Dear CheckMates,
Not long ago, if someone had asked me to explain the difference between a GPU, DPU, DGX, AI Factory, AI Agent, LLM, RAG and AI Firewall, I probably would have given them a confused look.
I've spent more than 25 years working with security technologies, much of that with data center security, but suddenly every AI presentation seemed to assume I already understood an entirely new vocabulary.
Even GPU required a change in thinking. Of course I knew what a GPU was, the graphics processor you wanted for gaming, 3D graphics or video editing. I wasn't thinking of it as the foundation for a new generation of data centers.
So I spent some time reading and trying to understand how these technologies actually fit together. Once I stripped away the acronyms and marketing terminology, I realized that the basic concepts aren't nearly as complicated as they first appear.
I'm not an AI researcher. I'm a security architect who wanted to understand enough about AI infrastructure to understand how we secure it. I'd like to share some of what I've learned from one security engineer to another.
Even if you don't work with AI infrastructure today, I think it's worth learning the basics. Something that looks specialized today may be sitting in one of your customers' data centers a few years from now.
So let's start with the infrastructure.
What Is an AI Factory?
An AI Factory is essentially a data center — or part of one — designed specifically for AI workloads.
Like any data center, it has servers, storage and networks. The difference is that the infrastructure is optimized for the enormous amount of computing and data movement required by AI.
NVIDIA's purpose-built AI computing systems are called DGX. So if the AI Factory is the data center, DGX systems are the servers inside it.
What's Inside a DGX?
A DGX contains several types of processors with different jobs:
CPU (Central Processing Unit) — The familiar general-purpose processor that runs the operating system and applications.
GPU (Graphics Processing Unit) — GPUs are exceptionally good at performing huge numbers of calculations in parallel, which makes them ideal for AI. In simple terms, GPUs do the heavy AI computation.
DPU (Data Processing Unit) — A processor designed to handle infrastructure functions such as networking, storage and security.
NVIDIA's DPU platform is called BlueField. Think of it as much more than a network card: it has its own processors and memory and can perform infrastructure functions independently from the host.
And this is where Check Point comes into the picture.
Check Point integrates its AI Factory Firewall (AIFF) with NVIDIA BlueField, allowing security enforcement to run directly on the DPU rather than consuming the CPU and GPU resources used by the AI workloads.
The whole picture can therefore be simplified to:
AI Factory
→ DGX server
→ CPU — general computing
→ GPU — AI computing
→ BlueField DPU — infrastructure processing
→ Check Point AIFF — security
Once I understood this hierarchy, AI Factory Firewall suddenly made much more sense.
It isn't some mysterious new type of perimeter firewall. It's Check Point security integrated directly into the infrastructure of the AI server. Remember that AIFF is not the same as the AI network firewall that we recently released.
Why Does Running Security on the DPU Matter?
The important thing about BlueField is where it sits.
In very simplified terms:
Network → BlueField DPU → DGX host and workloads
Traffic is already passing through the DPU, and BlueField has its own computing resources. This gives us an interesting place to enforce security without taking CPU or GPU resources away from the AI workloads.
AIFF can provide capabilities such as segmentation, threat prevention and workload protection close to the systems being protected.
Instead of bringing every workload to a firewall, we can bring security enforcement closer to the workload.
And because this enforcement runs separately from the host operating system, it also provides an additional level of isolation between the security infrastructure and the workloads it protects.
Why Is AI Security Different?
AI infrastructure still has the security problems we already know:
AI doesn't make any of those disappear.
It also introduces new risks involving AI applications, models and data, including things such as prompt injection, model theft, data poisoning, AI agent abuse and sensitive data leakage.
A useful way to think about it is:
AI doesn't replace traditional security problems. It adds new ones on top.
Fortunately, that also means much of what security engineers already know remains relevant.
This Is Only One Piece of AI Security
AIFF and BlueField are interesting, but securing the DGX server is not the same thing as securing the entire AI Factory.
We still need to think about familiar security architecture questions:
-
How do we protect the AI Factory perimeter?
-
How do we control east-west traffic and prevent lateral movement?
-
How do we implement segmentation and Zero Trust?
-
How do we protect management infrastructure?
-
How do users and applications securely access AI services?
-
How do we protect models, data and AI applications themselves?
This is where technologies such as traditional Check Point gateways, Maestro, segmentation, identity and other security controls continue to play important roles.
But trying to cover all of that in one post would bring us straight back to the problem I started with: too many new concepts at once.
Each of those areas is probably worth a post of its own.
For now, if you remember just this:
AI Factory = AI-focused data center
DGX = NVIDIA AI server
GPU = AI computation
DPU / BlueField = infrastructure processing
AIFF = Check Point security integrated with BlueField
...then the next AI Factory architecture diagram you see should hopefully make considerably more sense.
The infrastructure is changing, but many of the security principles we've been applying for years are not.
We just need to learn the new vocabulary.
