NVIDIA Cold Plates Explained: How GPU Liquid Cooling Evolved from H100 to Blackwell (B200/GB200)

Reviewed by APALTEK’s thermal engineering team · Last updated: October 2026

Introduction

In 2022, NVIDIA built direct-to-chip liquid cooling into one of its own PCIe data center GPUs for the first time. The pitch was simple: same performance, less power, less rack space. Liquid cooling was an option, and the NVIDIA cold plate was an accessory.

Three or four years later, the picture has flipped. A GB200 NVL72 rack won’t even run without cold plates. And with Vera Rubin, NVIDIA has pulled the fans out of the compute tray entirely.

Along the way, per-GPU power has roughly doubled, and each generation has rewritten what a cold plate looks like, what it covers, how it connects, and what coolant temperature it runs at. For server OEMs, system integrators, and data center engineering teams, following that evolution is the best way to decide how the cold plate on your next project should be built, and who should build it.

This article walks through how NVIDIA cold plate design changed from Hopper to Blackwell to Rubin, from the perspective of a cold plate manufacturer.

What Is an NVIDIA Cold Plate?

An NVIDIA cold plate is a direct-to-chip liquid cooling component mounted on NVIDIA data center GPUs and nearby chips like the Grace CPU. Usually made of copper, it routes coolant through internal channels to absorb chip heat and carry it back to the CDU. On rack-scale Blackwell systems like the GB200 NVL72, it’s required, not optional.

Key takeaways:

• Hopper (H100/H200): Up to 700 W per GPU. Air or liquid, with one cold plate per GPU on HGX boards.
• Blackwell HGX (B200/B300): 1,000–1,100 W per GPU. Air-cooled B200 systems exist, but they need far more rack space than liquid-cooled ones.
• Blackwell rack-scale (GB200/GB300 NVL72): About 1.2–1.4 kW per GPU. Liquid cooling is required, and cold plates are designed with the tray and rack.
• Vera Rubin: 100% liquid-cooled, fanless trays rated for 45°C inlet water, which leaves cold plates a tighter thermal resistance budget.

NVIDIA Cold Plate Evolution at a Glance

NVIDIA’s data center GPUs have evolved along two parallel product lines, and their cold plates have changed in different ways:

• HGX 8-GPU servers: GPUs plug into a baseboard as SXM modules, and the server vendor designs the chassis and cooling.
• Superchip / rack-scale systems: The CPU and GPU share a board, and the system ships as a complete rack.

Power figures are per GPU. Actual limits vary by system vendor. No reliable public figure exists yet for Rubin per-GPU power, so it’s left blank.

The Hopper Era (H100/H200): Liquid Cooling Was Still Optional

It Started with PCIe Cards

NVIDIA’s liquid cooling story began with PCIe cards. At Computex 2022, NVIDIA launched a liquid-cooled A100 PCIe and announced an H100 version to follow. The liquid-cooled card used the same hardware as the air-cooled version but took up a single PCIe slot instead of two. At this stage, the value of a cold plate was saving space and power, not keeping the system alive.

SXM5 and HGX 8-GPU Boards

The H100 SXM5 pushed per-GPU power to 700 W, and the H200 SXM held that line with an official spec of up to 700 W, configurable. On HGX 8-GPU baseboards, server vendors could choose large air-cooled heatsinks or liquid cooling.

A typical liquid-cooled design used one cold plate per GPU, mounted to the SXM module’s mechanical spec. Tubing tied the eight plates into a single assembly that connected to the rack manifold. The plates cooled the GPUs only. NVSwitch chips, memory, and power components mostly stayed on fans.

GH200: CPU and GPU on One Cold Plate Loop

During the same period, the GH200 Grace Hopper Superchip started a second path, with the CPU and GPU sharing a board over NVLink-C2C. In one 2U four-node server, for example, each node’s GH200 superchip is rated at up to about 1,000 W (CPU, GPU and memory combined), and the system uses four superchip cold plate loops.

That gave NVIDIA cold plates a new job: handling several different chips on one board. The CPU and GPU sit at different heights, draw different power, and put their hot spots in different places, yet they have to share flow on the same loop. That challenge set the stage for GB200 cold plate design.

Blackwell HGX (B200/B300): Air Cooling Gets Too Expensive

1,000 W: Air Still Works, but Density Pays the Price

The B200 pushed per-GPU power to 1,000 W. HGX B200 didn’t require liquid cooling, but the gap between air and liquid became hard to ignore. Supermicro is a good example: its air-cooled system needs a redesigned 10U chassis to fit eight 1,000 W GPUs, while the liquid-cooled version fits in 4U with newly developed cold plates and a 250 kW CDU.

That’s 2.5 times the rack height for the same eight GPUs. For data centers that price by the rack and plan around power density, liquid cooling stopped being the green option and became the economical one.

