“This is the fastest single computer humanity has ever created: one operating system, one programming model, you can program this as one computer.”
Speaking of the company’s new DGX-2, the first system built using the HGX-2 server platform, he added: “This is like a PC, except it’s incredibly fast.”
How fast? The 350-pound machine offers two petaflops of computing power and 512GB of HBM2 memory. (Just like a PC then).
Compared to its predecessor, DGX-2 represents a 10-fold leap in computing power in just six months, Huang said, as well as a 500-fold performance leap on the AlexNet image-recognition benchmark over what could be done five years ago on a pair of NVIDIA GPUs.
“There’s a new law in town,” Huang said. “This new law of computing says ‘If you are able, and if you are willing to optimize across the entire stack, the performance improvement you can achieve is incredibly fast.”
HGX-2 Gets a Fresh Push
The first system based on HGX-2, NVIDIA’s own DGX-2 system announced in March, features a GPU Memory of 0.5TB. It can singlehandedly replace 300 CPU servers for AI training and costs an estimated $399,999.
In released benchmarks the DGX-2 has obtained AI training speeds of 15,500 images a second using the ResNet-50 benchmark indicator
The HGX-2 comes equipped with Tesla Volta V100 GPU’s that bring together machine learning processor needs together with HPC requirements. Putting together 21.1 billion transistors on 815mm2 of silicon the Tesla Volta could be the start of reaching the limits of Moore’s law.
Huang detailed a “Cambrian explosion” of technologies driven by GPU-powered deep learning. In less than a decade, the computing power of GPUs has grown 20-fold — or growing 1.7 times each year, far outstripping Moore’s law,.
But demand for that power is “growing, not slowing,” thanks to AI, he added.
“Before this time, software was written by humans and software engineers can only write so much software, but machines don’t get tired,” he noted.
“As long as there is data, so long as there is knowledge in how to create the architecture, we can create absolute enormous software,” Huang said. “And every single company in the world that develops software will need an AI supercomputer.”
“I’ll Have One!”
Enterprise computing, storage and network solution provider Supermicro immediately announced that they would be incorporating the HGX-2 into their systems.
Charles Liang, president and CEO of Supermicro said: “To help address the rapidly expanding size of AI models that sometimes require weeks to train, Supermicro is developing cloud servers based on the HGX-2 platform that will deliver more than double the performance.’’
Also commenting on the announcement Steven Lu VP of Wiwynn said; ‘’Our collaboration with NVIDIA and the HGX-2 server building block will enable us to provide our customers with two petaflops of computing for computationally intensive AI and HPC workloads.”