gpt-oss offloading guide (Nov 2025); llama.cpp DGX Spark benchmarks (Feb 2026); Ollama DGX Spark blog (23 Oct 2025); NVIDIA product pages; Tom's Hardware (27 Feb 2026, Spark price; 26 Oct 2026, 5090 price tracker); eBay listing averages via GPU Poet (Nov 2026). Rows from different sources used different engines, so compare within a row.
Why is the RTX 5090 so much faster on smaller models?
When a language model generates text, each new token means reading most of the model's weights from memory. That makes token generation (decode) limited mainly by memory bandwidth, not by compute. The 5090 has about 6.6 times the Spark's bandwidth, and the measured gap tracks it closely: 61.4 against 9.4 tokens per second on Qwen3 32B at 4-bit is about 6.5 times. On gpt-oss-20b, tested by LMSYS on both machines, the 5090 was about 4.1 times faster at decode and prefill alike.
So if your daily models are 8B to 32B, the 5090 is the more responsive machine. Chat replies arrive faster, coding assistants feel snappier, and batch jobs finish sooner. It also serves several people well on small models. One academic study measured 4,844 tokens per second across 64 concurrent requests for gpt-oss-20b on a single 5090 (vLLM, short prompts). We go through where 32 GB starts to pinch in why AI developers still complain about 32 GB.
Where does the DGX Spark win?
Capacity. A model that doesn't fit in GPU memory has to borrow system RAM over PCIe, and that is far slower than the card's own GDDR7. gpt-oss-120b's weights are about 65 GB, so on a 5090 most of its expert layers have to sit in system RAM. Published results range from about 10 to 30 tokens per second, depending heavily on the host's memory. The Spark holds the whole model and generated 58.7 tokens per second in llama.cpp's own benchmark.
- 70B models: Llama 3.3 70B at 4-bit is 42.5 GB and won't fit on one 5090. The Spark runs it, but slowly: 4.4 tokens per second at 4-bit (Ollama) and 2.7 at FP8 (LMSYS, SGLang). It's fine for testing and batch work, but not for interactive chat.
- Fine-tuning: Unsloth lists 41 GB as the minimum for a 70B QLoRA run, beyond any 32 GB card. NVIDIA says the Spark can fine-tune models of up to 70 billion parameters. On a 5090, QLoRA tops out around 32B (26 GB).
- Several models at once: agent workflows that keep a large model, a small one and an embedding model loaded together fill 32 GB quickly. 128 GB gives room.
- Scaling up: NVIDIA supports linking up to four Sparks. Two or three connect with a single QSFP cable, and four need a 200 GbE switch. NVIDIA's own figure for Qwen3 235B across two Sparks is 11.7 tokens per second (vendor claim).
The RTX 5090 is the faster GPU. The DGX Spark is the bigger one. Which matters more depends on whether your model fits in 32 GB.
What about software, noise and desk space?
The Spark ships as a finished system: DGX OS (Ubuntu) on a 20-core Arm CPU, with NVIDIA's AI software stack installed and NVIDIA's own playbooks for common jobs. It measures 150 x 150 x 50.5 mm and weighs 1.2 kg. The trade-off is Arm Linux. Most AI frameworks and NVIDIA containers are fine, but x86-only desktop software and games are not.
The 5090 is a component. You need a PC to put it in, with a 1,000 W power supply, room for a long card and good case airflow. In return you get Windows or Linux on x86, and the same machine plays games and runs creative software. Two software notes: Blackwell GeForce cards need builds made for CUDA 12.8 or later (PyTorch 2.7 or newer), and NVIDIA's GeForce driver licence says the software "is not licensed for datacenter deployment" (version of 25 Feb 2025, section 2.8). That matters if you plan to put the card in a shared server.
What does each one cost to buy and run?
NVIDIA's RTX 5090 MSRP is still US$1,999, but late-2026 street prices have been several times that, and you still need the rest of a high-end PC. The Spark Founders Edition rose from US$3,999 to US$4,699 in February 2026, with NVIDIA citing constrained memory supplies. ASUS's version, the Ascent GX10, launched at a lower US price. We compare it with the 5090 in RTX 5090 vs ASUS Ascent GX10.
Running costs differ too. As an illustration, take a 5090 PC drawing 650 W under AI load and a Spark-class box drawing 200 W, each running 8 hours a day for 250 days a year. At SP Group's Q4 2026 tariff of 31.16 cents per kWh with GST, that's about S$405 a year against about S$125. These are our assumptions, not measurements, but the ratio holds: roughly three times the power for the 5090 system.
Which should an AI developer buy?
- Choose the RTX 5090 if your models are 32B or smaller, you want the fastest responses and image or video generation, you also game or render, or you want to serve a small team on compact models.
- Choose the DGX Spark if you need 70B–120B models, long context, or 70B fine-tuning on one quiet box, or if you're prototyping for NVIDIA DGX or cloud deployment, where its software stack matches.
- Look higher if you need big models and fast responses for many users. That's DGX Station territory, covered in RTX 5090 vs DGX Station GB300.
Our earlier GB10 vs RTX 5090 post covers the memory-versus-power trade-off in general terms, and the DGX Spark buyer's guide covers the Spark-class models sold in Singapore.
Buying in Singapore
SourceIT lists the NVIDIA DGX Spark 4TB and the DGX Spark Founders Edition 4TB at S$9,300 each, and the ASUS Ascent GX10 128GB/1TB at S$10,350 (all listed 6 Nov 2026). Other GB10 systems are in our DGX Spark collection. Desktop RTX 5090 cards are not in stock at the time of writing; ask us about availability. If you want RTX 50 graphics in a ready-built system, the ASUS ROG NUC 16 Edition 20 (S$11,888.40) uses the RTX 5090 Laptop GPU, which has 24 GB and runs at up to 175 W. It is a different chip from the desktop card. Prices change, so ask for a current quotation. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
Is the RTX 5090 faster than a DGX Spark for AI? For models that fit in its 32 GB, yes: about 4–6.5 times faster at generating text in published tests, in line with its roughly 6.6 times higher memory bandwidth. For larger models, the Spark is faster or the only one of the two that can run them.
Can an RTX 5090 run gpt-oss-120b? Only by offloading most of the model's expert layers to system RAM, which published tests put at roughly 10–30 tokens per second depending on the PC's memory. A DGX Spark holds the whole model and ran it at 58.7 tokens per second in llama.cpp's benchmark.
Why is the DGX Spark slow on 70B models? Its 273 GB/s memory bandwidth limits how fast it can read 40–70 GB of weights for each token. Measured decode was 4.4 tokens per second at 4-bit and 2.7 at FP8, which is usable for testing but slow for chat.
Does the DGX Spark run Windows or games? No. It runs DGX OS, a Linux distribution, on an Arm CPU. Use a GeForce PC for games and Windows software.
Which uses less electricity? The Spark. It has a 240 W power supply, while the 5090 card alone is rated at 575 W and NVIDIA requires a 1,000 W supply for the PC around it.
