"The mechanism for verifying the location of the accelerators isn't particularly sophisticated. From what we understand, it's essentially the equivalent of running a ping and comparing the response time to a lookup table to see whether it's within the expected range. This kind of infrastructure monitoring and management service isn't uncommon in modern datacenters. For example, this class of software might use telemetry to identify a failing component and flag the system or contain it for replacement."
"Nvidia already provides some of this functionality through its Datacenter GPU management suite. It's also important to note that the offering wouldn't rely on baking new functionality into its GPUs. Instead, it would rely on existing security functionality present in GPUs going back to the Hopper generation. Nvidia has previously denied allegations of backdoors, kill switches, or spyware in its chips and services raised by Chinese authorities."
Nvidia is developing a managed inventory service to stream telemetry on GPU fleet health to minimize failures and maximize uptime. The platform will provide optional, opt-in location verification that uses simple latency checks, essentially pinging accelerators and comparing response times to an expected lookup table. The service builds on existing security features present in GPUs since the Hopper generation and does not require new hardware functionality. Development was driven by customer demand rather than proposed legislation. Nvidia has denied allegations of backdoors, kill switches, or spyware in its chips and services. Lawmakers have proposed requiring hardware location verification.
Read at Theregister
Unable to calculate read time
Collection
[
|
...
]