Start With the Arithmetic: What 8 Cards at 600W Means
At up to 600W per card, eight dual-width GPUs draw 4,800W before counting the dual CPUs, rows of memory and front drive bays — a single 4U/5U chassis now dissipates heat on a scale that once took a dozen ordinary servers in a rack. The question a cooling scheme must answer is therefore concrete: who carries this heat away — chassis fans plus room air conditioning, or cold plates plus a water loop. The answer depends not on preference but on three boundaries: facility conditions, per-rack power budget, and cluster scale.
Two Ways to Stay on Air — and Why Air's Boundary Sits at the Rack
JH Semiconductor's 8-GPU RTX 6000D platform uses independent CPU/GPU airflow ducts: the CPU zone and GPU zone each take their own intake path, so the eight dual-width cards never breathe air pre-heated by the CPUs, and the 4U/5U chassis buys large fan cross-sections. The 10-GPU Thor T5000 platform takes the other route: up to ten dual-die cards across ten PCIe 5.0 x16 dual-width slots, pushed by high-volume air cooling, with four CRPS power supplies in 2+2 or 3+1 redundancy. Both routes show air cooling is not finished at 600W per card — but its boundary sits at the rack, not the chassis. First, density: at several kilowatts per machine, an ordinary rack's power and heat-rejection budget holds only a few units, and the empty U positions become a hidden cost once the cluster scales. Second, where the heat goes: air cooling hands everything to the room's air conditioning, whose energy use rises with load — exactly the margin liquid cooling exploits to push PUE down. Third, load shape: inference and training clusters run sustained full load rather than bursts, so an air-cooled system must be designed for the continuous worst case, with its headroom permanently occupied.
Three Things Cold-Plate Liquid Cooling Buys
The 8-10 GPU liquid-cooled platform uses cold plates with full CPU and GPU coverage, connects to a CDU and the facility water loop, and supports cluster-level supply/return water monitoring. It buys three things: heat from the main sources goes straight into the water loop, bypassing room air conditioning and opening room to lower PUE; the rack power-density ceiling rises, so the same rack positions hold more compute with shorter network distances between nodes; and supply/return temperatures and flow become observable, schedulable cluster parameters — cooling management turns from rules of thumb into data. The cost is equally explicit: the facility must have a water loop and a CDU, which is an infrastructure project, not a server swap.
Who the Lower-Barrier Air-Cooled Platform Is For
The line also carries an 8-GPU air-cooled platform: eight dual- or quad-width blower cards (up to 600W each), twelve PCIe 5.0 slots, twelve LFF bays, on Intel or AMD platforms — positioned precisely for rapid deployment in standard machine rooms and for lowering the liquid-cooling retrofit barrier. It fits three kinds of users: those with no water loop and no near-term retrofit plans; those running single digits to a dozen machines, before rack-density pressure appears; and those who want the workload running now, with the liquid-cooling decision deferred to the next facility plan. For them, forcing liquid cooling means prepaying for a density problem that does not yet exist.
The Dividing Lines
| Route | Deployment precondition | Fit |
|---|---|---|
| 8-GPU air-cooled platform (Intel/AMD) | Standard machine room, no water-loop retrofit | Rapid deployment; small-to-mid-scale inference and vision computing |
| Independent-duct / high-airflow platforms (RTX 6000D, Thor T5000) | Controlled cold aisle; spare rack power and space | Per-chassis GPU density first: LLM inference, AIGC, dense dual-socket inference |
| 8-10 GPU cold-plate liquid-cooled platform | Facility has, or will build, a CDU and water loop | High power density, low PUE, scaled clusters and full-rack delivery |
Run the selection in that order: first whether the facility has water — or is worth bringing water into; then how many machines the rack's power and heat budget can hold; only then the chassis specifications. Projects that reverse the order — pick the machine first, then bend the facility around it — usually pay the most expensive lesson at delivery.