LLMRequirements.com
All guides

RTX Spark vs DGX Spark vs RTX 5090 for Local LLMs

NVIDIA now sells two products called Spark, and they are not the same thing. The DGX Spark is the $4,699 Linux dev box that has been shipping since late 2025. The RTX Spark, announced at Computex on June 1, is a Windows-on-Arm platform that will power 30 or more laptops and 10 or more desktops starting this fall. Same family of silicon, very different products, and the name collision is already confusing buyers.

Here is the short version. If you want the fastest local tokens money can buy today, none of the Sparks is the answer, the RTX 5090 is. If you want to run 120B-class models on one box today, the DGX Spark and its GB10 siblings already do that, and so does AMD's cheaper Strix Halo. The RTX Spark's pitch is that same big-model capacity in a Windows laptop you can buy at a normal electronics store. Whether that is worth waiting for depends on which machine you were about to buy, so let's take it apart.

NVIDIA RTX Spark promotional render with the RTX Spark and NVIDIA logos above a stylized Windows-on-Arm laptop.
NVIDIA RTX Spark. Image courtesy of NVIDIA.

RTX Spark vs DGX Spark: same chip, different machine

The flagship RTX Spark systems are built on NVIDIA's N1X, a chip that is, on paper, the same superchip as the GB10 inside the DGX Spark: a 20-core Grace Arm CPU (10 Cortex-X925 performance cores plus 10 Cortex-A725), a Blackwell GPU with 6,144 CUDA cores (the same core count as a desktop RTX 5070), fifth-generation tensor cores with FP4 support, and up to 128 GB of unified LPDDR5X connected over NVLink-C2C. NVIDIA quotes about 1 petaflop of sparse FP4 AI compute for both.

The differences are in everything around the chip:

DGX Spark (GB10) RTX Spark (N1X systems)
What it is Compact dev desktop Laptops, mini PCs, desktops from Dell, HP, ASUS, Lenovo, Microsoft, MSI
OS DGX OS (Ubuntu-based Linux) Windows 11 on Arm
Memory bandwidth 273 GB/s (LPDDR5X-8533) Up to about 301 GB/s (LPDDR5X up to 9400 MT/s)
Networking ConnectX-7 200 GbE, clusterable No ConnectX-7
Power envelope Small desktop, about 240 W under load 45 to 80 W chip envelope in laptops
Software stack CUDA on Arm, NIM and TensorRT-LLM pre-baked, vLLM production RTX gaming and creator stack, DLSS 4.5, agentic AI features in Windows
Price $4,699 Founders Edition, in stores now Estimated from about $2,899 for N1X systems, fall 2026
NVIDIA DGX Spark, a small champagne-gold desktop AI computer with a textured metal-foam front panel, shown at an angle.
NVIDIA DGX Spark. Photo: Daniel Lu, CC BY-SA 4.0.

Three of those rows decide real purchases.

First, the operating system. The DGX Spark boots Linux and is sold as a development appliance. The RTX Spark is a consumer Windows PC, and Microsoft says it made kernel-level changes to Windows 11 for this platform. If your local AI life happens in llama.cpp, vLLM and Docker on Linux, the DGX Spark is the known quantity. If you want one machine for local models, games and normal Windows work, that is exactly the gap RTX Spark exists to fill.

Second, the networking. The DGX Spark carries a ConnectX-7 NIC, and two units cluster over 200 GbE to run 235B-class MoE models end to end (four units host 671B-class). RTX Spark systems drop the NIC. There is no published clustering story for them, so the DGX Spark remains the only Spark with an upgrade path beyond one box.

Third, the memory. Both use a 256-bit LPDDR5X bus, but the DGX Spark's spec sheet says 273 GB/s, while NVIDIA quotes up to about 301 GB/s for RTX Spark, which implies a faster 9400 MT/s memory grade. That is a 10 percent bump on the number that actually limits token generation on big models. Welcome, but it does not change the class of machine.

The $1,799 trap: N1 is not N1X

The headline that traveled after Computex was "RTX Spark laptops from $1,799." Treat that number carefully, for two reasons.

The first is that none of it is official. NVIDIA has announced no MSRPs and no retailer has opened preorders. The $1,799 and $2,899 floors come from Morgan Stanley channel checks with PC makers, reported in early June. Until Dell, HP and Microsoft publish configurations and prices, every dollar figure for RTX Spark is an estimate, and we will update this page when real listings appear.

The second reason matters more for local LLMs. The platform ships as two different chips. The N1X is the GB10-class part described above, with 16 memory channels and up to 128 GB. The cheaper N1 is a much smaller chip: 10 or 12 CPU cores, 2,048 or 2,560 CUDA cores, 8 memory channels, and a maximum of 64 GB. The $1,799 machines are N1 machines. Half the memory bus and half the maximum capacity puts them in a completely different class for inference, closer to today's mainstream AI laptops than to a DGX Spark. If you are buying RTX Spark to run big models locally, the configuration that delivers the pitch is N1X with the full 128 GB, and that is the one with the reported $2,899-and-up floor.

What the numbers say today

No RTX Spark has been independently benchmarked yet (systems land in the fall), but the DGX Spark has been on desks since October 2025, ours included, and the N1X is the same compute on a slightly faster memory bus. Powered and cooled properly, expect RTX Spark desktops to land at or a touch above DGX Spark numbers, and thin laptops below them, since a 45 to 80 W envelope cannot sustain what a 240 W desktop sustains.

Here is how the three machines people actually cross-shop compare on our dataset's numbers, all measured single-stream:

Palit GeForce RTX 5090 GameRock, a triple-fan partner graphics card, in a studio product shot.
Palit GeForce RTX 5090 GameRock (a partner RTX 5090 card). Photo: PantheraLeo1359531, CC BY 4.0.
Build Price Memory (usable) Bandwidth 8B dense 30B dense 120B MoE (gpt-oss)
DGX Spark $4,699 128 GB (119) 273 GB/s 90 tok/s 30 tok/s mid-50s tok/s
RTX 5090 build ~$4,500 32 GB (31) 1,792 GB/s 186 tok/s 70 tok/s does not fit
Strix Halo 128 GB from $2,799 128 GB (96) 256 GB/s 50 tok/s 16 tok/s 47 tok/s

Two notes on that table. The 5090's row is why "capacity vs speed" is the whole debate: more than six times the bandwidth of either unified-memory box, but 32 GB means 70B dense models at Q4 do not fit, let alone the 120B class. And the DGX Spark's 120B figure rewards tuning. The stock FP8 build of gpt-oss-120b decodes in the mid 50s on our box; switching to the NVFP4 build with multi-token prediction and an FP8 KV cache pushed the same hardware to 88 to 92 tok/s at short context and 72 to 77 tok/s with 43K tokens loaded. The N1X inherits exactly this software behavior, which cuts both ways: big wins are available, and the defaults will not give them to you.

Prices moved this year too. The DGX Spark launched at $3,999 and NVIDIA raised it to $4,699 in February, blaming memory supply. The 5090's $1,999 MSRP remains a paper number, with real cards between roughly $2,900 and $4,200 and a full build around $4,500. The RAM shortage that pushed both up is the same force that makes a $2,899 N1X laptop estimate feel optimistic to us, but that is a guess, and we flag it as one.

Is RTX Spark worth it for local LLMs?

Our read, by buyer:

The fall reviews will tell us what N1X actually sustains, and we will fold RTX Spark systems into the dataset the day real configurations and prices exist. Until then, pick the model you want to run in the model picker and it will show you which shipping machines handle it, or start from your budget on the best build for local AI page. Both use the same measured numbers as this article.

Sources