NVIDIA GB300 Full Cold Plate Liquid Cooling Architecture

As the computational demands of large-scale AI training and inference continue to surge, power densities in AI server racks have pushed well beyond the physical limits of traditional air cooling. Cold plate liquid cooling, with its superior heat transfer capability and compatibility with elevated supply water temperatures, is rapidly becoming the thermal management solution of choice for high-density compute platforms. NVIDIA's GB300 — the latest generation flagship AI computing platform — delivers full liquid cooling coverage within a 1U chassis, supporting a node-level thermal load of 6.9 kW, a cooling flow rate of approximately 8 LPM, and supply water temperatures up to 45°C. These specifications represent the current state of the art in cold plate liquid cooling architecture. This article traces the evolutionary path from GB200 to GB300, examining the core innovations and engineering value of the GB300 full liquid cooling design.

 

 

 

Cold Plate Liquid Cooling: Technology Overview and Engineering Challenges

Cold plate liquid cooling works by attaching a metal cold plate directly to the chip surface, transferring heat to a circulating liquid. Compared to air cooling, this approach can achieve heat transfer coefficients tens of times higher. However, as AI chip power densities continue to climb sharply, cold plate systems face an increasingly complex set of design challenges:

Proliferation of heat sources: Modern AI servers generate significant heat across a wide range of components — GPUs, CPUs, network ASICs, power distribution boards, optical transceivers, and storage devices. A liquid cooling system must manage all of these within a unified architecture.

Serviceability complexity: If the mechanical connections between cold plates and manifolds are not thoughtfully designed, module replacement becomes cumbersome and time-consuming, eroding data center operational efficiency.

Hydraulic resistance matching: In multi-cold-plate circuits with mixed series and parallel paths, the distribution of flow resistance directly affects how effectively each chip is cooled. Poor balancing leads to localized hotspots.

High-temperature supply water compatibility: Data center cooling water temperatures are often constrained by the use of free cooling or other natural cooling strategies. Liquid cooling systems must operate reliably at supply temperatures of 45°C or above.

The GB300 architecture offers a coherent and systematic engineering response to all of these constraints.

 

 

Nvidia GB200 Cold Plate Architecture: A Strong Foundation with Real Limitations

The Nvidia GB200 was the predecessor to the GB300 and represented a leading cold plate liquid cooling design in its time. However, real-world deployment at scale exposed several architectural limitations — limitations that directly informed the redesign efforts behind GB300.

 

GB200 Core Architecture

The GB200 integrated two compute units into a single module. Each module contained two GPU cold plates (B200) and one CPU cold plate (G200), arranged in a topology where the two GPU cold plates were connected in parallel, with their combined flow then routed in series through the CPU cold plate. Key characteristics included:

Two B200 GPU cold plates connected in parallel, with combined flow passing into a single G200 CPU cold plate;

The CPU cold plate also handled heat dissipation for secondary heat sources including LPDDR5 memory and the CX7 network chip, necessitating a physically large cold plate;

Both cold plate modules were equipped with quick-disconnect fittings for rapid fluid circuit interruption;

Cold plates were interconnected using rigid copper tubing or corrugated flexible hoses.

 

Key Engineering Shortcomings

While the GB200 successfully validated cold plate liquid cooling at production scale, several architectural weaknesses emerged:

Serviceability bottleneck: Three cold plates (2× GPU + 1× CPU) were bundled as a single unit, meaning that maintenance of any individual chip required disassembling the entire module — a complex, high-impact operation.

Elevated hydraulic resistance: The series-after-parallel topology resulted in significant pressure drop across the node. Under the same pump head, flow rates were constrained, limiting thermal margins.

Oversized and costly CPU cold plate: To accommodate multiple secondary heat sources, the CPU cold plate was engineered as a large monolithic copper block, driving up material costs and manufacturing complexity.

Limited scalability: With no internal manifold, adding new component cold plates in future iterations would require a complete re-evaluation of the system's hydraulic balance — a fundamentally inflexible approach.

These pain points created the impetus for a deep architectural rethink in GB300.

 

 

Nvidia GB300 Cold Plate Architecture: A System-Level Redesign

 

Package Upgrade Unlocks New Flexibility

Nvidia GB300 transitions the B300 GPU to the SXM7 package, which supports direct removal and reinstallation from the motherboard without requiring the entire module to be disassembled. This hardware-level change is foundational: it decouples individual GPU cold plates from one another, making independent design and maintenance of each GPU cold plate possible for the first time.

