DGX Spark/Guides/GB10 sustained performance by OEM
Performance · November 2026

GB10 thermal throttling and sustained performance by OEM: DGX Spark vs ASUS GX10 vs Dell, Acer, HP, Lenovo

By Samuel Seidel · Published November 7, 2026 · Sources checked November 7, 2026

No GB10 maker sustains meaningfully more tokens per second than another. In StorageReview's hour-long vLLM runs of GPT-OSS-120B, machines from ASUS, Dell, Acer, Gigabyte and HP landed within 2.6% of each other on the prefill-heavy test at batch 64 (2,707 to 2,778 tok/s), while peak CPU temperature ran from 74.7 °C (Acer Veriton GN100) to 90 °C (Gigabyte AI TOP ATOM). What did cost speed in 2026 was firmware: fan-curve and USB-C power-delivery faults reported on ASUS, Gigabyte and MSI units, and hard power-offs under sustained GPU load that NVIDIA calls a known issue on its own Founders Edition. Sources checked 7 November 2026.

MachinePeak CPUPeak GPUPeak GPU powerGPT-OSS-120B prefill-heavy, batch 64Decode-heavy, batch 64SourceDate
NVIDIA DGX Spark Founders Edition87 to 88 °C80 to 82 °C74 Wnot given as a numbernot givenStorageReview thermal test2026-01-25
ASUS Ascent GX1087.3 °C82 °C69.77 Wabout 2,700 tok/sabout 260 tok/sStorageReview2026-03-03
Dell Pro Max with GB1087.7 °C80 °Cabout 70 W2,707.15 tok/s273.54 tok/sStorageReview2026-03-03
Acer Veriton GN10074.7 °C69 °C69.18 W2,777.56 tok/s292.41 tok/sStorageReview2026-03-03
Gigabyte AI TOP ATOM90 °C81 °C75.54 W2,771.38 tok/s299.41 tok/sStorageReview2026-03-03
HP ZGX Nano G1n77.3 °C69 °C69 W2,773.3 tok/s262.9 tok/sStorageReview2026-04-24
Lenovo ThinkStation PGXnot publishednot published104 to 160 W sustained at the systema few percent ahead of the Founders Edition in many heat-soaked AI testsnot publishedServeTheHome2026-03-10
MSI EdgeXpert MS-C931No thermal review found. One of four units powered off at the start of a 120,000-token prompt after a September firmware update.Petronella2026-09-26

StorageReview's numbers come from its own monitoring script (nvidia-smi and kernel sensors, sampled every second) during vLLM serving of GPT-OSS-120B; it used no thermal probes or external power meter. The ASUS review gives rounded values, hence "about". The Founders Edition GPU-power figure is the ceiling quoted in the Dell review. The Gigabyte review's 90 °C sits above the 87 to 88 °C group range the same publisher gave Gigabyte in its January thermal test. Single units throughout.

Does any GB10 maker sustain more tokens per second?

No, not by more than run-to-run noise in any published test. The clearest evidence is the prefill-heavy column above: 4,096 tokens in and 512 out, the compute-bound case where a throttling GPU would show first. The five StorageReview machines span 2,707.15 to 2,777.56 tok/s at batch 64, a 2.6% spread by my arithmetic, even though Acer's CPU peaked 15 °C below Gigabyte's.

The decode-heavy column spreads wider, from 262.9 (HP) to 299.41 tok/s (Gigabyte), about 14%. The order does not follow temperature: the hottest box posted the highest figure and one of the coolest posted the lowest. StorageReview's own write-ups describe the inference results as closely grouped across Spark systems. ServeTheHome's Lenovo review found the same shape with a different test suite: heat-soaked, the PGX came in a few percent ahead of NVIDIA's DGX Spark in many of its AI tests, winning one of eight and tying three, while cold runs across five GB10 systems showed no consistent pattern.

Single-stream llama.cpp agrees. A user running llama-bench on an ASUS GX10 in llama.cpp discussion #15396 (25 November 2025) reported 57.19 tok/s decode on gpt-oss-120b MXFP4; the maintainers' DGX Spark results on build 7941 (5 February 2026) show 58.72. Different builds, 2.6% apart. That is expected: every machine here has the same 128 GB at 273 GB/s (on the DGX Spark, 256-bit LPDDR5X-8533 per NVIDIA's hardware page), and decode speed tracks bandwidth. The measured-bandwidth guide covers how much of that 273 GB/s engines reach, and the GB10 machines comparison has the spec and price differences I am not repeating here.

