H100 Water Cooling Guide: Liquid Cooling for AI Research GPUs

The NVIDIA H100 is the workhorse GPU of AI research in 2026, powering everything from university lab experiments to small sovereign training clusters. If you have acquired an H100 (or its successor, the H200) for your research lab, small team, or personal training setup, you have likely noticed that the stock cooling solution assumes a data center with forced airflow from front to back. In a desktop workstation, a lab bench setup, or a small multi-GPU cluster without enterprise-grade HVAC, the H100 needs a different thermal strategy. This guide covers water cooling the H100 and H200 for individual researchers and small labs -- not for data center operators, not for rack-level deployments, and not for anyone counting PUE ratios. Just practical thermal engineering for people who need a quiet, reliable H100 that fits in a research environment.

Key Takeaways

• The H100 PCIe variant (80GB HBM3) draws 350W TDP and is the realistic option for individual researchers and small labs.
• The SXM variant (designed for DGX/HGX baseboard mounting) requires a custom mezzanine and NVLink infrastructure that makes it impractical for most non-enterprise builders.
• Bykski produces a full-cover waterblock for the H100 80G PCIe / H800 80G, and an AIO liquid cooling solution for the H200 141G.
• A single H100 PCIe water-cooled build runs quietly at 28-30 dBA -- suitable for a shared office or university lab.
• For 2-4 card builds, plan for 1,000-1,400W of GPU heat: a 480mm + 360mm radiator combination minimum.

Who This Guide Is For

This guide is written for:

• AI researchers at universities who secured H100 hardware through academic programs, grants, or surplus sales and need to cool it outside a server room.
• Individual AI developers and researchers building sovereign AI infrastructure -- training models locally rather than paying cloud compute bills.
• Small teams (2-10 people) running a shared training cluster with 2-4 H100 cards in a workstation or small rack in an office.
• Privacy-focused builders who need on-premises training capability for sensitive data that cannot go to cloud providers.

This guide is NOT for data center operators deploying hundreds of H100s in rack-mount chassis with CDU (coolant distribution unit) infrastructure. That is a different engineering problem with different vendors (EK-Pro, Alphacool ES, etc.) and different budgets.

H100 Variants: PCIe vs SXM

The H100 comes in two main form factors, and the distinction matters for cooling:

H100 PCIe (Practical for DIY)

The PCIe variant fits a standard PCIe x16 slot. It uses a standard dual-slot form factor with a blower-style cooler designed for 1U-4U server chassis with front-to-back airflow. In a desktop workstation, this blower runs at high RPM and produces 55-65 dBA -- usable in a machine room, unbearable in an office.

The PCIe variant has 80GB of HBM3 memory with 3.35 TB/s bandwidth and draws 350W TDP. This is the variant that individual researchers and small labs should target. It fits in standard workstation motherboards, uses standard power connectors, and has a standard PCB layout that Bykski produces a waterblock for.

H100 SXM (Enterprise Only -- Honest Caveat)

The SXM variant is what goes into NVIDIA's DGX H100 and HGX H100 baseboards. It has higher memory bandwidth (3.35 TB/s), an NVLink 4.0 interface (900 GB/s), and draws 700W TDP. The SXM module does not have a PCIe connector -- it plugs into a custom mezzanine board that provides power, data, and NVLink connectivity.

Cooling an SXM H100 outside of the DGX/HGX chassis is technically possible but requires custom adapters, custom power delivery, and custom cooling solutions that are beyond the scope of a practical guide. If you have SXM modules, you almost certainly also have the baseboard they plug into, and the cooling solution is part of that system. We do not recommend attempting to run SXM modules in a DIY workstation.

The Thermal Reality of Sustained Training Workloads

The H100 PCIe at 350W generates significant heat under sustained training workloads. Unlike inference (which can be somewhat bursty depending on batch size), model training sustains maximum power draw for hours, days, or weeks continuously.

The stock blower cooler on the PCIe variant handles 350W thermally but does so by running the blower at 4,000+ RPM. At that speed, a single H100 produces 60-65 dBA of noise at 1 meter. Two or four H100 cards in the same workstation compound the noise problem exponentially -- four blowers at 65 dBA each produce a combined noise level exceeding 70 dBA, which is louder than a vacuum cleaner and well above OSHA's recommended 8-hour exposure limit for non-industrial settings.

