Acer Veriton GN100 Review: Coolest GB10 Box, 713.18 tok/s at Batch 64

The Acer Veriton GN100 serves gpt-oss-120b at 69.65 tok/s at batch 1 and 713.18 tok/s at batch 64 in StorageReview's vLLM tests, measured as total token throughput with equal input and output lengths, the same class of speed as every other NVIDIA GB10 box, while running the coolest of the group StorageReview compared.

That is the finding of StorageReview's Mar. 3 review, and it is why the lab's Best Desktops for Local AI leaderboard, updated Oct. 24, 2026, names the GN100 its "Best GB10 Implementation". ITPro's Apr. 20 review agreed on the essentials: its Geekbench 6 run put the GN100 level with the ASUS Ascent GX10, and it found the machine near-silent in use. Acer announced the GN100 on Sep. 3, 2025, from $3,999 in North America. That is still the only US price Acer has published, and we found no current Acer US listing to confirm it.

This review builds on our comparison of the six GB10 boxes sold in the US, which found that speed does not separate them; price, stock, storage and cooling do. The GN100 is the box that wins on cooling and storage. Whether it also wins on price depends on a figure nobody has confirmed this autumn.

What's New in the Acer Veriton GN100

The GN100 is Acer's version of NVIDIA's DGX Spark design. It uses the same GB10 Grace Blackwell Superchip: 20 Arm CPU cores and a Blackwell GPU sharing 128 GB of LPDDR5x memory. StorageReview lists that memory at 273 GB/s, the same figure NVIDIA gives for every GB10 system. Acer rates the box at up to 1 PFLOPS of FP4 AI performance. NVIDIA's own spec page labels that figure as theoretical and dependent on its sparsity feature, a detail Acer's release leaves out.

What Acer changes is the shell and the parts around the chip. The case measures 150 x 150 x 50.5 mm, and Acer quotes a weight under 1.5 kg. ITPro lists it at 1.2 kg against 1.4 kg for the ASUS GX10, making it one of the lightest GB10 boxes. The front is a wall of vertical slats cut by a horizontal accent line. Underneath, StorageReview found a raw cast-metal bottom plate that doubles as a heat spreader, less finished than the painted plates on NVIDIA's Founders Edition and several other makers' units.

The storage is the clearest hardware difference. Acer fits a 4 TB Samsung PM9E1 PCIe Gen5 drive in the single short M.2 2242 slot. StorageReview called it the fastest 2242 drive it had seen. The rest of the spec is standard GB10 fare: an NVIDIA ConnectX-7 network card for linking two units, four USB-C ports, HDMI, Ethernet, Wi-Fi 7 and a 240 W USB-C power adapter. It ships with NVIDIA's Linux-based DGX OS and AI software stack.

What the GN100 Solves

The problem the GN100 removes is memory. A model must fit in memory before it can run, and 128 GB holds models that a single consumer graphics card cannot. NVIDIA says one 128 GB GB10 unit runs models of up to 200 billion parameters. OpenAI's gpt-oss-120b, in its native MXFP4 four-bit format, fits with room to spare, so a developer can serve it at a desk with no cloud account.

Two units go further. Acer says two GN100s linked over ConnectX-7 handle models of up to 405 billion parameters; NVIDIA's own figure for a pair of 128 GB units is 400 billion. Nobody has published independent tests of a linked pair of GN100s.

Speed is the limit. Each new token requires reading the model's active weights from memory, so the 273 GB/s bandwidth caps how fast one conversation writes. That is why the batch-64 totals look large and the single-request figures do not. In StorageReview's decode-heavy gpt-oss-120b run, which is closest to one user waiting for a long answer, the GN100 started at 38.38 tok/s at batch 1 and reached 292.41 tok/s across 64 requests.

How Fast the GN100 Runs Local Models

StorageReview used the vLLM online serving benchmark, which sends concurrent requests to a running server and counts total tokens per second. It ran three workload shapes at batch sizes from 1 to 64. On gpt-oss-120b, the equal input and output test (Equal ISL/OSL) rose from 69.65 to 713.18 tok/s, and the prefill-heavy test, dominated by reading long prompts, rose from 303.61 to 2,777.56 tok/s.

Smaller models posted higher totals than gpt-oss-120b in the same test. gpt-oss-20b scaled from 91.91 to 1,565.62 tok/s in the equal test. Llama 3.1 8B Instruct at FP4 reached 2,834.70 tok/s at batch 64, against 2,262.73 at FP8, so the four-bit format was faster at batch 64 and at batch 1 (77.15 against 55.93 tok/s). Qwen3 Coder 30B A3B at FP8 climbed from 104.64 to 1,273.47 tok/s.