Which GB10 machine runs coolest under sustained load?

The Acer Veriton GN100, by 10 to 15 °C, in the only side-by-side test I found. StorageReview ran the Founders Edition, Gigabyte, Dell, Acer and ASUS units at the same time in one room on its thermal test (25 January 2026): GPT-OSS-120B on vLLM in three shapes (256/256, 4,096/512 and 512/4,096 tokens), batch 1 to 128, 24 stages with 30-second cooldowns.

MachineCPU peakGPU peakNVMe peakGPU power, prefill-heavy
Acer Veriton GN10074.6 °C68 °C51.8 °C69.3 W (lowest)
ASUS Ascent GX10a few degrees below the group below80 to 82 °C58 to 63 °C (1 TB Phison drive)not given
NVIDIA Founders Edition, Dell, Gigabyte87 to 88 °C80 to 82 °C58 to 63 °CGigabyte 76.0 W (highest)

StorageReview, 25 January 2026, one unit each, ambient controlled but not stated. Drives differed (ASUS 1 TB Phison, Dell 4 TB Phison, others 4 TB Samsung), so NVMe temperatures are not like for like.

The HP ZGX Nano, reviewed later with the same three-stage method, peaked at 77.3 °C CPU and 69 °C GPU, the second-coolest result. StorageReview's conclusion is that most partners matched NVIDIA's reference cooling rather than beating it, with Acer the exception. StorageReview reported no throttling in those runs, which is why a 13 °C advantage bought no speed. Headroom matters in the conditions reviewers do not test: a 30 °C room, which is the top of NVIDIA's operating range per the noise and cooling guide, dust, two units stacked, or a firmware release with a lazy fan curve.

How much power does each GB10 draw at the wall?

About 22 to 45 W idle and 104 to 228 W under load, depending on the maker and the workload, against a GB10 chip rated at 140 W. NVIDIA's hardware page specifies a 240 W external supply for the Founders Edition, 140 W of it for the GB10 and 100 W for the ConnectX-7, SSD, Wi-Fi and USB-C, and says a different or lower-rated supply may reduce performance.

MachineIdleUnder loadSupplySource, date
HP ZGX Nano G1n36 to 38 Wabout 228 W peak240 W adapterStorageReview, 2026-04-24
Lenovo ThinkStation PGX38 to 40 W104 to 160 W sustained, peaks just under 200 W240 W USB-CServeTheHome, 2026-03-10
Dell Pro Max with GB10around 30 Wnot measured at the wall280 WJeff Geerling, 2025-12-26
DGX Spark Founders Edition22 to 45 Wjust under 200 W fully loaded; 179.6 W just before one power-off240 Wpower guide; Petronella, 2026-09-26

Each figure is one reviewer's unit and meter; ServeTheHome does not name its meter. HP's noise figures are its own ratings, 22 dBA idle and 27.6 dBA under intensive load, quoted by StorageReview; NVIDIA declares 29 dB sound pressure at full GPU load for the Founders Edition. No review measured noise across makers.

Dell's 280 W supply is the only hardware difference here that touches power headroom. Geerling credits Dell's front-to-back airflow with a quieter box that avoids thermal throttling, but he published no clock or temperature trace, and StorageReview's Dell unit still reached 87.7 °C. The GPU itself drew 69 to 76 W at peak in every StorageReview test, far below the 140 W chip rating, so the extra 40 W of supply is margin for CPU and network load, not for faster tokens.

What actually throttles a GB10?

Firmware and power delivery, in every 2026 case with numbers that I found. These are owner reports on the NVIDIA developer forum, not controlled tests, but several include nvidia-smi counters.

The pattern is the useful part. Throttling follows the firmware track more than the chassis, and partner firmware ships on its own schedule: a forum regular noted that partner devices on the public firmware service sat several versions behind the Founders Edition. Petronella's update also moved two numbers in opposite directions. An FP16 burn test fell from 88.6 to 91.4 TFLOPS to 77.6 to 81.3, while Qwen3.8-27B NVFP4 decode rose from 23.1 to 25.0 tok/s. A synthetic test can therefore mislead you about LLM speed after a firmware change; the measured TFLOPS guide goes into compute numbers, and the software update log lists NVIDIA's releases.

Does capping the GPU clock cost tokens per second?

