Direct Liquid Cooling Design for AI and GPU Clusters
تصميم التبريد السائل المباشر لعناقيد الذكاء الاصطناعي ومعالجات الرسوميات
- Next date
- 18–22 January 2027Riyadh
- Fees from
- £2,950per delegate
- Formats
- Classroom · Online4 sessions
- Duration
- 5 daysCertificate
Overview
AI training and inference clusters have pushed rack densities far beyond what air cooling can handle. Modern GPU servers can draw more than 100 kW per rack, and chip-level heat flux keeps rising with every hardware generation. At these loads, cold plates fed by a dedicated coolant loop are becoming the default rather than the exception. For engineers, this shift changes almost every design decision, from the choice of facility water temperature to the layout of piping, manifolds and coolant distribution units inside the white space.
This course takes a design-first view of direct-to-chip liquid cooling. Participants learn how heat travels from the GPU die through the cold plate into the technology cooling loop, and how that loop connects to the facility water system through coolant distribution units. The course covers thermal and hydraulic sizing, coolant selection, water temperature classes, rack manifolds and quick disconnects, redundancy concepts, filtration and water quality, leak management and controls. Participants finish by developing a concept design for a GPU hall and reviewing the key decisions that determine capacity, efficiency and resilience.
Objectives
- Explain why AI and GPU workloads require direct liquid cooling at high rack densities
- Calculate heat capture, coolant flow rates and temperature rise for cold plate loops
- Select coolant distribution unit types, sizes and redundancy arrangements for a GPU hall
- Design technology cooling loops with manifolds, quick disconnects and balanced hydraulics
- Specify coolant chemistry, filtration and water quality limits for long-term reliability
- Integrate the liquid loop with facility water, air cooling and heat rejection systems
- Develop a concept design and design basis for a direct liquid cooled GPU cluster
Who should attend
Mechanical and HVAC engineers, data center designers, consultants and engineers delivering AI and high-density projects.
Training methodology
The course is built around engineering calculations and design exercises rather than product presentations. Each day combines short technical sessions with worked examples using realistic GPU rack data, flow and pressure drop calculations, and review of sample piping and CDU layouts. Participants work in small teams on a single design case that develops across the five days, from load definition to a complete concept design. Typical failure scenarios such as pump loss, leaks and fouling are analyzed in group discussions. Each day closes with a short design review and quiz.
Programme outline
Day1Thermal Basics of High-Density AI Computing
- How GPU and accelerator power has changed rack densities in modern AI data centers
- Heat flux at the chip level and why air cooling reaches its practical limits
- Thermal resistance from die to coolant and the role of the cold plate
- Liquid-to-chip versus air-assisted designs and typical heat capture ratios per rack
- Defining the IT load profile, diversity and peak conditions for a GPU cluster
- Key terms: TCS, FWS, CDU, approach temperature and supply temperature classes
- Exercise: building a rack-level heat balance for a typical GPU server configuration
Day2Coolant Loops and Distribution Units
- Architecture of the technology cooling system and its separation from facility water
- In-row, in-rack and liquid-to-air coolant distribution units and where each fits best
- Sizing a CDU for heat load, flow rate, approach temperature and pump head
- Facility water temperature classes and their impact on chiller-free operation
- Coolant options including treated water and propylene glycol mixtures with their trade-offs
- Redundancy concepts for pumps, heat exchangers and CDUs in mission-critical halls
- Exercise: selecting and sizing CDUs for a 2 MW GPU deployment
- Group review of CDU placement options and their effect on white space layout
Day3Hydraulic Design of the Technology Cooling Loop
- Calculating flow rates per rack from heat load and allowable temperature rise
- Pressure drop through cold plates, server hoses, manifolds and quick disconnects
- Rack manifold design, flow balancing and avoiding starved servers at the loop end
- Main piping routes, materials selection and joining methods for white space installations
- Expansion tanks, air removal, venting and pressure control in closed coolant loops
- Designing for future rack density increases without rebuilding the piping system
- Exercise: building a pressure drop budget and selecting pump head for a loop
Day4Water Quality, Leak Management and Controls
- Corrosion, erosion and biological growth risks in cold plate microchannels
- Filtration levels, strainers and side-stream treatment for technology cooling loops
- Coolant monitoring for conductivity, pH, inhibitor levels and particle content
- Leak detection methods at rack, row and room level with practical response logic
- Drip trays, containment, isolation valves and safe service procedures for servers
- Control strategies for flow, differential pressure and supply temperature setpoints
- Alarm design and integration of CDU data with BMS and monitoring platforms
Day5Integration, Commissioning and Concept Design
- Linking the liquid loop with chilled water, dry coolers and heat reuse opportunities
- Handling the residual air-cooled load from memory, power supplies and networking
- Flushing, cleaning, pressure testing and filling procedures before first IT load
- Commissioning tests for CDUs, failover scenarios and full-load thermal performance
- Common design mistakes seen in early liquid cooled deployments and how to avoid them
- Team project: presenting a complete concept design for a liquid cooled GPU hall
- Design review, final assessment and discussion of participants' upcoming projects
Dates and fees
Fees are in GBP. per delegate.
Certificate
Every delegate who attends the full programme receives a Harfard Technology certificate of completion stating the course title, duration, dates and venue.
Frequently asked questions
Will I receive a certificate?
What is the difference between classroom and online?
Can you run this course for our company?
How do I register?
Related courses
Data Centers & AI InfrastructureData Center Thermal Engineering: From Chip to Chiller
25–29 January 2027 · Abu Dhabi +3 datesData Centers & AI InfrastructureImmersion Cooling Systems: Engineering and Operation
11–15 January 2027 · Doha +3 datesData Centers & AI Infrastructure