NVIDIA announced today the next generation GPU architecture, “Pascal”, with unified virtual addressing, 3D Memory dimensional stacked integrated memory, NVLink high-speed interconnect channel etc. Seems killer. In fact, NVIDIA has had initial sample Pascal graphics core, but did not come up with a kind, but released a concept map.

This card design is very unique, but special attention is needed, at least understand the current situation, it is not the ordinary GeForce desktop, but high-performance computing Tesla.

NVIDIA's Next Generation Pascal architecture: The card would be so long

From the photo, this new design is not only compact (official said the current high-end cards only about 1/3), and is no longer vertically into the motherboard’s PCI-E expansion card into the motherboard lying on daughter card style, and connected by a newly designed NVLink bus and systems.

This alone can explain, this design does not work with common desktop, NVIDIA motherboard industry for no right to speak, but in the server space, especially those areas that need a compact computing, NVIDIA and vendors can collaborate, a new design.

Such a calculation using the card, not only can be used on higher-speed NVLink bus channel, but also greatly reduces the space occupied by the server, and can share the processor and CPU heatsink.

You can see that it is not the system interface PCI-E, so what is this ??

Nvidia NVLink

NVLink is one of the most critical areas, which is NVIDIA, IBM to design a new bus channel, plans to replace at least some extent PCI-E, better support for multi-GPU interconnection , because the current PCI-E 3.0 16GB / s bandwidth is too narrow, with memory frequently 250GB / s + fully immeasurably future PCI-E 4.0 is not likely to catch up.

NVLink actually learned a lot of features PCI-E, the programming is the same, but ituses a different signal systems, and belongs to the real point designed to allow multi-GPU direct interconnection, do not bypass the PCI-E. It’s a bit like AMD HT, Intel QPI.

The minimum unit NVLink called ” Block “(blocks), each containing eight channels, each channel 20Gbps, is the total bandwidth of 20GB / s, on the transfer rate is concerned about 20GT / s, far beyond the PCI-E 3.0 8GT / s.

NVIDIA did not let us see them specifically what new interfaces (certainly in the back of the card), but obviously requires a lot of changes, must motherboard, server vendors do it together.

NVIDIA knows it can not completely abandon the PCI-E, some command and control, communications NVLink actually to go PCI-E, said earlier that at least the beginning of a new design can be seen in high-performance computing server, the consumer level workstation graphics market even have honestly with PCI-E.