Quick answer: The NVIDIA DGX Spark is a GB10 Grace Blackwell mini-workstation with 128GB of unified memory at a modest 273 GB/s of bandwidth. That memory lets it hold models a 4090 or 5090 physically cannot, a 70B dense model or a 120B MoE, in one small quiet box. But the low bandwidth means slow token generation on big dense models (reviewers saw only a few tok/s on 70B-class). It launched at $3,999 and has since risen to a $4,699 MSRP. Worth it only if you specifically need that capacity or the NVIDIA dev stack; otherwise a used 3090, a 5090, or a Ryzen AI Max box serves most people better.
NVIDIA pitched the DGX Spark as a personal AI supercomputer you can put on a desk. It shipped on October 15, 2025, and it is a genuinely interesting machine. But “interesting” and “the right buy for running local LLMs” are different questions. Below is the honest split: the real specs, what the giant unified memory makes possible, exactly where the bandwidth hurts, and the three cheaper machines most people should look at first.
The real specs
Here is what you get, verified against NVIDIA’s listings and launch coverage:
| Spec | DGX Spark (GB10) |
|---|---|
| SoC | GB10 Grace Blackwell Superchip |
| CPU | 20-core Arm (10x Cortex-X925 + 10x Cortex-A725) |
| GPU | Blackwell, 6,144 CUDA cores, 5th-gen Tensor Cores |
| Unified memory | 128 GB LPDDR5X |
| Memory bandwidth | 273 GB/s |
| AI performance | up to 1 PFLOP at FP4 (sparse) |
| Storage | 4 TB NVMe |
| Networking | ConnectX-7 |
| Launch price | $3,999 (Oct 15, 2025); MSRP later raised to $4,699 |
Two numbers on that list do the heavy lifting, and they point in opposite directions. The 128GB unified memory is the reason to want one. The 273 GB/s bandwidth is the reason to be careful. And read the “1 petaFLOP” line with a grain of salt: that figure is sparse FP4, roughly 500 dense TFLOPS, a theoretical prefill peak rather than sustained throughput, per LeCompute’s breakdown.
What 128GB unified memory actually buys you
On a normal PC the model has to fit in the GPU’s dedicated VRAM: 24GB on a 4090, 32GB on a 5090. Spill past that and you fall off a performance cliff. The DGX Spark has no separate VRAM. CPU and GPU share one 128GB pool, and the GPU can address most of it directly.
That changes what is loadable. A 70B dense model at a 4-bit quant needs roughly 40 to 45GB (see how much VRAM a 70B model really needs); a 120B-class MoE like gpt-oss-120b needs even more total memory. Neither fits on a single 5090. The Spark holds both with room to spare. Reviewers highlighted exactly this: it democratizes running Llama-3 70B or GPT-OSS 120B locally, work that previously meant a multi-GPU server or a cloud instance. If your bottleneck is “the model will not even load,” 128GB solves it in a box that stays quiet on a desk.
The catch is what happens after it loads.
Where the low bandwidth hurts
Token generation speed is set by memory bandwidth divided by the active model size: to produce each token, the machine streams the relevant weights out of memory. At 273 GB/s, roughly a sixth of an RTX 5090’s bandwidth per LeCompute’s comparison, the Spark is bandwidth-bound on any large dense model. The measured results (attributed to their sources, and your mileage varies with quant, backend, and context):
- ~50 tok/s on a 32B reasoning model, per DGX Spark benchmark write-ups.
- 2.7 tok/s decode on Llama 3.1 70B in FP8, and around 4.6 tok/s on Qwen 2.5 72B / Llama 3.2 90B, per reviewer measurements.
- On prompt processing (prefill), which is compute-bound rather than bandwidth-bound, it is fast: about 1,723 tok/s on a 120B model versus roughly 340 tok/s for a Ryzen AI Max (Strix Halo) box, per LeCompute.
Read that carefully. The Spark ingests a long prompt quickly, then generates the answer slowly on big dense models. A few tok/s is fine for a one-shot reference query and painful for a long draft, which is the tradeoff our guide on usable tokens per second is built to calibrate. This is the same pattern MoE models exploit, only a few billion parameters are active per token, so they sidestep the bandwidth ceiling; that is why the same class of machine can post strong MoE throughput while crawling on dense 70Bs. The mechanism is in our MoE on low VRAM explainer.
One more nuance the single-stream numbers miss: fed many concurrent requests, the Spark’s aggregate throughput climbs sharply (one benchmark reported around 695 tok/s across 256 streams, roughly 120x the single-stream figure). That helps a developer serving a small model to many clients, not a solo user chatting with one big model.
Who should buy something else
For most people running local AI, one of these is the better buy:
| Option | Usable memory | Bandwidth | Holds 70B / 120B MoE | Price ballpark |
|---|---|---|---|---|
| DGX Spark (GB10) | 128 GB unified | ~273 GB/s | Yes / Yes | $3,999 launch, $4,699 now |
| RTX 5090 | 32 GB | ~6x the Spark | No / No | ~$2,000+ |
| Used RTX 3090 | 24 GB | ~936 GB/s | No / No | ~$700 used |
| Ryzen AI Max 395 (128GB) | ~96-110 GB usable | ~256 GB/s | Yes / Yes | ~$2,000 |
- Run models up to ~32B and want them fast? A used RTX 3090 is the value king, and a 5090 versus 4090 is the choice if you want the fastest single card. Both beat the Spark on tokens per second for anything that fits their VRAM, at less money.
- Want to hold huge models cheaply in a quiet box? The Ryzen AI Max+ 395 (Strix Halo) does nearly the same 128GB-unified trick at similar bandwidth for around $2,000, roughly half the Spark’s current MSRP. It is the Spark’s closest rival on the capacity-per-dollar pitch, and it usually wins that fight.
- Want polish and the highest bandwidth in a unified box? A large Mac Studio pushes far more bandwidth than either, at a higher price; see Apple silicon versus NVIDIA for local AI.
The narrow case for the DGX Spark is real but specific: you want a single, silent, low-power machine that holds very large models and you value its full CUDA stack and ConnectX-7 networking because you are prototyping software meant to run on a real NVIDIA DGX server later. That developer-mirroring story is the Spark’s genuine edge over a Ryzen box. For a pure “run big local LLMs at home” goal, the cheaper capacity king or a faster discrete GPU is the smarter spend. For where each budget lands overall, our best GPU by budget guide maps the full field.
The verdict, and the no-hardware detour
The DGX Spark is a well-built machine solving a real problem, holding models too big for consumer VRAM, in a small quiet box with a real NVIDIA stack. It is not a value play. At $3,999, and now $4,699, you pay a premium over a Ryzen AI Max that does the same capacity trick, and you get slower dense-model generation than a discrete GPU that costs less. Buy it for the capacity-plus-CUDA niche, not as a general local-AI box.
And if the whole reason you are pricing a $4,000 machine is an AI companion, know that the floor is far lower: Ember is an uncensored 18+ companion that runs on a single 8 GB NVIDIA card, so that particular goal does not need this class of hardware. If privacy-by-architecture is why you want the box, build local; if you just want the companion tonight, no supercomputer required.