StorageReview also used a GN100 as its DGX Spark stand-in when it reviewed MSI's GB300-based XpertStation WS300 in August. That review counts output tokens only, with different prompt lengths and up to 128 streams, so its numbers sit on another basis from the March sweep. On it, the GN100 peaked at 500 output tok/s on gpt-oss-120b against 9,294 for the WS300, and flattened at 199 with 8,192-token prompts. The gap shows what a desktop with far more memory and bandwidth buys, at a far higher price.

ITPro published no language-model speeds. Its Geekbench 6 scores of 3,106 single-core and 19,192 multi-core sit close to the 3,104 (multi-core is 4.3% below the ASUS) and 20,048 it recorded on the ASUS GX10, which fits the picture of identical silicon.

Pros and Cons

The thermal case rests on StorageReview's one-hour test, which ramped the load through three stages and compared the GN100 with NVIDIA's Founders Edition and boxes from Dell, ASUS and Gigabyte. The GN100's CPU peaked at 74.7°C and its GPU at 69°C during prefill bursts. Its NVMe drive peaked at 56.8°C and its network card at 61°C. GPU power topped out at 69.18 W, slightly below the highest peaks in the group, and StorageReview tied the lower temperatures to that more conservative power tuning. In its reviews of those boxes, the same lab logged CPU peaks of 87.3°C on the ASUS GX10, about 87.7°C on the Dell and 90°C on the Gigabyte.

Storage results follow the Gen5 drive. In NVIDIA's GPU Direct Storage test, which lets the GPU read the SSD without going through the CPU, 16K reads scaled to 6.91 GiB/s at 128 threads with latency nearly flat. Large 1M transfers saturated at about 11.2 GiB/s for reads and 12.2 GiB/s for writes, though writes fell to 8.94 GiB/s at 128 threads.

The cons are mostly shared with the platform. ITPro notes that every GB10 box runs DGX OS, and StorageReview notes every one tops out around 4 TB on a single short M.2 slot. ITPro's caveats were price and the lack of Windows 11, which it said would frustrate buyers who mainly work in Windows applications.

Who the GN100 Is For

The GN100 fits a developer or small team that wants to serve 70B- to 120B-class models locally on NVIDIA's CUDA stack and plans to load large checkpoints often, where the Gen5 drive should help. It also fits buyers who will put the box on a desk next to people, given ITPro's report of near-silent running and StorageReview's thermal headroom. A team that may later pool two boxes for a 400B-class model can start with one.

It does not fit someone who wants a fast personal chat assistant. At 38.38 tok/s on decode-heavy gpt-oss-120b at batch 1, one user waits on the memory, as on every GB10 box. Buyers who need Windows should look elsewhere; ITPro pointed to AMD's Ryzen AI Max+ 395 systems, which run Windows 11.

How It Compares

On StorageReview's gpt-oss-120b Equal ISL/OSL test at batch 64, the March series put the GN100 at 713.18 tok/s, the ASUS Ascent GX10 at 680, the Dell Pro Precision with GB10 at 670.83 and the Gigabyte AI TOP ATOM at 649.62. The HP ZGX Nano reached 722.9 in a separate April StorageReview review with its own run, so it is not ranked against the March series. All share 128 GB and 273 GB/s, so a gap under 10% in one series likely reflects software tuning more than hardware; StorageReview did not say so.

Price is where they part. Acer's only US figure is $3,999 from September 2025. The ASUS GX10 listed at $5,999 with 1 TB on Oct. 30 and showed out of stock on Nov. 2. Notebookcheck reported the HP ZGX Nano at $4,759 on Amazon and $7,399 at the HP Store. Dell.com listed its 128 GB box at $7,967 on Nov. 2, CDW advertised the Gigabyte at $7,312, and NVIDIA's DGX Spark showed $6,950 on its marketplace on Oct. 31. These are different configurations (1 TB, 2 TB and 4 TB) on different dates, and NVIDIA's listing is a cached marketplace page with stock unverified. If Acer's launch price still holds, the GN100 undercuts every current US listing above; that is unconfirmed.

NVIDIA said that since Oct. 23 its six partners, Acer among them, would sell a 64 GB GB10 configuration from $4,999. We found no Acer page for a 64 GB GN100.

What to Watch

The first thing to watch is a dated US price for the GN100 from Acer or a major US reseller, and whether a 64 GB version appears. Independent single-user figures from llama.cpp, Ollama or LM Studio would show what one person gets, since no lab has published them for the GN100. Nor has anyone published its power draw at the wall: StorageReview reports GPU power only. NVIDIA reports its third-quarter fiscal 2027 results on Nov. 17, and its GTC Washington, D.C. keynote is set for Dec. 1, either of which could bring GB10 software or pricing news.