Supermicro Super AI Station Measures 39-42 dBA in ServeTheHome Review

Supermicro's Super AI Station, a GB300 machine that works as a deskside tower or a 5U rack server, matched the other DGX Stations on CPU tests in ServeTheHome's Nov. 5 review, which measured 39-42 dBA at full load.

It is the first independent measurement Voltgrid found of Supermicro's build and the third DGX Station through ServeTheHome's lab, after Gigabyte's on Oct. 6 and MSI's on Oct. 23. The platform underneath is the one Voltgrid's DGX Station review covered on Nov. 1: NVIDIA's GB300 Desktop Superchip, with 252 GB of HBM3e on the Blackwell Ultra GPU and 496 GB of LPDDR5X on the 72-core Grace CPU, 748 GB in all. Supermicro's datasheet rates it at up to 20 PFLOPS of FP4 without a sparsity label; NVIDIA's own figure is 20 PFLOPS sparse and 15 PFLOPS dense. Supermicro says it handles models of up to 1 trillion parameters.

What Sparse and Dense FP4 Figures Mean

Peak FP4 figures count how many four-bit operations a chip could finish per second. NVIDIA's DGX Station table lists 20 PFLOPS with sparsity and 15 PFLOPS without it. Sparsity lets the hardware skip zeroed weights, so the sparse figure is higher but applies only to models pruned into the right pattern. Supermicro's datasheet says "up to 20 PFLOPS of FP4" and gives no label.

A measured result is a third thing. ServeTheHome's more than 3,000 tokens per second at 32 users on Qwen3.8 Flash Next depends on the model, the batch and the software, so it ranks neither a 20 nor a 15. Only compare machines on the same label and the same test.

Same speed, low noise

ServeTheHome found the three DGX Stations effectively tied on the CPU. Single-core results on its AgentSTH suite varied by less than 3%. Supermicro's unit posted the best single-core score yet, then placed last, by a small margin, in an estimated SPEC CPU2026 integer-rate run that the outlet compiled at -O3 rather than with vendor tuning.

On the GPU, the outlet ran Qwen3.8 Flash Next with several tool-call input and output patterns, with and without multi-token prediction (MTP), from 1 to 32 concurrent users. Several scenarios passed 3,000 tokens per second at 32 users, ServeTheHome said, and were still rising when the sweep stopped. That is a different model, workload and concurrency from StorageReview's 9,280 tokens per second on gpt-oss-120b at 128 streams on an ASUS build, so the two figures do not rank the machines against each other.

Noise is where the convertible design shows. Because the box must also work in a rack, Supermicro cools it with server-style front-to-back airflow: a thick front radiator with two 140mm fans and two separate liquid loops, one for the GPU and one for the CPU and ConnectX-8 network chip. ServeTheHome measured 36-37 dBA at idle and 39-42 dBA at full load. The same outlet measured MSI's XpertStation WS300 at 46-49 dBA under full load in October, though neither review states the measuring distance. ServeTheHome's verdict is that the Super AI Station is as fast and as quiet as any DGX Station it has seen.

The outlet and the rack

Every DGX Station draws on a 1,600 W budget, and NVIDIA requires a 20 A circuit. Supermicro meets it with a Seasonic 1600W 80 Plus Titanium supply, the first better-than-Platinum unit ServeTheHome has received in a DGX Station, and its datasheet says the system plugs into a NEMA 5-20 outlet in North America. That differs from HP's ZGX Fury, which HP says needs a dedicated 200-240 V circuit and is not meant for a standard 120 V outlet.

Idle draw stays high. ServeTheHome read about 152 W at idle on the GPU, which is capped at 1.3 kW, and about 78 W on the Grace module, capped at 300 W, and saw the system never idle below 250 W. Its sample shipped with a 70 W RTX PRO 2000 Blackwell card for display output, which did not move most results; ServeTheHome puts the RTX PRO 6000 at 300 W and says it would eat into the AI hardware's budget, and the outlet advises adding a card only when a local desktop is needed.

The rack option matters because of the networking. Linking systems over two 400G links, ServeTheHome moved 784.3 Gbps of payload between the Blackwell Ultra GPUs but only 229.0 Gbps between the Grace CPUs, which reach the ConnectX-8 over a PCIe Gen5 x8 link. ServeTheHome sees a strong case for racks of these linked at 800G, and measured a 5U chassis that fits between rack posts. The tower is 701 mm deep and weighs 88 lb, and it needs feet attached to stand on the floor.

ServeTheHome also listed the server habits: no USB-C ports, two 5Gbps USB-A ports on the front, and no labels on the rear I/O of its early unit. It found two of the four M.2 slots wired at PCIe Gen6 x4 through the ConnectX-8, while Supermicro's datasheet lists four M.2 PCIe 5.0 NVMe slots and NVIDIA's DGX Station page lists 4x M.2 Gen 5.

What is not confirmed yet

Supermicro's US store shows no price, and TechRadar reported on Sept. 1 that Supermicro was selling it for $92,887.47 before tax and shipping. The store showed the ARS-511GD-NB-LCC-01-G2 as in stock and ready to ship on Nov. 8. For scale, Dell lists its GB300 build at $160,195 on Dell.com, and MSI's started from $99,999 at Newegg as a request-for-quote listing, as listed Nov. 4. ServeTheHome gave no wall-power figure under load, and Voltgrid found no second outlet that has measured this build. NVIDIA said on Oct. 21 that DGX Station for Windows is due from partners in December.