In a university lab or shared office, this noise level is not acceptable. Water cooling the H100 drops noise to 28-32 dBA per card, making multi-card builds viable in shared spaces.

Bykski H100 Waterblock and H200 AIO

H100 80G / H800 80G Waterblock

The Bykski H100 80G waterblock is a full-cover block designed for the PCIe variant. Key specifications:

• Material: High heat resistance POM top with full metal (copper/nickel) cold plate construction.
• Coverage: Full cover -- the cold plate makes direct contact with the GPU die, HBM3 stacks, and VRM area. This is critical for the H100 because HBM3 (High Bandwidth Memory) runs hot under training loads and benefits significantly from direct liquid cooling contact.
• Backplate: Full metal backplate included for structural support and passive rear-side cooling.
• Compatibility: Fits the H100 PCIe 80G and the H800 80G variant (the China-market version with reduced NVLink). Verify your exact model before ordering.
• Fittings: Standard G1/4 threads -- compatible with any custom loop hardware.

H200 141G AIO

The Bykski H200 141G AIO is a complete liquid cooling solution for the H200 (141GB HBM3e) that includes the waterblock, radiator, pump, tubing, and fans in a single package. This is the simpler option for researchers who want liquid cooling without building a custom loop. Available in 240mm and 360mm radiator configurations.

Single-Card Loop Plan

For a single H100 PCIe in a workstation or lab bench, a simple custom loop provides the best noise-to-performance ratio:

• Waterblock: Bykski H100 80G full-cover block
• Radiator: 360mm minimum. A FormulaMod 360mm copper radiator handles 350W with fans at 800-1,000 RPM.
• Pump: Barrow D5 pump + reservoir combo at 40-50% PWM -- effectively silent.
• Fittings: G1/4 compression fittings for 10x16mm soft tubing. Budget 8-10 fittings for a single-card loop.
• Coolant: Distilled water + biocide, or premixed clear coolant. Avoid colored coolant in research builds -- it can leave deposits that reduce thermal performance over years of operation.

Expected results for a single H100 PCIe with 360mm radiator:

Metric	Stock Blower Cooler	Custom Water Loop (360mm)
GPU Temperature (sustained training)	78-83C	50-55C
HBM3 Temperature	85-92C	55-62C
Noise at 1m	62 dBA	28 dBA
Sustained Training Stability	Stable (but loud)	Stable (and quiet)
Suitable for Shared Office	No	Yes

Based on community testing and manufacturer specifications -- actual results vary by loop configuration.

2-4 Card Sovereign Cluster Cooling

For researchers building a small training cluster with 2-4 H100 cards, cooling scales linearly: each card adds 350W of heat to the loop.

2-Card Build (700W GPU heat)

• Radiator: 480mm + 240mm = 720mm total radiator area
• Pump: Single D5 at 60-80% PWM handles the increased flow restriction
• Loop order: Pump → GPU 1 → GPU 2 → Radiator 1 → Radiator 2 → Reservoir → Pump
• Expected noise: 30-32 dBA at 1m with fans at 900-1,100 RPM

4-Card Build (1,400W GPU heat)

• Radiator: 2x 480mm = 960mm minimum, or 480mm + 360mm + 360mm = 1,200mm for quiet operation
• Pump: Dual D5 pumps recommended for reliable flow through 4 waterblocks in series
• Consider splitting into two independent loops (2 GPUs per loop) for redundancy -- if one loop develops a leak, only 2 of 4 GPUs are affected
• Expected noise: 32-35 dBA at 1m. At 1,400W, fans need to run at moderate speeds even with substantial radiator area.
• Case: At this point, you are likely using a rackmount chassis (4U), a large open-frame case, or an external radiator mounting solution.

Noise Budgets for Research Environments

Different research environments have different noise tolerances:

• Dedicated machine room / server closet: 65+ dBA acceptable. Stock blower cooling works here. Water cooling is a convenience, not a necessity.
• Shared university lab (with other equipment): 45-50 dBA background. Two water-cooled H100s at 30-32 dBA are below the ambient noise floor.
• Shared office / desk setup: 35-40 dBA maximum. A single water-cooled H100 at 28 dBA works. Two or more cards push the limit and may need an enclosure or separate room.
• Presentation / meeting room (adjacent): 30 dBA maximum through a wall. Water cooling is the only option for any H100 build in this scenario.

