【The Biggest Compute Black Hole in 10,000-GPU Clusters: Why Does 80% of the Electricity Bill for the World’s Large Models Go to Computing 0 × 0?】
Shenzhen AI chip upstart Muxin Artificial Intelligence has just secured nearly 1 billion yuan in Series C+ funding, bringing its total raised across two rounds in just six months to nearly 2 billion yuan! The company has exposed one of the most hidden vulnerabilities in the world’s computing infrastructure: the Zero-Value Compute Tax.
The human brain has 86 billion neurons, but only 1% to 2% fire and activate during any given thought, consuming just 20 watts of power. Traditional GPUs, by contrast, are built on dense-computing architectures designed for 3D game rendering. Even when more than 80% of the weights and activations in a trillion-parameter model are zero during inference, traditional GPUs still have to move every zero multiplied by X into registers and compute it—wasting enormous amounts of memory bandwidth and electricity!
Muxin’s self-developed Antoum architecture pioneered a dual-sparsity algorithm (static weight sparsity + dynamic activation sparsity), skipping useless zero-value operations directly at the silicon level. It supports sparsity rates of up to 32× and delivers computing power equivalent to 3774 TOPS at just 300 watts.
This is also a major turning point for Web3 and decentralized DePIN computing networks: as AI Agents enter an era of high-frequency, 24/7 autonomous on-chain interactions, the winner will no longer be whoever piles up the most expensive GPUs, but whoever can use sparse computing to cut the cost of each Token inference by an order of magnitude! Do you think architectural innovation can break the computing power monopoly?
#BTC #Crypto #AI #Web3
Shenzhen AI chip upstart Muxin Artificial Intelligence has just secured nearly 1 billion yuan in Series C+ funding, bringing its total raised across two rounds in just six months to nearly 2 billion yuan! The company has exposed one of the most hidden vulnerabilities in the world’s computing infrastructure: the Zero-Value Compute Tax.
The human brain has 86 billion neurons, but only 1% to 2% fire and activate during any given thought, consuming just 20 watts of power. Traditional GPUs, by contrast, are built on dense-computing architectures designed for 3D game rendering. Even when more than 80% of the weights and activations in a trillion-parameter model are zero during inference, traditional GPUs still have to move every zero multiplied by X into registers and compute it—wasting enormous amounts of memory bandwidth and electricity!
Muxin’s self-developed Antoum architecture pioneered a dual-sparsity algorithm (static weight sparsity + dynamic activation sparsity), skipping useless zero-value operations directly at the silicon level. It supports sparsity rates of up to 32× and delivers computing power equivalent to 3774 TOPS at just 300 watts.
This is also a major turning point for Web3 and decentralized DePIN computing networks: as AI Agents enter an era of high-frequency, 24/7 autonomous on-chain interactions, the winner will no longer be whoever piles up the most expensive GPUs, but whoever can use sparse computing to cut the cost of each Token inference by an order of magnitude! Do you think architectural innovation can break the computing power monopoly?
#BTC #Crypto #AI #Web3
