- Openai’s hardware manager warns future AI models need real-time hardware killing contacts
- Richard Ho highlights networks, memory and power challenges in scaling infrastructure
- He calls for benchmarks, observability and cross -industry partnerships to tackle reliability and trust
A senior Openai director has warned that future AI infrastructure will require hardware level security features, including Kill Switches.
Richard Ho, head of hardware in the company, made the comments under his keynote speaker at AI Infra Summit in Santa Clara.
“It has to be built into hardware,” Ho said. “Today there is a lot of security work in the software. It assumes your hardware is safe. It assumes your hardware will do the right thing. It assumes you can pull the plug on hardware. I don’t say we can’t pull the plug on that hardware, but I tell you these things are dishonest, the models are really dishonest, and then like a hardware -I want to make sure.”
Silicon security measures
Ho argued that the growth of generative AI forces a reconsideration of system architecture and described how future agents will be long lasting and interact in the background, even when a user is not actively engaged.
This shift requires memory -rich infrastructure with low latency to control continuous sessions and communication across multiple agents.
Networking, Ho said, becomes a bottleneck. “We need to have real-time tools in these-what means these agents are communicating with each other. Some of them may look at a tool, some may have a site search. Others think and others have to talk to each other.”
Ho outlined several hardware challenges to be addressed, including limits on high bandwidth memory, the need for 2.5D and 3D chip integration, progress in optics and extreme power requirements that could reach 1 megawatts per year. Rack.
The safety measures Openai Taken included include real-time killing contacts built into AI clusters, telemetry to detect signs of abnormal behavior and secure execution paths in CPUs and accelerators.
Ho wrapped things by saying, “We don’t have good benchmarks for agent-noticing architectures and hardware, and I think it’s important to know about latency walls and latency tails, what is the efficiency and power and things like that. We need to have good observability as a hardware feature, not just as a debug tool, but built-in and constantly monitors our hardware.”
“Networking is a real important thing, and when we go towards optical, it is unclear that the reliability of the network is there today. We have to get there with enough testing of these optical test beds and these other communication tests that show that we actually have reliability.”
Via The next platform
