Why you can trust Tom's Hardware Our expert reviewers spend hours testing and comparing products and services so you can choose the best for you. Find out more about how we test.
The Spark comes with a hefty 240 W USB-C power supply. Nvidia officially budgets 140 W of this for the GB10 SoC itself and leaves the rest for the ConnectX 7 NIC and other peripheral devices when they’re active.
The Spark idles at about 35 W as a headless system, and at about 40 W with a 4K 160Hz panel connected. This strikes us as a bit high given GB10’s Arm DNA and Nvidia’s mobile power management chops, but given that the typical enthusiast desktop pulls 75 W to 100 W at idle nowadays, it’s downright frugal by comparison.
We saw typical power usage of about 160 W at the wall under a primarily GPU-driven load, which suggests that overall system performance will be reduced somewhat if the CPU, GPU, and ConnectX NIC are all under heavy load at the same time.
In any case, instantaneous power draw isn’t useful for understanding efficiency by itself. Fast chips with high power draw can get a job done quickly, resulting in low overall task energy, while slow chips can consume lots of power over time to get the job done slowly.
To test token generation efficiency, we once again turned to llama.cpp. This time, we set up the test to generate 4096 output tokens from GPT-OSS 120b and logged cumulative power usage from both systems over the course of the benchmark. That lets us calculate the number of joules consumed to produce each token.
With an empty context buffer, both platforms expend about as much power to generate 4096 tokens from GPT-OSS 20B using llama.cpp. The Spark holds a small advantage here, but it’s not significant.
At the extremes of the context sizes we tested, however, Strix Halo takes twice as much time to generate the same amount of output, and so it expends twice as many joules per token as the Spark does to finish the job.
Our Strix Halo system only consumes about 112 W under an LLM inference load, so neither it nor the Spark are going to inflate your power bill or trip breakers when you put them to work. But we think it’s important to challenge the widely held notion that Strix Halo and GB10 are essentially “the same” when the reality is that GB10 is a faster and more efficient platform once you test anything that isn’t zero-context-depth LLM token generation.
The Spark remains quiet while delivering its solid performance. Under an extended image generation load, the loudest noise level we could measure from the Spark was 37.5 dBA at 18” from the system. (For reference, our testing environment has a noise floor of 33 dBA.)
The Spark isn’t silent under load, but it has a relatively neutral noise signature with only a hint of the high-pitched whine typical of small axial fans. It won’t disturb you or others around you when it’s hard at work.
The outer shell of the Spark does get quite hot under stress. A large amount of the heat is concentrated in the rightmost corner of the case, which appears to be the home of a big inductor in the power-delivery subsystem.
That corner of the Spark does get too hot to comfortably touch under load, but since this is a desktop system, it’s not likely that you should need to touch it under full load anyway.
We want to know whether the GB10 SoC itself is adequately cooled, of course. We don’t have the same wealth of system monitoring options on Linux that we do on Windows, but a glance at the nvtop utility suggests the GB10 GPU doesn’t exceed 82 °C under sustained load.
That’s quite a bit higher than the relatively chilly temps we usually see with quad-slot graphics cards that occupy more volume than the Spark itself, but for a compact system of this class, it’s well within acceptable limits.
nvtop also shows typical GPU clock speeds of about 2450 MHz, or about 10% lower than the RTX 5070 Founders Edition desktop card can sustain. All in all, for the types of workloads this system is built for, its achieved performance, thermals, noise levels, and efficiency are commendable.
MORE: Best Graphics Cards
MORE: GPU Benchmarks and Hierarchy
MORE: All Graphics Content
