The ASUS ExpertCenter Pro ET900N G3 is the second GB300 DGX Station we’ve tested and the first we evaluated hands-on in our lab; our MSI XpertStation WS300 testing was conducted remotely. It uses the same NVIDIA silicon seen in the MSI XpertStation WS300: a Grace Blackwell Ultra Desktop Superchip with 72 Grace cores, a B300 GPU, 252GB of HBM3e, 496GB of LPDDR5X, and a ConnectX-8 SuperNIC equipped with dual 400GbE ports. ASUS packages this fixed core platform inside a compact deskside AI tower, adding features absent from the WS300: two top-mounted carry handles, a dedicated fan targeting the ConnectX-8 optics cages, and a 1,600W 80 PLUS Titanium power supply.

ASUS sent the ET900N G3 to our lab for roughly one week of benchmarking, imaging and physical inspection, preconfigured with an RTX PRO 2000. Details covering the base platform, B300, Grace, NVLink-C2C, MIG and DGX Station fundamentals are covered in our earlier MSI XpertStation WS300 review. This analysis focuses on ASUS’s implementation around the Superchip and whether its performance matches the first GB300 tower we tested.

ASUS ExpertCenter Pro ET900N G3 Specifications

Design and Build

ASUS positions the ET900N G3 as a business workstation with brushed silver chassis, dark mesh front panel, chrome trim on caps and feet, plus a low-profile ASUS front badge. Its dimensions are dictated by the internal platform: 584 x 232 x 565mm, slightly taller and marginally narrower than the WS300. At 27kg, it cannot be moved easily, which motivates ASUS’s practical design choice: two top metal handles (front and rear) allow two people to lift the tower onto desks or carts without gripping chassis panels. These handles look out of place on an office workstation until you have transported a heavy GB300 tower across a lab.

The front panel houses the power button, two 10Gbps USB Type‑A ports, two 10Gbps USB Type‑C ports, a USB 2.0 port and separate audio jacks arranged vertically at the mesh upper right. The side panel is a large vented plate, with a long perforated front section for radiator intake and a smaller rear grille aligned to system fans. There is no window or RGB lighting, matching the target buyer’s preferences.

At the rear, the layout separates workstation I/O from underlying server hardware. The upper cluster includes four 10Gbps USB Type‑A ports, 10GbE and 1GbE RJ45 ports, three audio jacks, BMC Micro‑USB console and BMC Mini DisplayPort for setup. The two QSFP112 cages for ConnectX-8 sit atop the I/O shield next to an exhaust fan; three expansion slot covers run down the center, and the C19 inlet power supply sits at the bottom. Like the WS300, the Mini DisplayPort is BMC-only for initial setup and troubleshooting; desktop display output requires the RTX PRO add-in card.

Inside the ET900N G3

With the side panel removed, the ET900N G3 closely resembles the WS300, as expected given both share the same NVIDIA baseboard. The GB300 Superchip sits under a copper cold plate on the chassis left, SOCAMM modules have dedicated copper plates beside it, and ConnectX-8 is cooled by a third cold plate near rear I/O. Braided coolant tubing runs from cold plates to a chassis crossmember manifold, then onward to radiators. Three full-length PCIe slots lie below the Superchip, the power supply occupies the lower front corner, and a metal shroud covers bottom cable routing.

Internal airflow layout mirrors the WS300: one vertical radiator behind the front mesh, a second radiator mounted along the top, each fitted with fan arrays, plus a rear chassis fan. Coolant lines are sleeved and secured with hook-and-loop straps, and the manifold consolidates lines from four cold plates so only two tubes feed each radiator.

The optics fan

One cooling component unique to the ET900N G3 is a small bracket-mounted fan above the ConnectX-8 cold plate, aimed directly at the QSFP112 cages. The liquid loop cools the SuperNIC silicon itself, but 400G optical transceivers generate significant heat inside metal cages that only receive conductive cooling from the cold plate. Our prior testing noted ConnectX-8 runs warm under sustained load and optical modules can get hot, so dedicated airflow over the cages is a valuable design feature for users running continuous fiber connections. When using DAC cables, our setup for an ongoing clustered inference test linking this unit to a second GB300 Station, this fan sees lighter workloads.

Storage, expansion, and power

ASUS shipped the ET900N G3 with one 2TB NVMe drive for OS and NVIDIA software on the pair of PCIe 5.0 x4 M.2 2280 slots connected via Grace. The second pair of M.2 slots taps into ConnectX-8’s integrated PCIe switch and comes empty from factory.

