Fun Facts of Heterogeneous Computing

Deeply engaged in our lab's explorative environment, we have thoroughly researched on the use of GPUs and computational chips for diversely running smart contracts. Our setup, which included a pair of 3090, consumer-level GPU , remarkably achieved the feat of 200 million smart contract calls within a mere second. Amid the DeFi summer buzz, we conducted a series of intriguing experiments. One notable venture was harnessing CUDA technology to process 200 million paths every second, enabling lucrative arbitrage across various swaps. Our approach focused primarily on the liquidity pools present in these swaps, bypassing subjective trading strategies. We identified a myriad of potential paths across thousands of these pools, with simple ones like USDT->BTC->USDT. Our objective was to rapidly enumerate about 200 million paths, a task achieved using a comprehensive algorithm. Remarkably, our CUDA-powered GPU could navigate complex paths up to the 6th loop, like USDT->A->B->C->D->E->USDT, resulting in significant arbitrage gains.

As of today, propelled by the AI-induced computational race, we've been able to effortlessly establish several terabytes of high-speed storage and exponentially boost our computational capacity. By drastically enhancing Trie and integrating hundreds of thousands of CUDA parallel computational units, we are poised to massively expand the capabilities of the Computing Platform.

We transitioned Trie to high-speed VRAM from its conventional storage form, boosting its concurrent processing ability through innovative sharding algorithms. Furthermore, we achieved seamless state synchronization across various regions via temporal synchronization.

Although transforming computing units from sequential to asynchronous parallel processing and shifting to synchronized algorithms for state access might still hinge on the efficiency of state storage, we anticipate that separating algorithms and storage for well-designed contract codes will lead to substantial scalability.

Looking ahead, 0xVM plans to migrate the VM onto a heterogeneous computing platform, delving into the myriad applications made possible by AI's immense computational strength. This move signifies a critical advancement in leveraging AI and heterogeneous computing's power for blockchain and smart contract technologies.

Challenge

The experiments and reasoning we've conducted are logically sound, but they come with their own set of significant challenges. A notable issue is the unpredictability in the outcomes produced by complicated smart contracts. This begs the questions: How can we verify these results and maintain consistency? This becomes particularly pressing after integrating CUDA technology, which has led to an incredible increase in performance by a factor of a thousand. A crucial dilemma we face is how to effectively merge heterogeneous parallel processing with serial state synchronization, without compromising on efficiency.

Our lab has previously conducted tests on the TON blockchain, which operates on an asynchronous basis and does not guarantee strict consistency. If the TON model is deemed feasible, it suggests that our theoretical approach of running smart contracts using CUDA is also achievable. We are currently exploring the hypothesis that a single GPU equipped with 20,000 CUDA cores might offer a performance enhancement of up to 5,000 times. The prospect of such an advancement is indeed very exciting and eagerly awaited.