NVIDIA DGX Spark NVDGXSPARK-PB Gold 2025

The DGX Spark hits a ridiculous 99th percentile for RAM and 98th for storage, but its standard GPU score is a disappointing 11th percentile. It's a 1 petaFLOP AI monster for local model training with 128GB of unified memory, and it's completely useless for gaming. Only buy this if you're fine-tuning 200-billion-parameter models and know exactly what the NVIDIA AI stack is.

Let's be clear: our standard benchmark suite is almost the wrong tool to measure the DGX Spark. Its CPU score is underwhelming for the price, falling behind most mid-range desktops. The GPU score is one of the worst we've seen for a system in this price bracket, because it's not built for rasterization or gaming. It's built for tensor math. The 1 petaFLOP of FP4 performance is the only number that matters here, and it's a number that makes most workstations look like pocket calculators. For AI inference and fine-tuning, this thing is a standout, chewing through large language models locally with the help of that ConnectX-7 Smart NIC and 128GB of unified memory.

In real terms, you're getting a system that can handle models up to 200 billion parameters without breaking a sweat. The 10GbE port means feeding it data won't be a bottleneck. For a developer prototyping a new transformer model, the compile and training times will be dramatically faster than any traditional desktop. But try to run a standard Cinebench run or boot up Cyberpunk, and you'll wonder where your twelve grand went. This is a single-purpose machine, and it performs that single purpose exceptionally well.

Stacking the DGX Spark against a Lenovo Legion 34IAS10 or an HP Omen GT22 is like comparing a scalpel to a sledgehammer. Those competitors will obliterate the Spark in any gaming or general productivity benchmark, with GPUs that aren't stuck in the 11th percentile. An ASUS Republic of Gamers GM700TZ will run circles around it in creative apps. But none of them can even load a 200-billion-parameter AI model, let alone train one. The Dell Tower Plus EBT2250 might offer better all-around CPU performance, but it lacks the unified memory architecture and the dedicated AI tensor cores. You're choosing between a machine that does one thing perfectly and machines that do everything else pretty well.

Value is a tricky concept here. The price spread across vendors is a staggering $122,342, from $4,680 to $127,022. That's not a typo. At the low end, if you can actually find it for that price, it's a steal for an AI development box with this much unified memory. At the high end, you're being charged a premium that makes a cloud GPU cluster look affordable. Memory Express Inc. seems to have the more reasonable listing, but you'll need to hunt for the actual street price. For the right developer, the time saved on model iteration pays for the hardware quickly. For anyone else, it's a very expensive paperweight.

The DGX Spark is a weird beast. It's a desktop that thinks it's a data center, packing 128GB of unified memory and a 4TB NVMe drive into a 1.2kg box. Those specs land it in the 99th and 98th percentiles respectively, which is frankly absurd for a mini PC. But the real story is the 1 petaFLOP of FP4 AI performance from the GB10 Grace Blackwell Superchip. This isn't for gaming or spreadsheets. It's a local AI powerhouse for developers who need to wrangle 200-billion-parameter models without renting cloud GPUs by the hour.

But here's the catch: as a general-purpose computer, it's a bit of a mess. The 20-core Arm CPU is a custom beast, but it lands in a mediocre 39th percentile for raw CPU grunt against traditional x86 desktops. And the GPU, while an AI monster, scores a disappointing 11th percentile in our standard graphics benchmarks. This thing is a surgical instrument, not a Swiss Army knife. If your workflow doesn't revolve around the NVIDIA AI stack, you're paying a massive premium for hardware you won't fully use.

Skip this without a second thought if you aren't a developer working directly with large language models or the NVIDIA AI stack. The CPU is a mediocre 39th percentile performer, and the GPU is a real letdown for anything outside of tensor compute, landing in the 11th percentile. Reliability is also a weak spot at the 12th percentile, so this isn't a machine you want to depend on for mission-critical tasks outside its narrow comfort zone. If you just want a fast desktop, a Lenovo Legion or HP Omen will be dramatically faster and cheaper for everyday work.

The DGX Spark is a purpose-built AI development box that makes no apologies for its narrow focus. If you're a developer actively fine-tuning large language models and want to escape hourly cloud GPU fees, this is one of the best local solutions you can buy right now. The 128GB of unified memory and 1 petaFLOP of AI performance are genuinely impressive. But if your work involves anything outside the NVIDIA AI software stack, the weak general CPU and GPU performance make it a terrible investment. Buy it for the tensor cores, or don't buy it at all.