Nvidia has unveiled two new graphics cards based on its Pascal architecture, the Tesla P40 and Tesla P4.
The cards are aimed at the data centre market for deep learning applications.
The Tesla P4 is designed to deliver efficiency in scale-out servers, meeting the density and power-efficiency requirements of modern data centres.
Tesla P4’s Pascal GP104 GPU offers high floating point throughput and efficiency, and features optimised instructions for deep learning inference computations.
The new IDP2A and IDP4A instructions provide 8-bit integer (INT8) 2-element and 4-element vector dot product computations with 32-bit integer accumulation.
The P4 is capable of a peak 21.8 INT8 TOP/s (tera-operations per second).
Nvidia’s Tesla P40 accelerator was made to offer the highest possible throughput for scale-up servers, with performance taking preference over space or power consumption.
The P40 has a peak 32-bit floating point throughput of 12 Teraflops and accelerates INT8 vector dot products (IDP2A/IDP4A instructions), with a peak throughput of 47.0 INT8 TOP/s.
The Tesla P40 will launch in October, and the Tesla P4 in November.