The death of Moore’s Law is old news at this point, no matter what Intel might say. A few years back, NVIDIA CEO, Jensen Huang was the first one to admit this. Now, the company has gone a step further ahead, introducing what it calls “Huang’s Law”, named after Jensen himself.
Moore’s Law stated that the number of transistors on a chip doubled with each generation (every 2 years or so). That hasn’t happened in the last half-decade or so, with Intel being stuck on the same 14nm node it launched in 2016 with the Skylake parts.
TSMC which is now the leading foundry with the 5nm EUV process hasn’t been able to crack that figure either. It’s latest two nodes, namely the 5nm and 7nm were roughly 25-35% denser than their preceding process nodes, nowhere close to the 2x figure dictated by Moore’s Law.
NVIDIA’s newly introduced law states that GPU performance in AI will double every year. According to Huang’s Law, it is artificial intelligence, not the process node, that will play a key role in improving performance with every next-generation.
This falls in line with NVIDIA’s recent policies in both the gaming Data Center market. DLSS 2.0 is being called the most advanced upsampling technology, thanks to the use of neural networks paired with a temporal algorithm.
In the Data Center market, the company’s A100 accelerator is second to none, offering insane mixed-precision performance gains compared to Volta thanks to the use of structured sparsity and the new TF32 data format
To keep this law consistent, NVIDIA has developed the MAGNet tool which is a modular generator-accelerator of neural networks, with a rated performance of 100 trillion ops per W or 100 TOPs. This is more than 10x higher than any accelerator available on the market.
This info was disclosed in a conference attended by the CEOs of all major IT and semiconductor companies and OEMs including Alibaba, Baidu, China Telecom, Dell Technology, Xinhua Three, Inspur, Kuaishou, Microsoft, Ping An, Xiaomi, AWS, ByteDance, Didi, Lenovo, and Tencent.