NVIDIA A100 now also in PCI Express format with 80GB of memory

NVIDIA A100 now also in PCI Express format with 80GB of memory

From technical documentation NVIDIA leaks the existence of a new version of A100 video card, top of the range proposal intended for use in datacenters and certainly not in gamer systems. This is the top card of the Ampere GPU family, based on GPU characterized by an area of ​​826 square millimeters and equipped with 6,912 CUDA cores inside.

This card has the particularity of being combined with a amount of video memory really very high: let’s talk about 80GB with HBM2e modules as maximum, a feature that allows to reach a theoretical maximum bndwidth of 2TB / s. NVIDIA offers a second version of this GPU combined with the lower quantitative 40GB HBM2 (1.6TB / s bandwidth), without prejudice to the other basic technical characteristics.

We all know that a large amount of video memory combined with a GPU allows you to increase performance, avoiding that these can encounter a bottleneck when the resolution is pushed to very high values. In the field of parallel processing, target audience of products like this, the greater amount of onboard memory is instead a feature that it allows to significantly increase the overall performance in the specific areas of parallel computing. From this an onboard quantity so high that it would be overabundant as a frame buffer even thinking of using the 8K video resolution.

You Might Want To Read This:   AMD B550 close to debut, here is the first motherboard for new Ryzen

This second version of the A100 card is offered in the traditional PCI Express form factor and joins the one offered to customers up to now, in SXM4 format: the latter is proprietary to NVIDIA and is intended to install video cards directly on the server motherboard. In this way NVIDIA can expand its product offering, giving the possibility to customers who want to install this card inside a traditional workstation without necessarily having to opt for a rack server.

The PCI Express version, obviously in standard Gen 4.0, boasts the same specifications as the SXM4 model, differing only in one detail: it is possible to connect in parallel only a maximum of 2 cards of this type via the NVLink connection, while with the SXM4 versions you can go further. It should also be noted that the SXM4 form factor allows you to provide the card with up to 400 Watts of power, a lower value (between 250 Watts and 300 Watts) for the PCI Express declination. At the moment the prices of this card are not known but thinking about the SXM4 version list (approximately 15,000-20,000 dollars) it will certainly not be a cheap card.

You Might Want To Read This:   Samsung launches app so you remember to wash your hands