Skip to main content
AI Learning Resources

Quick Overview

This video is a detailed hardware and benchmark review presented by Harrison Kinsley on the Sentdex channel. It examines the Asus ET900N G3 workstation, an implementation of Nvidia's DGX Station architecture powered by the Grace Blackwell GB300 superchip. The demonstration explores local large language model inference performance, system thermals, and concurrency scaling using high-end open-weight models.

Key Points

  • 1.The Asus ET900N G3 DGX Station combines an Nvidia GB300 GPU with a Grace CPU and an RTX Pro sidecar card to deliver 748 gigabytes of coherent memory in a single workstation chassis.
  • 2.Under a custom DeepSeek-V4.1-Flash deployment recipe, the system achieves cold prefill speeds averaging 44,000 tokens per second, with peaks exceeding 60,000 tokens per second.
  • 3.In tests running GLM 5.3 Flash using QAD quantization, the GB300 matches a four-way RTX Pro 6000 rig in decode speeds while tripling its prefill throughput.
  • 4.Testing Nvidia Nemotron 3.5 Super VL 120B in NVFP4 demonstrated scaling up to 64 concurrent streams, achieving an aggregate generation rate of roughly 7,000 tokens per second.
  • 5.Integrated liquid cooling maintains GB300 thermals at an average load temperature of 61 degrees Celsius and a maximum peak of 75 degrees Celsius while operating under a single 1,600-watt power supply.

Summary

The review opens with Harrison examining the physical build and internal architecture of the Asus ET900N G3 workstation, commonly referred to as the Asus DGX Station. He details its core hardware, which centers on the Nvidia GB300 GPU boasting 252 gigabytes of high-bandwidth HBM3e memory capable of over 7 terabytes per second of bandwidth. This processor connects to an Nvidia Grace CPU featuring 72 Neoverse V2 cores and 496 gigabytes of LPDDR5X memory, creating a coherent memory pool of 748 gigabytes. The chassis also houses an RTX Pro sidecar GPU, NVMe RAID storage, and a ConnectX-8 SuperNIC providing dual 400-gigabit ports for 800-gigabit inter-station networking, all running off a single 1,600-watt power supply.

Moving to software execution, Harrison benchmarks DeepSeek-V4.1-Flash using an offloading recipe from EL8 of the Local Inference Lab. The recipe distributes the model across tiers, placing the base weights and 285 hot experts into the 230 gigabytes of GB300 HBM, assigning 99 cold experts to the RTX Pro sidecar GPU, and keeping 264 gigabytes of n-gram tables and 139 gigabytes of page cache in system LPDDR5X RAM. On a single stream, the setup produces 237 tokens per second for general text and over 400 tokens per second for Python code generation. At 16 concurrent streams, output scales to between 1,300 and 1,900 aggregate tokens per second, while input prefill throughput achieves a consistent 44,000 tokens per second, peaking at 63,000 tokens per second.

Harrison next evaluates GLM 5.3 Flash quantized with QAD, comparing the single GB300 workstation against a dedicated four-way RTX Pro 6000 rig named Andromeda. The decode throughput between the two architectures proves comparable across early concurrency levels, with the quad-card system remaining competitive at higher concurrencies due to its larger combined VRAM pool for KV caching. However, the GB300 achieves dramatically superior prefill numbers, delivering 33,700 tokens per second on an eight-thousand input context compared to 9,400 tokens per second on the quad-card setup, while avoiding the complex PCIe switching and thermal tuning issues typical of multi-card desktop rigs.

The final benchmark focuses on Nvidia Nemotron 3.5 Super VL 120B in NVFP4 precision, highlighting unusual concurrency scaling. In interactive generation tests, a single stream generates between 450 and 522 tokens per second. Stepping up to four concurrent streams yields 1,200 aggregate tokens per second, scaling to 2,800 tokens per second at 16 streams, and reaching between 6,500 and 7,200 aggregate tokens per second across 64 concurrent streams, with individual streams still maintaining over 100 tokens per second. Telemetry recorded across the full benchmarking suite shows the liquid-cooled GB300 idling at 35 degrees Celsius, averaging 61 degrees Celsius under load, and peaking at 75 degrees Celsius, while the sidecar GPU peaks at 55 degrees Celsius.

Hardware Architecture and Power Constraints

The Asus ET900N G3 workstation features an Nvidia GB300 GPU with 252 gigabytes of HBM3e delivering over 7 terabytes per second of bandwidth, paired with a 72-core Grace Neoverse V2 CPU, 496 gigabytes of LPDDR5X memory, an RTX Pro sidecar GPU, and a ConnectX-8 SuperNIC offering up to 800 gigabits per second of station-to-station networking. The full chassis is powered from a single wall outlet via a 1,600-watt power supply, with the GB300 consuming between 200 watts at idle and 1,300 watts at maximum load.