 

The Manifold Architecture: The Cornerstone of Decoupling and Parallel Flow

The most significant architectural innovation in GB300 is the introduction of an internal inlet/outlet manifold at the node level. Each manifold features 10 pairs of quick-disconnect fittings (NVQD03 type), serving as a centralized distribution and collection hub for all cold plate branches within the node.

This design delivers three fundamental improvements:

True parallel flow: All four GPU cold plates connect independently to the manifold, achieving a fully parallel hydraulic circuit. Compared to GB200's hybrid series-parallel topology, hydraulic resistance is substantially reduced and flow distribution is more uniform.

Independent serviceability: Any cold plate can be disconnected via its quick-disconnect fitting without disrupting the flow to adjacent cold plates. Maintenance shifts from "full module shutdown" to "single cold plate hot-swap."

Built-in expandability: Adding a new component cold plate requires only connecting it to an available manifold port — no global re-routing or hydraulic re-balancing is needed. The system's scalability is fundamentally improved.

The manifold effectively transforms the node's liquid cooling system from a fixed-topology circuit into a configurable, platform-style architecture.

 

CPU Cold Plate: Lightweight Redesign

Nvidia GB300 significantly simplifies the CPU cold plate. Rather than covering LPDDR5 memory and the CX8 network chip directly with an oversized cold plate, GB300 routes heat from these secondary components to the CPU cold plate via heat pipes. The heat pipes have sufficient flexibility to absorb installation tolerances, eliminating the need for a large monolithic copper cold plate block.

The engineering benefits are threefold:

Lower cost: A smaller cold plate requires less copper and reduces machining complexity;

Standardized design: A single-function cold plate is more amenable to standardized mechanical design and manufacturing;

Easier maintenance: Since thermal coupling to adjacent components is achieved via heat pipes rather than physical contact, cold plate removal does not disturb neighboring devices.

 

Full-Component Liquid Cooling Coverage

Nvidia GB300 achieves a qualitative leap in liquid cooling scope. Every high-power component within the node is brought under liquid cooling management:

OSFP transceiver cold plates: The two upper OSFP cold plates are connected in series via flexible hoses. This allows each cold plate a degree of independent float to accommodate the mechanical displacement of transceiver insertion and removal while maintaining effective liquid-cooled contact.

SSD cold plates: An independent liquid cooling circuit covers all eight M.2 SSDs, supporting hot-swap operation. Thermal Interface Material (TIM) is applied to the cold plate surface to minimize contact thermal resistance, and wear-resistant structural features are integrated to withstand the mechanical demands of frequent SSD insertion cycles.

Power Distribution Board (PDB) cold plate: A dedicated cooling circuit maintains thermal stability for the high-power distribution components.

BlueField-3 NIC and lower OSFP shared circuit: The two lower OSFP cold plates are connected in series and then routed into the BF3 cold plate. The BF3's QSFP transceivers are cooled via bent copper tubing embedded in a thermally conductive plate — an efficient design that minimizes piping count while maximizing thermal utilization of the circuit.

The overarching design logic is clear: with the manifold as the central hub and branch-level quick-disconnects as the operational interface, all six core cold plate circuits (4× GPU + 2× CPU) and all auxiliary component circuits are unified into a fully parallel, modular, and highly maintainable liquid cooling architecture.

 

 

Key Technical Specifications and Engineering Significance

The 45°C supply water temperature capability deserves particular attention. Conventional cold plate systems typically require supply temperatures in the 20–35°C range, necessitating dedicated mechanical chiller plants with substantial ongoing energy costs. At 45°C, Nvidia GB300-equipped data centers can rely on free cooling or indirect evaporative cooling in the vast majority of climate conditions — eliminating or significantly reducing compressor-based refrigeration. The impact on PUE (Power Usage Effectiveness) is meaningful, especially for large-scale deployments in temperate and subtropical regions, and directly supports green data center objectives.

 

 

 

Implications for the Future of AI Server Liquid Cooling

The architectural evolution from GB200 to GB300 reflects several defining trends in AI server liquid cooling design:

From chip-level to system-level coverage: The boundary of thermal design is no longer limited to primary compute devices. Every power-significant component in the server is now a candidate for liquid cooling, placing new demands on system-level design precision and plumbing complexity.