Our review unit arrived with an NVIDIA RTX PRO 2000 Blackwell installed in the PCIe 5.0 x16 slot, one of ASUS’s supported configurations. This card supplies display output without drawing heavily from the GB300 power budget; ASUS states the chassis supports only one RTX PRO Blackwell card. We removed the RTX PRO 2000 before benchmarking to reserve the full accelerator power budget for the Superchip, matching our WS300 test conditions. We did not test higher-power RTX PRO cards in this chassis; their impact on shared power headroom will be covered in a future GB300 Station review.

The 1,600W ATX power supply carries 80 PLUS Titanium certification, an upgrade over the WS300’s Platinum unit. Titanium requires 94% efficiency at 50% load on 115V input versus Platinum’s 92%, and it enforces a minimum efficiency rating at 10% load, so the ET900N G3 wastes fewer wall watts under full GB300 load. Power remains shared: Grace, B300, pumps, fans, storage and any RTX PRO card draw from the same 1,600W pool. NVIDIA’s vsloshd service dynamically shifts available headroom between Superchip and add-in GPU based on power draw, with no fixed hard cap for either component. This power framework is unchanged from the WS300 review. North American deployments require a dedicated 20A electrical circuit.

Connectivity

ConnectX-8’s dual QSFP112 ports serve as the ET900N G3’s primary fabric link to storage, a second DGX Station or cluster networking. In our lab, these ports connect via 400G DACs to another GB300 Station for separate upcoming clustered inference analysis. The Marvell 10GbE port handles regular host traffic, while the Realtek 1GbE port is reserved exclusively for BMC management.

Testing Notes

All results use the lab sample’s stock memory configuration: 252GB HBM3e plus 496GB LPDDR5X. Testing ran without the RTX PRO card installed, granting the GB300 Superchip full accelerator power budget and matching WS300 test conditions. The software stack, vLLM configuration, workload profiles and concurrency sweep are identical, enabling direct side-by-side tower comparison. This review focuses purely on inference performance; we did not measure power draw or thermal telemetry on this unit.

We ran two standard vLLM live-inference profiles used across all local AI hardware: a balanced workload with 512 input / 512 output tokens, and a prefill-heavy workload with 8,192 input / 1,024 output tokens, sweeping concurrency from 1 to 128 streams. For frontier-scale models, we compare against a GPU server with two or four RTX PRO 6000 cards, the closest single-box alternative for these model checkpoints. For smaller models, we benchmark against a single RTX PRO 6000 in the same server, a single H200 NVL in another enterprise server, and a GB10 DGX Spark system identical to the one used in the WS300 review. ET900N G3 metrics are raw log values; comparison figures come from our archived result charts.

Inference Performance

Frontier models on the coherent memory pool

The GB300 is built to host models that straddle the B300’s 252GB HBM3e limit. We tested four checkpoints spanning this boundary: DeepSeek V4 Flash and MiniMax M2.7 fit fully within HBM with leftover capacity; MiniMax M3 barely fits and offloads a subset of expert layers to Grace memory; GLM-5.2 offloads roughly half its weights. Each chart contrasts results against the nearest single-box alternative: RTX PRO 6000 cards inside the enterprise server, offloading experts to host memory when needed.

DeepSeek V4 Flash is a straightforward test: a ~20GB native FP8 checkpoint handled easily by the B300. Balanced workload output throughput rose from 152 tokens/sec at 1 stream to 1,766 tokens/sec at 32 streams, total token throughput hitting 3,532. For the prefill-heavy profile, ET900N G3 scaled from 148 to 954 output tokens/sec, total throughput reaching 8,587. A dual RTX PRO 6000 setup peaked near 800 output tokens/sec on balanced workloads and ~490 on long prompts, with uneven performance at low concurrency.

MiniMax M2.7 in NVFP4 quantization occupies ~125GB HBM and delivered the strongest scaling of the four models. Balanced workload output climbed from 214 tokens/sec on one stream to 4,793 at 128 streams, total throughput reaching 9,586. Prefill-heavy testing scaled to 1,744 output tokens/sec and 15,700 total tokens/sec at 64 streams. Two RTX PRO 6000 cards followed the same general curve but reached less than half the throughput, topping near 1,930 output tokens/sec for balanced workloads and ~510 for long prompts.