Honest Caveats

Before investing in H100 water cooling, consider these practical limitations:

• SXM modules are not DIY-friendly. If you have SXM H100s, the cooling solution is part of the DGX/HGX baseboard design. Custom water cooling an SXM module requires a custom adapter board, custom power delivery, and is not covered by this guide.
• H100 PCIe cards are expensive. Used H100 PCIe cards sell for $15,000-25,000 as of April 2026. Adding a $200 waterblock to a $20,000 card is proportionally insignificant, but the total build cost is substantial.
• Warranty considerations. Removing the stock cooler may void the manufacturer warranty on the GPU. For academic and research use, this is often acceptable. For production deployments, consult your procurement team.
• When to consider alternatives. If your budget allows for a DGX Spark ($4,699, 128GB unified memory) or a Mac Studio M4 Ultra ($6,999, up to 512GB), and your workload is inference-only rather than training, those turnkey solutions avoid the complexity of custom cooling entirely.

Coolant and Material Considerations for Long-Term Research Builds

Research builds often run for months or years without interruption. Coolant selection matters more here than in a gaming PC that gets shut down regularly:

• Distilled water + biocide: The simplest and most reliable option. Distilled water has the best thermal conductivity of any coolant, and a biocide additive (e.g., PTNuke or a silver kill coil) prevents biological growth. Change the water annually.
• Premixed clear coolant: Convenient but more expensive over time. Use non-colored varieties to avoid dye deposits in the waterblock's microfin channels.
• Avoid colored coolant: Colored coolants can leave deposits that reduce flow and thermal performance over months of continuous operation. In a research build that runs 24/7, clear coolant is the practical choice.
• Avoid mixing metals: If your waterblock is copper/nickel (as the Bykski H100 block is), ensure your radiator is also copper or brass. Do not mix aluminum radiators with copper waterblocks -- galvanic corrosion will degrade both components within months.

The Sovereign AI Training Case

In April 2026, Vitalik Buterin published his guide to building a self-sovereign local LLM setup, describing the hardware and software needed to run AI models without depending on cloud providers. His post resonated because it articulated what many AI researchers already feel: there are compelling reasons to train and run models on hardware you physically control.

For university labs subject to data governance policies, for researchers working with sensitive medical or legal data, and for organizations in regions where AI regulations restrict cross-border data transfer, local training is not a preference -- it is a requirement. The H100 is the standard GPU for this work because of its HBM3 bandwidth and training-optimized architecture.

Water cooling makes sovereign AI training practical in the physical environments where researchers actually work. A lab with four blower-cooled H100 cards is a lab that nobody wants to spend time in. The same lab with four water-cooled H100 cards is a comfortable workspace where researchers can monitor training runs, adjust hyperparameters, and have conversations without shouting over fan noise.

Comparison with Pre-Built Solutions

Solution	GPU(s)	Price Range	Noise Level	DIY Required
DIY H100 PCIe + Bykski waterblock	1-4x H100 PCIe	$15,000-80,000+ (GPUs) + $200-800 (cooling)	28-35 dBA	Yes
Lambda Vector (dual 5090, AIO cooled)	2x RTX 5090	~$12,000	40-45 dBA	No
BizonTech Z5000 (liquid cooled)	1-4x various	$7,000-150,000	35-45 dBA	No
NVIDIA DGX Spark	GB10 Grace Blackwell	$4,699	~30 dBA	No

Prices as of April 2026. DGX Spark uses integrated cooling. Lambda and BizonTech noise levels depend on configuration.

The DIY path with custom water cooling is not the cheapest option -- the GPU cost dominates. But it provides the quietest operation, full control over the cooling solution, and the ability to expand or modify the system over time. Pre-built solutions like Lambda Vector or BizonTech trade DIY effort for convenience, but they lock you into their cooling design and often produce more noise than a well-designed custom loop.

Ready to Build?

The Bykski H100 80G waterblock and the Bykski H200 141G AIO are both available in our catalog. For custom loop components -- radiators, D5 pumps, fittings, and coolant -- browse the Sovereign AI Rig collection. Whether you are building a single-card research workstation or a 4-card sovereign training cluster, quiet liquid cooling makes the H100 practical in environments where the stock blower is not.