B300: More Power, More Standardized Interfaces

With Blackwell Ultra, per-GPU power on the HGX line kept climbing. In Supermicro’s liquid-cooled HGX B300 system for OCP racks, each of the eight Blackwell Ultra GPUs is rated at up to 1,100 W TDP, and the system uses blind-mate manifold connections and modular GPU/CPU trays.

For cold plates, this stage brought two changes:

• Each plate has to handle higher heat flux, so the base and channels need to be reoptimized.
• The rack’s blind-mate interface now dictates where the plate’s inlet and outlet go and which connectors it uses.

GB200 NVL72: The Cold Plate Becomes a Rack-Scale Component

The Tray Is the Cooling Unit

GB200 NVL72 makes the entire rack the product. Each rack holds 18 compute trays and 9 NVLink switch trays. Every compute tray carries two GB200 superchips, which adds up to two Grace CPUs and four Blackwell GPUs.

NVIDIA contributed the design to OCP, including 1RU liquid-cooled compute and switch trays and a floating blind-mate tray connection that helps liquid quick disconnects line up and mate reliably in the rack.

That changes the cold plate’s role. In the HGX era, a cold plate was one part inside a server. In NVL72, the cold plates, in-tray tubing, quick disconnects, rack manifold, and CDU form one system that has to be designed and validated together.

Superchip Cold Plates: One Plate, Many Chips

GB200 trays use large cold plates that cover both the CPU and the GPU, and some designs extend to NICs and other components. Fewer connectors and a simpler loop are the upside. The cost is a much tougher design problem:

• The chips sit at different heights with different tolerances, so one plate has to press evenly on several contact surfaces at once.
• Heat flux over the GPU runs far higher than over the CPU, so the channels have to target hot spots instead of spreading flow evenly.
• The plates are much larger and heavier, which makes stress harder to manage during assembly, shipping, and service.

Supplier-published GB200 superchip plates, for example, are brazed copper assemblies built in left- and right-hand versions to fit the tray layout.

Rack Power, and Why It Isn’t 100% Liquid Yet

A GB200 NVL72 rack draws in the 120–140 kW range. HPE’s spec sheet lists a rack TDP of 132 kW, with compute trays cooled by cold plates mounted on the CPUs and GPUs.

But GB200 isn’t fully liquid-cooled. Supply chain analysis shows eight fans still sit in each compute tray, so part of the tray’s heat still leaves through air. That points to a clear goal for the next generation: extend liquid cooling to everything.

GB300: A Reported Shift to Per-Chip Cold Plates

GB300 raises per-GPU power to about 1.4 kW. The cold plate change that drew the most industry attention: GB300 reportedly dropped GB200’s large shared plates in favor of a dedicated plate for each GPU, with its own inlet and outlet, along with a reworked quick-disconnect interface.

Accounts from the supply chain differ, though, and some reports say certain configurations went back to a GB200-style design. The safest reading: GB300 cold plate design varies by configuration and ODM, and there’s no single standard answer.

Whichever design a system uses, the impact on cold plate makers points in the same direction:

• Per-chip plates: Each part is smaller, but the part count multiplies. Every connection is a potential leak point, which raises the bar for consistency and leak testing.
• Large shared plates: Fewer parts, but bigger surfaces, harder tolerance stack-up, and more complex channel design.

Cold plate design still depends heavily on the specific system, even within a single NVIDIA platform generation. Off-the-shelf parts rarely cover it.

What Comes Next: Vera Rubin, 45°C Inlet Water, and Fanless Trays

Vera Rubin is the latest step in NVIDIA’s liquid cooling evolution, and it draws about twice as much power overall as Grace Blackwell. It brings three changes that matter most for cold plates.

1. 100% liquid cooling, no fans in the tray. NVIDIA says the Vera Rubin NVL72 compute tray has no cables, fans, or hoses, which cuts tray assembly from hours to about a minute. Every heat-generating component now has to be liquid-cooled, so cold plate coverage expands like never before.

2. 45°C inlet water. Rubin is designed to accept coolant entering at 45°C, which in many climates lets dry coolers replace chillers. But the warmer the inlet, the smaller the margin between coolant and chip junction temperature, and the lower the cold plate’s thermal resistance has to be. That’s pushing designs toward finer channels, such as microchannel cold plates.

3. Simpler loops. NVIDIA rerouted the cooling loop so that several high-power chips share a single inlet and a single outlet, simplifying the tray layout. That means flow distribution between plates now has to be solved inside the tray’s channel design rather than in external tubing.