MiniMax M3 illustrates operation right at the memory threshold. The NVFP4 checkpoint is smaller than 252GB, but runtime state and KV cache consume additional space, so part of the expert layers reside in Grace memory; every decode step accessing those experts crosses NVLink-C2C. With EAGLE3 speculative decoding, ET900N G3 hit 1,041 output tokens/sec at 32 streams for balanced workloads. The prefill-heavy profile peaked at 282 tokens/sec at two streams, held 271 at four streams, then dropped to 103 at eight streams as long prompts exhausted available cache. Four RTX PRO 6000 cards using the same speculative decoder kept pace through eight streams, pulled ahead at 16 streams to ~1,180 output tokens/sec, then fell to roughly 760 at 32 streams. On prefill-heavy workloads, the four-GPU system maintained ~349 output tokens/sec at four streams versus 271 for the DGX Station. This model sitting exactly at HBM capacity is the only case in this set where 384GB total VRAM across four cards competes with 252GB HBM plus Grace offload.

GLM-5.2 represents a workload only the coherent memory pool enables. Using MTP speculative decoding, the NVFP4 checkpoint keeps ~215GB expert weights in Grace memory. ET900N G3 delivered 35 output tokens/sec for one stream and 139 for 32 streams under balanced workloads, total throughput reaching 277. For prefill-heavy tests up to 32 streams, it hit 120 output tokens/sec and 1,079 total tokens/sec. Four RTX PRO 6000 cards, each offloading ~30GB to host memory, managed ~40 output tokens/sec at 32 streams and flattened to roughly 9–10 tokens/sec from four streams onward for long prompts. Neither system makes this 433GB model feel fast, but the GB300’s 396GB/s LPDDR5X and 900GB/s NVLink-C2C link to offloaded experts maintains scaling where PCIe-attached host memory across four GPUs stalls.

Compared to MSI WS300, ET900N G3 frontier-model results sit within roughly 1% at all matching test points: 1,766 output tokens/sec for DeepSeek V4 Flash at 32 streams on both towers; 4,793 versus 4,801 for MiniMax M2.7 at 128 streams; 1,041 for MiniMax M3 at 32 streams on both; and 139 for GLM-5.2 at 32 streams on both.

Shared models against RTX PRO 6000, H200 NVL, and DGX Spark

The second benchmark group uses smaller models all test systems can fit fully in memory: GPT-OSS-20B and 120B native MXFP4; Llama 3.1 8B BF16 and FP8; Mistral Small 24B BF16 and FP8; Qwen3 Coder 30B BF16 and FP8. New compared to the WS300 review is a single H200 NVL, NVIDIA’s 141GB Hopper accelerator, adding a prior-generation data center GPU to the comparison set.

GPT-OSS-20B is the easiest workload, comfortably loaded by every hardware candidate. Under balanced workloads, ET900N G3 scaled to 22,041 output tokens/sec at 128 streams (44,082 total), versus ~7,920 for RTX PRO 6000, 6,010 for H200 NVL and 1,420 for DGX Spark. Single-stream performance reached 536 output tokens/sec for the Station, compared to 354 on the RTX PRO 6000, 489 on H200 NVL and 120 on Spark. Prefill-heavy runs widened the gap: 9,555 output tokens/sec and 85,994 total tokens/sec at 128 streams for the ET900N G3, against ~2,480 for RTX PRO 6000, 3,410 for H200 NVL and 445 for Spark.

GPT-OSS-120B exposes memory limitations across competing hardware. ET900N G3 hit 9,280 output tokens/sec on balanced workloads at 128 streams, around three times RTX PRO 6000’s 3,060, 2.8 times H200 NVL’s 3,370, and over 19 times Spark’s 480. For long prompts, smaller systems exhausted KV cache capacity early: RTX PRO 6000 peaked near 1,170 output tokens/sec and H200 NVL nearly 1,820, while the Station continued climbing through 128 streams and finished at 5,023 output tokens/sec with total throughput of 45,205 tokens/sec.

Llama 3.1 8B FP8 delivers the highest single throughput number in this test suite. ET900N G3 pushed 25,602 output tokens/sec across 128 streams on balanced workloads (51,205 total), compared to ~8,060 on RTX PRO 6000 and 11,040 on H200 NVL. Switching from BF16 to FP8 boosted Station throughput by roughly 50% (17,074 to 25,602), while RTX PRO 6000 gained ~70% and H200 NVL ~37%. Prefill-heavy results compressed all scores: Station hit 6,164 output tokens/sec, RTX PRO 6000 1,480, H200 2,040.

