HOT CHIPS 2025: Meta Catalina (NVL72), When AI Systems Start Being Designed for Existing Data Centers
Why Meta built its own NVL72 instead of going by the book
In this case study, Meta was not simply showing a custom version of NVIDIA's NVL72; it was making one point clearly:
AI system design must bow to real-world data center conditions.
Meta has deployed AI at data center scale for more than a decade
Most existing data centers were not built for high-density liquid cooling
AI is scaling faster than new data centers can be built
A system that can't be deployed quickly is meaningless
Catalina's starting point is not maximum performance but "maximum deployability."
What is Catalina? Logically an NVL72, physically not
Catalina is Meta's reinterpretation of NVLink 72.
Logical level: a single 72-GPU scale-up domain
Interconnect level: a full NVLink switch fabric
Software level: mapped to Meta's existing AI stack
But in the physical design, Meta made a key split.
One NVL72 is split into two IT racks
Connected by cross-rack NVLink cables
Still presented externally as a single scale-up domain
This is not a compromise but a deliberate engineering choice.
Why Meta doesn't accept a "single-rack NVL72"
For Meta, a single ultra-high-density rack is not a cure-all.
Most existing halls are primarily air-cooled
The power architecture centers on rack-level BBUs
No traditional UPS
Power, cooling and operations processes are highly standardized
If NVL72 could only live in "brand-new liquid-cooled data centers," it couldn't scale quickly.
CPU and memory are starting to constrain AI systems
Catalina's most important design difference comes from the CPU and memory.
NVIDIA NVL72: 1 CPU per 2 GPUs
Meta Catalina: 1 CPU per 1 GPU
The impact is very direct.
Double the number of CPUs
LPDDR memory rises from about 17 TB to 34 TB
GPU + CPU coherent memory approaches 48 TB
This isn't for theoretical performance but for real AI workloads.
Two racks buy greater system flexibility
Splitting NVL72 into two racks lets Meta:
Increase the CPU and memory ratio
Spread out power and heat density
Fit the power envelope of existing halls
Lower single-rack deployment risk
From Meta's perspective,
system stability and deployment speed matter more than peak single-rack density.
The real role of air-assisted liquid cooling
Catalina is neither purely air-cooled nor traditional direct liquid cooling.
GPU, CPU, NVLink switch: liquid-cooled
Heat exchange: rejected to air through the ALC
Essentially a large radiator system
This lets Catalina move into existing air-cooled data centers.
No full facility retrofit needed
AI capacity can scale quickly
Well suited as a transitional-generation solution
The real challenge of liquid cooling isn't cooling, it's "risk control"
Meta devoted substantial system resources in Catalina to handling liquid-cooling risk.
Tray-level leak detection
Rack-level leak detection
ALC-level monitoring
Coordinated control with the facility valve train
These are not single-point protections but a layered strategy.
Small leak: stop only the sled
Medium leak: stop the tray
Large leak: stop the rack and shut off the pumps
The point of the system design is "keeping an incident from spreading."
The Rack Management Controller is the system's hub, not an accessory
The RMC plays a key role in Catalina.
Monitors all trays, switches and manifolds
Coordinates the liquid cooling system with the facility
Orchestrates power and cooling shutdown
This represents a shift:
The AI rack is no longer passive equipment
but a system unit capable of protecting itself
Catalina's network design has three layers, not one
Catalina clearly separates three networks.
Front-end network: CPU-facing external services
Back-end network: GPU scale-out
Management / console network: operations and debugging
The heavy front-side cabling is not chaos but the result of design.
Why Meta is willing to sacrifice cost and simplicity for AI
In the Q&A, someone asked: are these designs really for AI?
The answer is written in the trade-offs.
Willing to add more logging and telemetry
Willing to add control modules such as the CLD and RMC
Willing to accept higher hardware complexity
Because what AI production systems fear most is not low performance, but unpredictability.
The Simple Tech Trend View
Catalina shows a route that differs from NVIDIA's but is just as reasonable.
NVIDIA is solving "extreme single-rack engineering"
Meta is solving "deployable systems at scale"
When AI reaches the 100,000-GPU class,
the systems that come online fastest and most reliably are the real winners.
Catalina wasn't born for benchmarks; it was born for real-world data centers.
Comments