Before You Trust a Liquid Cooling Claim, Ask These 5 Questions
By Maurizio Miozza, Co-Founder, GemaTEG
The liquid cooling market is moving faster than most buyers can process, and every week brings a new announcement promising to solve the data center heat problem once and for all. The claims are bold. The fine print often tells a different story.
I have spent years in thermal management. What concerns me is not the pace of innovation. It is the growing gap between what vendors announce and what operators need to understand before making a decision that can cost millions.
It is why we recently published a new white paper. This article shares some of the thinking behind it, and I also sat down for a video conversation to walk through it in more depth. The paper itself goes much further still, with the full data, methodology, and a complete framework for evaluating any thermal management solution on the market.
Watch the full conversation here
A good place to start is a distinction that gets lost in most of the noise.
Cooling vs. thermal management
Cooling extracts heat from a chip. Done well, it:
- Removes heat efficiently under controlled conditions
- Is well understood and simple to procure
But it only answers one question: how much heat can you remove? It does not tell you whether the chip stays within its rated temperature, or what that performance actually costs in energy and infrastructure.
Thermal management controls the chiplet's temperature, zone by zone, in real time. Done well, it:
- Keeps every zone within spec
- Deploys cooling energy only where the thermal gradient demands it
- Adapts automatically as workloads shift between training and inference
- Reduces dependency on chillers and oversized infrastructure
Most of what is marketed today as thermal management is, technically speaking, cooling. That gap matters more than ever, because today's chips were not designed for yesterday's solutions.
Five questions to ask before believing the headline number
The market makes this hard to evaluate on your own, so over the years I have settled on a practical set of questions. They will not tell you everything, but they will tell you whether a claim is built on solid engineering or built to sound good in a press release.
1. What flow rate is required, and what does it cost at the CDU level?
A reasonable operating range is 60 to 120 liters per hour at the cold plate level, up to 200 liters per hour for the new generation of high-power GPUs. When requirements run ten times that figure, ask what it means for pumping power, CDU sizing, infrastructure investment, and long-term reliability.
2. What fluid temperature does this assume, and can my facility supply it without a chiller?
Solutions requiring low fluid temperatures (below 35°C) impose a refrigeration dependency. That energy cost and capital expenditure belong in the evaluation.
Important Note: Some vendors now claim chiller-free operation, but they also concede that it is geography-dependent: in warmer climates or during peak summer months, mechanical refrigeration remains necessary. Geography-dependent chiller-free operation is not the same as chiller-free operation.
3. Does this solution differentiate between the GPU die and the HBM memory stack?
A system applying uniform cooling across the entire package cannot respond to the distinct thermal signatures of co-located compute and memory. For current and next-generation chiplet architectures, this is a fundamental limitation.
4. What happens to my infrastructure as chip TDP increases over the next three to five years?
A solution optimized for today's chips may require full replacement when the next GPU generation arrives. The transition cost is part of the true cost of the solution.
5. What are the long-term reliability implications at the stated operating conditions?
High flow rates, low fluid temperatures, and continuous high-load operation all carry maintenance and reliability consequences that rarely appear in product announcements.
These are not abstract concerns. They are the questions that determine whether an infrastructure investment will hold up over the next several chip generations, and they are the same questions we used to stress-test our thinking while developing DaTEG® 2.0.
Why we wrote this white paper
We spent three years developing our answer to these problems, and two more years testing it before saying a word publicly. In this industry, claims must be backed by data, and data must be reproducible.
If you are specifying infrastructure, evaluating vendors, or trying to understand the impact of decisions made in your organization, this paper was written for you. It goes well beyond these five questions, with the data and modeling behind each one.
