DGX Spark vs RTX 5090
This is not a close fight on speed. An RTX 5090 is many times faster than a DGX Spark on anything that fits in its 32 GB of VRAM. The entire question is whether your model fits, and what you do on the days it doesn't.
The verdict, up front
Buy the RTX 5090 if… your models live under ~30 GB, 7B–32B at 4-bit, Stable Diffusion / Flux, video generation, most fine-tuning of small models, or you also game. 1,792 GB/s of GDDR7 versus the Spark's 273 GB/s is a 6.5× bandwidth advantage, and it shows up in the benchmarks below as roughly an 18× lead in tokens per second on the one like-for-like test we have: 575 tok/s against 25–32 tok/s on a Qwen3-class 27B at concurrency 8. Nothing about the Spark closes that gap.
Choose the Spark if… you need to run or fine-tune models that simply will not load on 32 GB, 70B, gpt-oss-120b, Llama 3.3 70B, big MoEs, long-context KV caches, or you want a quiet 240 W appliance with DGX OS and datacenter-identical CUDA rather than a 575 W card in a tower you maintain. Capacity is the product. Speed is not.
Side-by-side specs
| NVIDIA DGX Spark | GeForce RTX 5090 | |
|---|---|---|
| Memory for models | 128 GB unified (≈110 GB usable) | 32 GB GDDR7 |
| Memory bandwidth | 273 GB/s | 1,792 GB/s |
| Architecture | GB10 Grace Blackwell superchip (CPU+GPU coherent) | Blackwell GB202, discrete PCIe 5.0 card |
| Peak AI compute | ~1 PFLOP FP4 | ~3.4 PFLOPS FP4 |
| Power | ~240 W total system, ordinary wall socket | 575 W card; 1,000 W+ PSU for the rig |
| What you get | Complete system: 20-core Arm CPU, NVMe, DGX OS, 200 GbE | A graphics card. You build the rest. |
| Scaling out | Two units link over ConnectX-7 into a 256 GB pool | Multi-GPU possible; lanes, PSU and heat get hard fast |
| Price | $4,699 complete | $1,999 MSRP, but $4,400–$4,900 street in August 2026 |
| Also good for | DGX OS ships a full Ubuntu desktop over HDMI, so it doubles as an Arm workstation; no gaming, no CUDA-on-Windows apps. | Gaming, rendering, video, everything CUDA |
Which models fit where
This table is the whole article, really. Assume 4-bit quantisation and a modest context window.
| Model (4-bit) | ~Weights | RTX 5090 (32 GB) | DGX Spark (128 GB) |
|---|---|---|---|
| Llama 3.1 8B | ~5 GB | Fits, very fast | Fits, slower |
| Qwen3 32B | ~20 GB | Fits | Fits |
| Qwen3 32B + 64K context | ~30 GB+ | Tight, KV cache is the problem | Comfortable |
| Llama 3.3 70B | ~40 GB | No | Fits |
| gpt-oss-120b (MXFP4) | ~60 GB | No | Fits |
| Llama 3.2 90B / Qwen2.5 72B | ~45–50 GB | No | Fits, ~4–5 tok/s |
| Two Sparks linked | 256 GB pooled (~240 GB usable) | — | 200B+ class |
Note the row that catches people out: it isn't only weights. Long contexts and concurrent users inflate the KV cache, and 32 GB fills faster than the parameter count suggests. That's the most common reason a 5090 owner ends up renting something bigger.
The speed gap, with numbers
Published community and vendor benchmarks agree on the shape of this:
- Float16 Cloud and independent testers running Qwen3-class ~27B models with vLLM AWQ-INT4 at concurrency 8 report the RTX 5090 up around 575 tok/s aggregate, against roughly 25–32 tok/s for the Spark with llama.cpp MTP and about 15 tok/s under vLLM speculative decoding. Same model, same quantisation, an order of magnitude apart.
- On gpt-oss-120b: which the 5090 cannot load at all, NVIDIA developer-forum users report the Spark at roughly 52 tok/s with SGLang (MXFP4), with tuned llama.cpp and vLLM builds in the same thread reported near 58 tok/s. That's a usable interactive speed on a 120B model from a 240 W box, and it is the single best argument for the machine.
- On dense 70–90B models (Qwen2.5 72B, Llama 3.2 90B) ProxPC's GB10-vs-5090 testing measured a steady ~4.6 tok/s on the Spark. Dense 70B at 4-bit is where the Spark feels slow, MoE models like gpt-oss-120b suit it far better because only a fraction of weights are read per token.
- Time to first token is the Spark's other weak spot: long prompts on a 70–90B dense model were measured by ProxPC at 133 s (Llama 3.2 90B) and about 187 s (DeepSeek R1 70B) before the first word appears. Know this before you buy.
Total rig cost, not sticker price
The comparison that gets quoted, "$4,699 versus $1,999", hasn't been true for a while. The RTX 5090's $1,999 MSRP is a Founders Edition number; through 2026 street prices have climbed steadily amid the memory squeeze: Tom's Hardware's GPU price tracker showed a best price of $4,381 in early August 2026, a median of $4,699.99 later that month, and the cheapest in-stock card at a reputable retailer at $4,829. Then add the machine around it:
# a realistic 5090 LLM box, mid-2026 RTX 5090 (street) $4,400 - $4,900 CPU + motherboard ~$600 64 GB DDR5 + 2 TB NVMe ~$450 1000W+ PSU, case, cooling ~$400 # ------------------------------------- Total $5,850 - $6,350 → 32 GB of model memory # DGX Spark Complete system $4,699 → 128 GB of model memory
So the honest framing is: at 2026 street prices a 5090 rig now costs more than a Spark, and you are choosing 32 GB that is very fast over 128 GB that is not. If you also game or render, the 5090 rig's dual use tips it decisively; the Spark's Ubuntu desktop is a workstation, not a games machine.
What the Spark is actually for
NVIDIA positions it as a development machine, and that's fair. It runs DGX OS with the same CUDA, the same containers and the same NCCL as a DGX B200, so a fine-tuning script or a vLLM config you validate on your desk deploys to rented datacenter GPUs without a rewrite. A 5090 gets you close, but consumer cards diverge on FP8/FP4 kernel support, memory model and multi-GPU behaviour, and you'll find that out at the worst moment. Add native two-unit clustering over ConnectX-7, near-silent operation and a plug that works in any office, and the picture is coherent, it just isn't "fast".
What we'd tell a friend
- Everything you run is ≤32 GB: buy the 5090. It's not close, and we won't pretend otherwise.
- You need 70B+, big MoEs, or long contexts: the Spark is the cheapest way to hold them locally at all.
- Dense 70B at interactive speed: neither. Rent datacenter GPUs for that workload.
- Not sure: rent a Spark for an hour and load the exact model you care about. Ten minutes of measurement beats a week of forum threads.