Security Week

Intel TDX Connect Bridges the CPU-GPU Security Gap

Kevin Townsend Published: March 4, 2025 | Updated: March 4, 2025 3 min read

The use of AI by companies is expanding rapidly. This requires the collection and processing of vast amounts of corporate data. The threat of sensitive company data and PII leaking is serious and heavily regulated by governments.

One problem is that AI data processing is performed on devices with GPUs (such as Nvidia), while the data source (as in parameters and prompts) is delivered through connected devices more commonly using standard CPUs (such as Intel). Mapping the data from one device to another has been achieved in software with the use of Bounce Buffers. But these add overhead to the data transfer and cannot be secured as effectively as hardware protection. Direct memory access, from one device to the other, is a better solution.

Intel is addressing this by extending its TDX Connect technology on its Xeon 6 processors. TDX is the basis for Intel’s Confidential Computing – isolated and hardware-protected Trust Domains within VMs providing greater data confidentiality and integrity in cloud and virtualized environments.

TDX Connect extends this concept beyond the Intel CPU to any supporting device, including GPUs, Smart NICs, and storage devices. Its relevance is primarily to Intel’s wider concept of confidential computing – but in the current technology environment, much interest will focus on the potential for confidential AI.

The data security problem for burgeoning AI applications lies in AI methodology. “AI is all about data,” explains Anand Pashupathy, VP & general manager of Intel’s security software & services division. “Parameters going in, prompts going in, data being processed, and the results coming back. A lot of this is happening without confidential computing protection.”

For him, confidential AI is the application of confidential computing to the rapidly growing use of gen-AI applications. It is a partnership between the trusted execution environment (TEE) on the CPU (that is, TDX on Intel) and the GPU’s own TEE. Data is kept confidential between the two via TDX Connect’s high performance, encrypted connection and secure direct memory access.

“This helps ensure end-to-end compliance and data security,” he writes in Intel’s announcement.

The realization of Confidential AI will not be immediate – it will require specific partnerships with the manufacturers of the GPUs processing the AI. “Intel is collaborating with partners to expand the deployment of Intel TDX Connect capabilities throughout the partner ecosystem,” writes Pashupathy. The big one, of course, is Nvidia. Although there are no details of precisely when or how, “Nvidia plans to support TEE-IO [that is, TDX Connect] on its Blackwell platform.”

Advertisement. Scroll to continue reading.

Microsoft has also declared support without timeline details. “Microsoft is excited to productize Intel TDX Connect into future generations of Azure confidential VMs, which is in early development between our hardware and software developers,” said Vikas Bhatia, head of product for Azure confidential computing at Microsoft.

“Intel TDX Connect represents a major milestone in our journey to improve confidential computing performance and extensibility. It enables existing and net-new workloads to benefit from the privacy assurances of confidential computing without compromising on price or performance.”

OpenAI, of course, currently uses both Azure and Nvidia for its AI processing – although it is worth noting the company is developing its own custom AI chips to reduce reliance on Nvidia. The design phase is expected to be finalized this year, with production by Taiwan Semiconductor Manufacturing Company (TSMC) following in 2026.

While TDX Connect is generally expanding the realm of Intel’s confidential computing, it is also opening the potential for hardware protected confidential AI. AI is a rapidly advancing technology. Organizations know they need to use AI without necessarily knowing how to secure AI data in transit between different types of processors produced by different hardware manufacturers. TDX Connect, when adopted by Intel’s hardware partners, promises to provide surety — keeping confidential data safe and regulators at bay.