According to a leak from Chinese site ByCare, it appears that Nvidia is working on a new single-GPU compute graphics card, the Tesla K40, codename "Atlas". To be based on a new GK108 GPU, the Tesla K40 should pack 2,880 CUDA cores and provide over 4 TFLOP/s of single-precision compute performance and pack some quite impressive features.

According to the leaked slide, the GK180 behind the Tesla K40 does not look that much different from the GK110, the GPU behind the Tesla K20X, as it packs 2880 CUDA cores. On the other hand, the peak floating point compute performance rose from 3.95 TFLOP/s for single-precision and 1.32 TFLOP/s for double-precision, up to over 4 TFLOP/s for single-precision and over 1.4 TFLOP/s for double-precision floating point performance.

Nvidia appearently also decided to double the amount of memory to 12GB of GDDR5 with a memory bandwidth of 288GB/s. The new Tesla K40 should also end up with some sort of dynamic overclocking feature that will work on ANSYS and AMBER workloads and is also made with support for PCI-Express 3.0.

The slide also details the "total board power" which should end up at around 235W (245W SXM) depending on the form factor. The slide does note that these are preliminary specifications so we will probably wait a bit longer before we see see it.



Source: Via Wccftech.com.