Verifiable AI: Cysic’s GPU-Accelerated Leap in Trusted Machine Learning

Jul 21, 2025

Artificial intelligence is powering critical decisions in healthcare, finance, and more — but how do we trust that these AI models run correctly? Verifiable AI offers a solution, using cryptographic proofs (specifically zero-knowledge proofs) to mathematically guarantee an AI’s outputs are correct without revealing sensitive data. Historically, generating these proofs for large neural networks was painfully slow, but Cysic’s cutting-edge platform changes the game. By harnessing massive GPU parallelism and custom algorithms, currently Cysic delivers real-time verifiable AI inference for a bunch of useful AI models. In fact, verifying a large model’s computations can be extremely expensive without hardware acceleration — a challenge Cysic meets head-on with GPU-centric optimizations. Below, we break down how our technology makes trustworthy AI both fast and easy to use.

Our Approach

Our initial results show impressive results for several useful models. For instance, CNN seems to be the model used widely in the community as the comparison baseline. We benchmarked our proving system for CNN on different parameter settings, as shown in the following table. It achieves about 10x speedup compared with the state-of-the-art on CNN-4M. The server we use contains AMD EPYC 7773X 64-Core Processor, one NVIDIA GeForce RTX 4090 and 512GB RAM. Below, we break down how our technology makes trustworthy AI both fast and easy to use.

Comparison Table

GPU-Parallel Sumcheck: Thousands of Threads, One Unified Proof

One cornerstone of our system is the sumcheck protocol — a cryptographic method used in zero-knowledge proofs to verify huge polynomial calculations efficiently. Cysic’s prover divides this traditionally heavy process into tiny tasks spread across thousands of CUDA threads running in parallel. By distributing the sumcheck computations (and the associated lookup table checks) over the GPU’s many cores, we keep every shader busy and eliminate bottlenecks. Research shows that launching on the order of 10,000 threads can fully saturate a modern GPU (e.g. 10,240 threads on an NVIDIA V100 with 5,120 cores) , ensuring maximum throughput. In our architecture, all those threads work in concert on different parts of the proof — continuously and without idling — so the sumcheck and lookup steps finish blazingly fast. The result is a prover that scales with GPU horsepower: as GPUs get more cores, our sumcheck protocol automatically gains speed by using them all. This massive parallelism translates to lower latency and higher proof-per-second performance, bringing real-time verifiable AI closer to reality.

Optimized Field Arithmetic: Maximizing GPU Efficiency

Another technical breakthrough is our optimized prime-field arithmetic on GPUs. Neural network proofs involve intensive math operations on large integers in a prime field (the math underpinning our cryptographic proofs). Instead of relying on generic GPU routines, Cysic developed custom CUDA kernels tailored for these finite-field calculations. The focus is on minimizing memory usage and maximizing arithmetic throughput on the GPU. Concretely, our kernels keep data close to the GPU’s ALUs (Arithmetic Logic Units) by leveraging fast on-chip memory (like registers and shared memory) and avoiding slow global memory wherever possible . For example, we pack frequently used values into GPU registers (which are the fastest storage) so that computations aren’t bottlenecked by memory fetches . We also coalesce memory accesses and use warp-level parallelism smartly, so dozens of threads fetch and process data in lockstep for efficiency . Altogether, these low-level optimizations mean the GPU spends more time crunching numbers and almost no time waiting on memory. Our prime field math kernels squeeze every drop of performance from the hardware, thus performing more operations per second and speeding up proof generation. The bottom line: Cysic’s verifiable AI runs fast because it treats GPU architecture not as an afterthought but as an integral part of the design, achieving superior utilization of GPU compute resources.

Seamless Framework Integration: Proofs at the Push of a Button

Speed isn’t the only highlight — we’ve made verifiable AI incredibly easy to use. Cysic’s platform offers out-of-the-box support for popular ML frameworks like PyTorch and TensorFlow. This means you can take any existing nn.Module (or TensorFlow model), wrap it in our provided VerifiableModule, and instantly get cryptographic proofs alongside every forward pass. There’s no need to rewrite your model or understand complex cryptography. Under the hood, our system automatically transforms your model’s computations into a verifiable format and invokes the GPU-accelerated prover to generate a proof in tandem with the model’s output. This seamless integration allows AI developers to continue working in familiar frameworks while gaining the added trust of verifiable results

Next Steps: What’s Coming in Verifiable AI

• CNN & Large-Scale LLM Support: We’re extending our framework to support more complex architectures, including convolutional neural networks and large open-source language models. Soon, models like Llama and the DeepSeek series will run with verifiable inference, just as easily as our current demos. This expansion will bring the power of verifiable AI to cutting-edge vision and language tasks, ensuring even massive CNNs and transformers can be proven correct in their predictions. As the scale of ML models grows, our commitment to acceleration grows.
• Application Demos: To showcase the technology, we’re rolling out end-to-end verifiable AI demos for real-world applications. Expect to see examples like face recognition with FaceNet, object detection in images, and more — each one producing cryptographic proofs live alongside the AI results. These demos will illustrate how verifiable AI can bolster trust in everyday AI services.
• Open-Source Launch: We believe in the power of community. That’s why we’re preparing a comprehensive open-source release of our verifiable AI platform in the next months. By open-sourcing our technology, we invite researchers, engineers, and enthusiasts to collaborate, audit, and build upon our work. This will include thorough documentation, example projects, and integration guides to make adoption frictionless. Our open-source launch isn’t just a code drop, while it’s the start of an ecosystem. We envision a growing community pushing the boundaries of what verifiable AI can do, from optimizing proof algorithms to exploring new applications. Transparency and community engagement will also reinforce trust in the system itself. We’re incredibly excited to see the creative ways the world will use and improve this technology once it’s in everyone’s hands.

In summary, Cysic’s verifiable AI platform merges the rigor of cryptography with the power of modern GPUs, delivering a solution that is fast, developer-friendly, and poised to scale. The future of AI is one where every inference can come with a mathematical guarantee. Verifiable AI is a core pillar of Cysic’s ComputeFi vision. At its heart, ComputeFi aims to decentralize and democratize access to cryptographic and AI computing power, ensuring not just availability, but also trustworthiness of compute. This aligns seamlessly with Cysic’s ComputeFi infrastructure, where users can tap into a decentralized pool of compute providers to run AI workloads with strong cryptographic guarantees. It transforms AI from a black box into a verifiable service layer, enabling a new class of trust-minimized applications in areas like decentralized identity, on-chain agents, and secure multi-party AI collaboration.