5 Engineering Shifts Each NVIDIA Generation Forced on Cold Plates

Put these platforms side by side, and five shifts in cold plate design stand out:

1. Heat flux keeps rising. Per-GPU power doubled from 700 W to about 1.4 kW, and local heat flux over the die rose with it. Channels have to get finer and sit closer to the hot spots.
2. Coverage grew from the GPU to the whole tray. HGX Hopper boards used one plate per GPU, rack-scale Blackwell covers CPU and GPU together, and Rubin is fully liquid-cooled. Cold plates keep taking on more components.
3. Warmer inlets mean tighter thermal resistance budgets. As inlet temperatures rise toward 45°C, the same junction temperature limit leaves less thermal resistance budget for the cold plate.
4. Interfaces became a system-level design problem. Quick disconnects, blind-mate connections, and rack manifold specs now drive where the plate’s inlet and outlet go and how it’s built.
5. The reliability bar keeps rising. More plates and more connections mean stricter pressure testing, leak testing, and long-term reliability validation.

If you’re choosing a cold plate for a specific platform, our server GPU cold plate selection guide explains how to turn these shifts into selection criteria.

Beyond NVL72: Custom Cold Plates for HGX, PCIe and Non-Reference Platforms

Cold plates for rack-scale systems like NVL72 typically come from vendors in NVIDIA’s validated supply chain. But many real-world projects don’t use the NVL72 reference design directly:

• AI servers built on HGX 8-GPU baseboards in a chassis the server vendor designs
• Inference servers and workstations that use PCIe GPUs
• Custom platforms whose board layout, chassis height, inlet temperature, or connectors differ from the reference design

Almost all of these projects need custom cold plates built for the specific platform.

APALTEK is a liquid cold plate manufacturer focused on R&D and production. All of our server and GPU cold plates are custom projects, built with brazed copper construction:

• 8-GPU liquid cooling modules. Our AI-8GPU and GPU-OP-8K open-loop modules, for example, were developed for 8-GPU AI server platforms, with quick-disconnect brands selectable to match the platform.
• PCIe GPU cold plates. Low-profile designs built around single-slot thickness and the spacing between neighboring cards.
• Microchannel cold plate capability. For projects with higher heat flux and warmer inlet temperatures.
• Thermal resistance targets set from junction temperature. We work back from the GPU’s maximum junction temperature, inlet temperature, and power to a thermal resistance budget for the cold plate, instead of quoting a standalone number. See our cold plate thermal resistance guide for the method.
• Validation defined by your requirements. Pressure hold, leak testing, thermal cycling, and other test items are agreed with you at the RFQ stage.

FAQ

Does the NVIDIA B200 require liquid cooling?

Not always. HGX B200 8-GPU servers come in both air-cooled and liquid-cooled versions, but air cooling costs far more rack space. In Supermicro’s lineup, for example, the air-cooled system is 10U and the liquid-cooled one is 4U. Rack-scale systems like the GB200 NVL72 do require liquid cooling.

Why did NVIDIA move to liquid-cooled cold plates?

Because per-GPU power climbed past what air can remove efficiently in a dense rack. Liquid cooling also cuts facility resource use: NVIDIA says the liquid-cooled GB200 NVL72 is over 300x more water-efficient than traditional air-cooled architectures.

What coolant inlet temperature does NVIDIA’s Rubin platform use?

NVIDIA says Vera Rubin supports 45°C inlet water, which in many climates removes the need for chillers. The warmer the inlet, though, the lower the cold plate’s thermal resistance has to be.

What’s the difference between an HGX B200 cold plate and a GB200 cold plate?

An HGX B200 cold plate is designed around the SXM module, usually one plate per GPU, with tubing that joins eight plates into an assembly the server vendor integrates. A GB200 cold plate is designed around the superchip and the 1U tray. One plate covers several chips, including the CPU and GPU, and connects to the rack manifold through blind-mate quick disconnects.

What’s the difference between an NVL72 reference cold plate and a custom cold plate?

A reference cold plate follows NVIDIA’s reference design and fits standard NVL72 racks. Custom cold plates are for HGX, PCIe, and non-standard platforms. Any project whose board layout, chassis height, inlet temperature, pressure drop budget, or connector specs differ from the reference design needs one.

Conclusion

From optional liquid cooling on H100, to rack-scale cold plates on GB200, to fully liquid-cooled trays with 45°C inlet water on Rubin, every NVIDIA generation has rewritten the cold plate spec. Higher power, wider coverage, tighter thermal resistance budgets, and more system-level interfaces aren’t going away.

For cold plate suppliers, the real test isn’t whether you have a ready-made part. It’s how fast you can turn a new platform’s specs into a design that’s ready for production and validation.

If you’re developing a liquid-cooled server built on NVIDIA HGX, PCIe GPUs, or a custom board, send APALTEK’s engineering team:

• Target platform and package type
• Power per GPU and maximum junction temperature
• Coolant type and inlet temperature range
• Available flow rate and pressure drop budget
• Chassis height and quick disconnect specs

We’ll get back to you within one business day with a cold plate design assessment and a prototyping plan.

Request a Quote