The drive to lock model weights and user prompts within the "black box" of hardware encryption is a noble goal, but it comes with a staggering price tag. A new benchmark from Mozilla engineers Wei Wang, Abdul Hayee Waqas, and Burns Smith reveals that data security demands payment not just in capital, but in a significantly degraded user experience. The researchers tested Mistral-7B and Qwen3-30B-A3B on a single NVIDIA H100 80GB within an Intel TDX confidential instance, and the findings are a sobering reality check for any CTO.
The results show a direct correlation between privacy levels and latency. Under a fixed workload, the average Time to First Token (TTFT) jumped by 21.8% for the lightweight Mistral-7B and a massive 27.8% for Qwen3. Total throughput also took a hit, dropping by 17.7% and 21.1% respectively. In other words, by enabling Confidential Computing, you are effectively throttling your GPU’s performance by a fifth just to ensure data remains within a trusted execution environment.
Key performance metrics
Average response time (TTFT) increases by 22–28%. Total throughput drops by an average of 20%. Performance saturation occurs significantly earlier in protected mode.
Closed-loop experiments with parallel requests confirm that the performance gap fluctuates between 11.5% and 20.2%. Perhaps most concerning is the saturation effect. Models in protected mode hit their ceiling much faster: adding more requests fails to scale throughput and instead turns the processing queue into an endless wait.
Encryption overhead is not a flat one-time fee; it is a progressive tax that scales with model size and traffic volume.
These benchmarks make one thing clear: companies planning a move to a secure cloud must budget for significant computational overprovisioning. Until the industry optimizes data transfer between Intel TDX and GPUs, high-level confidentiality will remain a luxury for fintech and the public sector—industries willing to trade speed for the peace of mind of their legal departments.