Ampere unveiled: Nvidia unveils monster GPU for servers and AI
Nvidia has finally unveiled the Ampere architecture on which the company's future graphics cards will be based. Although a consumer card is still months away, the new A100 chip for servers and supercomputers offers a glimpse into the future.
Ampere is, according to Nvidia, a significant leap over the Turing architecture: the new A100 chips are based on a 7nm process and are up to twenty times faster than the Tesla V100, the manufacturer's current top product in this category.
Normal consumers will have to be patient
Nvidia officially unveiled its next-generation Ampere GPU architecture in a recorded keynote. The A100 GPU is designed for cloud computing, AI and scientific number crunching. It will initially be available for servers and supercomputers. As expected, the GTC keynote is not about graphics cards for private use. You'll have to wait a little longer for a GeForce RTX 3080 for your computer.
According to the company, the A100 is the biggest generational leap in GPUs in a long time, with 20 times the speed of the previous Volta-based solution and third-generation Tensor Cores. The A100 is a beast made of silicon with 54 billion transistors and 6912 CUDA cores. By comparison, the current flagship consumer card GeForce 2080 Ti has 4352 CUDA cores. As expected, Nvidia's new Ampere GPU is based on a 7 nm process.
Nvidia sees the biggest advantage of the new chip in the cost reduction for large data centres. According to the manufacturer, a system that currently costs 11 million dollars and requires 25 server cabinets and 630 KW of power can be replaced by an Ampere system that fits into a single cabinet, costs one million dollars and requires 28 KW of power.
Nvidia's CEO Jensen Huang takes such a system out of the oven himself at the presentation - the keynote was recorded in the kitchen of Huang's home. Nvidia's monstrous HGX motherboard is an impressive piece of engineering. Eight A100 GPUs together with 30,000 other components and a kilometre of conductors are contained in one part. This makes it one of the most complex motherboards in existence. The board is used in Nvidia's DGX A100 system. According to the manufacturer, this delivers 5 petaflops of AI computing power and 320 GB of GPU memory with 12.4 TB per second of bandwidth in a relatively small housing.
What does this mean for the next GeForce generation?
Huang then went on to say a few words about the Ampere-based GeForce generation of graphics cards. The chip is configured slightly differently. The A100 has been designed to perform floating point calculations with twice the accuracy. Consumer-orientated Ampere GPUs will be focused on graphics and less on computation.
The GTC 2020 keynote and press releases reveal little about clock speeds. But this much is clear: Ampere cards will support PCIe 4.0 and the switch to 7nm will allow Nvidia to utilise many more RT cores for improved ray tracing performance. A significant increase in performance can therefore be expected in the next generation of consumer cards. We will probably only find out just how big in a few months' time when Nvidia officially unveils the new GeForce series.
Engadget has summarised the keynote in ten minutes. Here you can find the whole keynote to watch.
From big data to big brother, Cyborgs to Sci-Fi. All aspects of technology and society fascinate me.