Mistral Small 24B FP8 produced the largest performance ratios. ET900N G3 peaked at 11,075 output tokens/sec under balanced workloads, 3.4x RTX PRO 6000’s 3,240, 2.4x H200 NVL’s 4,610 and nearly 20x Spark’s 559. On long prompts, RTX PRO 6000 peaked at 32 streams around 480 tokens/sec before declining; H200 NVL topped near 854 tokens/sec, while the Station reached 2,302 tokens/sec at 128 streams.

Qwen3 Coder 30B, a mixture-of-experts model with low active parameter count, narrows single-stream gaps. At one stream balanced workload, RTX PRO 6000 delivered ~232 output tokens/sec versus the Station’s 319 and H200 NVL’s 308. Batching restores the performance hierarchy: at 128 streams FP8, ET900N G3 hit 12,439 output tokens/sec, 3.1x the GPU’s 3,990 and 1.7x H200 NVL’s 7,450. Prefill-heavy runs reached 3,837 output tokens/sec for the Station; RTX PRO 6000 peaked near 880 at 64 streams, H200 NVL near 1,820.

Across these shared models, ET900N G3 delivered 2.8–3.4x throughput of a single RTX PRO 6000 and 1.7–3.7x H200 NVL at full concurrency. Its lead expands on long prompts as competing hardware depletes KV cache headroom. Versus WS300, these four models again sit within 1%: 22,041 vs 22,161 for GPT-OSS-20B; 9,280 vs 9,294 for GPT-OSS-120B; 11,075 vs 11,157 for Mistral Small FP8; 12,439 vs 12,425 for Qwen3 Coder FP8. Full comparison across all 14 models appears in the next section.

Two Towers, One Baseboard: ASUS ET900N G3 vs MSI WS300

ET900N G3 and MSI XpertStation WS300 share the identical NVIDIA GB300 DGX Station baseboard packaged into different OEM chassis. Both ran the same 14 models, two workload profiles and concurrency sweep using matching vLLM builds. The table below compares ET900N G3 peak output throughput against WS300 across all models.

Of 28 total model-workload pairs, 24 results fall within 2% of each other, most showing a fractional percentage advantage for WS300. Four sit outside this band, three on balanced workloads: Llama 3.1 8B FP8 at 3.5% below WS300 (25,602 vs 26,536), Qwen3 Coder 30B BF16 at 4.6% lower (10,000 vs 10,486), Nemotron 3 Ultra 550B at 5.2% lower (159 vs 168), plus Qwen3 Coder 30B BF16 again at 2.1% lower on long prompts. All data points are single-run measurements for each tower, so 2–5% variance across three out of 14 models cannot be attributed to chassis design; we simply document the observed delta. Models dependent on coherent memory pooling, MiniMax M3 and GLM-5.2, match or exceed WS300 performance on both workloads, confirming the Grace offload path behaves identically in ASUS and MSI enclosures.

GPT-OSS-20B, the highest-throughput dense model, and GLM-5.2, the offload-focused workload, show near-perfect overlap: WS300 performance traces ET900N G3 nearly point-for-point.

We will continue this benchmark framework as additional GB300 towers arrive in our lab. HP’s ZGX is next in line, followed by other OEM builds. Every unit runs this identical concurrency sweep, so any differences between OEM implementations will appear in this benchmark framework, with individual model charts retaining raw performance curves.

Conclusion

The ASUS ExpertCenter Pro ET900N G3 fulfills what a second reference-platform implementation should do: it validates the original benchmark results. Across 14 models and two workload profiles, 24 out of 28 peak vLLM throughput comparisons sit within 2% of MSI XpertStation WS300 results, ranging from 1,766 output tokens/sec for DeepSeek V4 Flash up to 22,041 for GPT-OSS-20B. Four single-run outliers fall 2–5% lower, and the system successfully runs GLM-5.2 with 215GB expert layers stored in Grace memory at throughput four RTX PRO 6000 cards cannot match. The GB300 Superchip defines the performance ceiling, and ASUS delivers solid execution.

ASUS’s value additions are practical engineering improvements: the handles let two people move the 27kg tower without gripping chassis panels; the dedicated fan over QSFP112 cages addresses thermal risk for long-duration 400G optical runs; the Titanium-rated PSU reduces wall power draw. The Arm host, mandatory 20A circuit, BMC-only display output and shared 1,600W power budget when large add-in GPUs are installed are inherent platform constraints applying equally to both towers.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!