Nvidia’s masterpiece

Nvidia has produced a 5,000 core GPU delivering 7.5 FP64 Teraflops aimed at running deep learning algorithms.

IMG_0619

Called the Volta, the chip measures 815 sq mm and contains 21 billion transistors. It was made on TSMC’s 12nm process.

Nvidia CEO Jen-Hsun Huang (pictured) called Volta a ‘masterpiece’.

“To make one chip work per 12-inch wafer, I would characterise it as unlikely,” said Huang, ” so the fact that this is manufacturable is an incredible feat.”

The Volta can handle 120 Tensor Teraflops for deep learning. It can transfer data at 300 Gb/s.

Nvidia says that, this year, it will train 100,000 developers to use deep learning.

Microsoft is putting together 64 Voltas for an image recognition Deep Residual Network application.


Comments

2 comments

  1. Maybe Huang wasn’t joking when he said he gets one working die per wafer, Pete.

  2. The MOAC – Mother of all Chips. Our first chip at Picochip was 270 sq mm and we thought that was insanely big. We needed to have rundundancy to get sane yields. With 5000 cores, I’n sure Nvidia keep a few in their back pocket aa well, otherwise the yield would be near zero.

Leave a Reply

Your email address will not be published. Required fields are marked *

*