From fixed topology to manifold-based platforms: The introduction of the manifold as a structured distribution layer transforms liquid cooling from a one-time design artifact into a reusable platform that can absorb future component changes without wholesale redesign.

From module-level to device-level serviceability: The proliferation of quick-disconnect fittings and independent branch circuits reduces the granularity of maintenance operations from the module level to the individual component level — a shift that aligns naturally with the trend toward automated data center operations.

From chilled water dependency to warm water compatibility: The ability to operate at 45°C supply temperature removes a longstanding barrier between liquid-cooled servers and free cooling infrastructure, a prerequisite for the technology's broad-scale adoption.

 

 

Conclusion

The Nvidia GB300 cold plate full liquid cooling architecture marks a meaningful milestone in the evolution of AI server thermal management. Through the introduction of a manifold-based parallel distribution architecture, comprehensive liquid cooling coverage across all high-power components, a simplified and cost-effective CPU cold plate design, and device-level hot-swap serviceability, GB300 achieves measurable improvements in hydraulic efficiency, thermal performance, maintainability, and system scalability. For data center operators planning or upgrading liquid cooling infrastructure, the architectural principles embodied in GB300 offer a compelling and practical reference framework.

As next-generation AI chips continue to push power envelopes higher, the convergence of immersion cooling and cold plate liquid cooling — alongside deeper integration between server-level thermal systems and data center infrastructure — will define the next frontier of thermal management. GB300 represents a confident and well-executed step in that direction.

 

 

Quick Tip:

Lian Li Liquid Cooling is a global leader in liquid cooling solutions. Backed by its own R&D team, the company has developed specialized liquid cooling solutions for the NVIDIA GB300 series. Its product lineup—ranging from cold plates and liquid-cooled server racks to CDUs and liquid-cooled data center containers—is deeply optimized for the GB300, achieving a PUE of below 1.15 and meeting authoritative international standards such as RoHS, CE, and UL.

 

 

 

FAQ About 

1.Why did Nvidia GB300 introduce an internal manifold rather than continuing with the direct cold plate interconnect approach used in GB200?

In GB200, cold plates were connected directly in series-parallel configurations, meaning that removing or replacing any single cold plate required disconnecting an entire module's plumbing — a high-effort, high-impact operation. The introduction of a node-level manifold in GB300 changes this fundamentally. Every cold plate branch connects to the manifold independently via a quick-disconnect fitting, enabling true parallel flow. The practical benefits are threefold: any individual cold plate can be removed without affecting others, eliminating full-module downtime; the parallel topology eliminates the pressure drop associated with series segments, improving both flow balance and overall thermal margins; and unused manifold ports provide a ready connection point for future component cold plates, eliminating the need to re-engineer the system's hydraulic layout with each product iteration.

 

2.Nvidia GB300 supports a 45°C supply water temperature — what does that mean for data center infrastructure?

Most conventional cold plate liquid cooling systems require supply water in the 20–35°C range, which means mechanical chillers must run continuously to maintain the necessary temperatures, consuming substantial energy. GB300's 45°C supply water compatibility changes the infrastructure calculus significantly. In most climates, data centers can now use free cooling — outdoor air-to-water heat exchangers or indirect evaporative cooling systems — as the primary cooling source without any mechanical refrigeration. This directly reduces PUE, operating costs, and carbon footprint. For large-scale data centers in temperate or subtropical regions, the operational savings over a facility's lifecycle can be substantial, and the alignment with green data center sustainability goals is considerable.

 

3. Nvidia GB300 uses liquid cooling for its SSDs — what are the key deployment considerations compared to conventional air-cooled SSDs?

GB300 dedicates an independent liquid cooling circuit to its eight M.2 SSDs. The cold plates use Thermal Interface Material (TIM) to fill the contact gap and minimize thermal resistance, with wear-resistant structural features to handle the mechanical wear associated with repeated drive insertions. In practice, three areas warrant close attention: first, the TIM must be compressible enough to maintain consistent contact pressure across SSD insertions and removals — any degradation in contact quality risks thermal resistance increases that can push drive temperatures out of spec; second, liquid-cooled SSD swaps should always be performed after isolating and draining the relevant cold plate branch to eliminate leak risk; third, since the SSD cold plate circuit is independently routed through the manifold, its flow rate requirements must be included in the overall manifold hydraulic budget during system capacity planning to ensure balanced flow across all branches.