DeepSeek-V4.1-Flash Hybrid Offloading and Prefill Speeds

Using an inference setup developed by EL8 from the Local Inference Lab community, the GB300 hosts 230 gigabytes of active model weights and hot experts, while an RTX Pro GPU acts as a sidecar holding 99 cold experts, and the Grace CPU memory stores 264 gigabytes of n-gram tables alongside 139 gigabytes of page cache. This configuration achieves decode speeds of roughly 237 tokens per second for general text and over 400 tokens per second for code generation on a single stream, while cold prefill throughput reaches an average of 44,000 tokens per second and peaks at 63,000 tokens per second.

Comparison Against Quad RTX Pro 6000 Rigs

Evaluating GLM 5.3 Flash compressed with QAD demonstrates that the GB300 workstation functions as a direct match in decode output to a four-card RTX Pro 6000 rig across low to moderate concurrencies, though larger KV cache headroom on the multi-card build can offer advantages at higher stream counts. The primary operational difference lies in cold prefill rates, where the GB300 processes 33,700 tokens per second at an eight-thousand-token context compared to 9,400 tokens per second on the quad-GPU setup, while eliminating complex multi-GPU PCIe interconnect bottlenecks.

Nemotron 3.5 Concurrency Scaling and Thermal Performance

Running the 120-billion-parameter Nemotron 3.5 Super VL model in NVFP4 reveals extreme scaling characteristics, moving from over 450 tokens per second on a single stream to roughly 1,200 tokens per second across four streams, 2,800 tokens per second at 16 streams, and reaching between 6,500 and 7,200 tokens per second across 64 concurrent streams. Throughout continuous heavy execution, the workstation's liquid cooling loops hold the GB300 GPU to an average load temperature of 61 degrees Celsius with a 75-degree peak, while the sidecar GPU peaks at 55 degrees Celsius.

The Bottom Line

The video establishes that the Asus ET900N G3 workstation delivers exceptional prefill throughput and high-concurrency inference capability that rivals or exceeds complex multi-card custom systems within a single quiet chassis. It demonstrates that combining HBM3e with Grace system memory enables large mixture-of-experts models to run efficiently off ordinary wall power. What remains open is the precise architectural cause behind the unusual concurrency scaling of the Nemotron 3.5 model compared to rival open-weights architectures, as well as the practical feasibility of deploying local multi-agent swarms using this hardware.

FAQ

What is the Asus ET900N G3 DGX Station and what are its core hardware specifications?

The Asus ET900N G3 is a liquid-cooled AI workstation based on Nvidia DGX Station architecture, equipped with an Nvidia GB300 GPU featuring 252 gigabytes of HBM3e, an Nvidia Grace CPU with 72 Neoverse V2 cores, 496 gigabytes of LPDDR5X system memory, an RTX Pro sidecar GPU, and a ConnectX-8 SuperNIC.

How does the Asus ET900N G3 manage power consumption during heavy machine learning workloads?

The entire system is powered from a single wall outlet using a 1,600-watt power supply. The GB300 GPU idles at approximately 200 watts and draws up to 1,300 watts at peak load.

What prefill speeds does the Nvidia GB300 achieve when running the DeepSeek-V4.1-Flash language model?

The GB300 workstation achieves cold prefill speeds averaging around 44,000 tokens per second, with recorded peak measurements reaching up to 63,000 tokens per second.

How does the Nvidia GB300 GPU compare to a four-card Nvidia RTX Pro 6000 system on GLM 5.3 Flash?

Decode speeds between the two systems are roughly comparable across low to moderate concurrencies, but the GB300 provides more than triple the cold prefill speed, reaching 33,700 tokens per second on an eight-thousand-token context compared to 9,400 tokens per second on the quad-card setup.

How does Nvidia Nemotron 3.5 Super VL scale across high concurrency levels on the GB300 workstation?

Nemotron 3.5 Super VL scales from roughly 450 tokens per second on a single stream to 1,200 tokens per second across four streams, 2,800 tokens per second across 16 streams, and peaks between 6,500 and 7,200 aggregate tokens per second across 64 concurrent streams.

What operational temperatures were recorded for the Nvidia GB300 GPU under continuous load testing?

Over more than an hour of continuous testing, the liquid-cooled GB300 idled at 35 degrees Celsius, maintained an average load temperature of 61 degrees Celsius, and reached a maximum peak temperature of 75 degrees Celsius.

Worth watching for

Machine learning engineers, AI infrastructure architects, and technical professionals evaluating dedicated local workstations for high-concurrency LLM inference and agentic workflows.

  • hardware
  • nvidia
  • benchmarks
  • llm
  • workstation
  • asus
Watch on YouTubeMore videos from Sentdex