General Technics delivers enterprise scale AI power in a 4U workstation package
The Challenge
How much raw AI compute can genuinely fit in a single workstation and still be practical to install on-premises? And when proprietary data can’t leave the building, how do research teams do serious AI work on-site? General Technics pushed the limits of what’s possible with a system that packs 384GB of VRAM across four GPUs, each running at full PCIe 5.0 x16 bandwidth, with no exotic liquid cooling required.
A GT-designed chassis with CFD-validated airflow, paired with the latest workstation platform, makes it possible to load and train even your largest models on an in-house machine. It’s cooled with ambient air and mounts in a standard rack alongside your other business systems. You no longer have to risk uploading private datasets, or shrink them to fit your hardware. Scroll or click through the tour below to see how it works.
The Architecture
Most multi-GPU systems quietly compromise by forcing cards to share PCIe lanes or run in bifurcated slots. This build eliminates that tradeoff:
- Uncompromised Data Paths: The Intel Xeon w5-3435X and ASUS Pro WS W790E-SAGE SE provide 112 native PCIe 5.0 lanes, running four NVIDIA RTX PRO 6000 Max-Q GPUs at full x16 simultaneously, with lanes to spare for fast M.2 storage, networking, or capture.
- Massive Throughput & Reliability: 384GB of combined GPU memory keeps large models and long context windows entirely resident. Eight channels of DDR5 ECC RDIMM keep the 16-core Xeon feeding the GPUs, while hardware error correction protects week-long training runs from silent bit-flip failures.
- Direct-to-Card Power: A Corsair WS3000 delivers four native 600W 12V-2×6 cables, one per GPU, straight from the supply with no daisy chains or adapters. Each 300W card runs its connector at half its rated load.
Designed to Scale
In this maximum reference configuration, the system typically draws 1500~1700W during inference or training runs, peaking around 1800~2100W under heavier workloads if every component were to run at full load. That’s well within the capacity of the 3000W Corsair supply, which requires a 220/240VAC circuit.
Scaling to a power envelope suitable for 110/120VAC outlets does not mean compromising on performance. Throttling back the GPUs in software to about 75~80% of thier 300W TGP costs only some inference speed. The CPU can also be capped at 150~200W in BIOS. Its main job in AI workloads is supplying PCIe lanes and memory bandwidth to the GPUs, not raw compute. Alternatively, dropping to three 96GB GPUs cuts about 300W and still leaves 288GB of combined VRAM, which is still more than most systems this size offer. Combining some of these options would bring the system within range of a 1600W power supply and a common 120VAC circuit.
Custom Workstations, Built to Order
This reference AI workstation is designed for teams doing serious AI work on their own infrastructure: research groups whose datasets are too large to stage from disk, engineering teams fine-tuning models on proprietary data, and security-conscious organizations whose compliance policies rule out the cloud.
Every workload is different. Model size, concurrency, power availability, and budget all matter. We’ll design a system that fits your workload today and scales with you tomorrow.