Very little for generation, which is why owners use a cap to stop power-offs. In a June 2026 forum test, one owner ran the same coding benchmark on Qwen3.6-35B-A3B NVFP4 twice: stock clocks between 2,400 and 2,470 MHz gave about 113 tok/s at 43 W of GPU power, and a 2,000 MHz lock gave the same 113 tok/s at 25 W. Another owner measured image generation at 21.52 against 22.16 seconds per iteration, about 3% slower. Decode waits on memory, so a lower core clock barely shows; prefill is compute-bound and should lose more, though nobody in these threads published a prefill number with and without a cap.

# cap the GPU clock (values posted by owners range from 2,100 to 2,200 MHz)
sudo nvidia-smi -lgc 200,2150

# watch clock, power, temperature and the active throttle reasons every second
nvidia-smi --query-gpu=clocks.gr,power.draw,temperature.gpu,clocks_event_reasons.active --format=csv -l 1

# remove the cap
sudo nvidia-smi -rgc

The cap does not survive a reboot unless you set it again at boot. If a unit fails NVIDIA's field diagnostic, an RMA is the better fix. All of this applies to a GB10 you own. A rented GPUwerk Spark is a non-privileged container that takes its driver from the host (see the tenant environment), so clock locks and firmware updates are host settings, not instance settings.

Which GB10 should you buy for sustained work?

For a machine that will run long jobs in an ordinary office, I would buy the Acer Veriton GN100. It matched the others on throughput and ran 13 °C cooler at the CPU in the same room, which is margin you will want in August or with two units side by side. The HP ZGX Nano is my second choice: the next-coolest temperatures, and a reviewer-measured idle of 36 to 38 W and peak of about 228 W. Do not pay extra for any badge expecting more tokens per second: the published data shows none.

Store pages for the two I would pick: Acer Veriton GN100 and HP ZGX Nano G1n. For the ASUS-specific differences, see ASUS Ascent GX10 vs DGX Spark.

FAQ

Does the DGX Spark thermal throttle?

Usually not in reviews, but owners have reported it. StorageReview's side-by-side test saw the Founders Edition peak at 87 to 88 °C CPU and 80 to 82 °C GPU on GPT-OSS-120B without a throttling note. NVIDIA forum threads from 2026 report SW and HW thermal slowdown after embedded-controller firmware updates, and NVIDIA calls hard power-offs above about 90 W of sustained GPU load a known issue, with a lower GPU clock cap as the workaround.

Is the ASUS Ascent GX10 slower than the DGX Spark?

Not measurably in published tests. StorageReview's ASUS review put GPT-OSS-120B at about 2,700 tok/s on its prefill-heavy test at batch 64, within a few percent of Dell, Acer, Gigabyte and HP. A llama.cpp user run on a GX10 gave 57.19 tok/s decode on gpt-oss-120b against 58.72 on the maintainers' DGX Spark, on different builds. The GX10 has its own firmware track, and several 2026 forum reports of throttling and crashes came from GX10 owners.

Which GB10 machine runs coolest?

The Acer Veriton GN100 in StorageReview's tests. Running five GB10 systems side by side on GPT-OSS-120B, StorageReview recorded an Acer CPU peak of 74.6 °C and GPU peak of 68 °C, against 87 to 88 °C and 80 to 82 °C for the Founders Edition, Dell and Gigabyte. The HP ZGX Nano, tested separately in April 2026, peaked at 77.3 °C CPU and 69 °C GPU.

Why does my GB10 shut down under load?

The common reported causes in 2026 were firmware and power delivery, not the chassis. NVIDIA staff called hard power-offs under sustained GPU load above about 90 W a known issue and advised lowering the GPU clock cap. Petronella Technology Group saw two Founders Editions and one MSI EdgeXpert power off at the start of 120,000 to 131,000-token prompts. NVIDIA says a failing Field Diagnostic qualifies for an RMA.

Does locking the GPU clock reduce tok/s on a DGX Spark?

Barely for token generation. One NVIDIA forum user measured about 113 tok/s on Qwen3.6-35B-A3B NVFP4 at stock clocks of 2,400 to 2,470 MHz and the same 113 tok/s with the clock locked at 2,000 MHz, while GPU power fell from 43 W to 25 W. Another saw image generation about 3% slower. Prompt processing is compute-bound, so expect a cap to cost more there.

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

Pick the chassis for your room, not for speed.

GPUwerk's store lists GB10 machines from eight makers, with the Lenovo ThinkStation PGX on sale now.

GB10 machines shipped by DHL Express from Czechia, or hosted in GPUwerk's EU datacenter and rented out for you.

Browse machines Compare all GB10 machines