The GeForce RTX 4090 is currently the best performing video card for gamers on the market. Due to its high power and generous onboard memory allocation, many users have decided to use it for non-gaming uses as well, such as content creation, machine learning, and scientific work.
However, An NVIDIA employee has confirmed that the P2P (peer-to-peer) feature, which allows data to be transmitted between the memory of two NVIDIA graphics cards while bypassing system memory, is not supported. This speeds up memory access and data transfer, which is especially useful for those working with CUDA programs.
Puget Systems has revealed that P2P-related workloads fail or are corrupted, causing potential problems for who use this feature for their work. The problem with P2P appears to be that the card does not support the NVLink bridge, which is required for P2P to work. Earlier GeForce RTX 30-series (Ampere) graphics cards also do not support P2P, as demonstrated by Puget Systems tests.
Photo Credit: NVIDIA It's easy to see why the Californian company has decided to phase out support P2P on GeForce graphics cards. NVIDIA prefers GeForce products to be used only for gaming and nothing else. However, in the past some system integrators have started using blower versions of the GeForce RTX 3090, which are less expensive than their professional counterparts, to offer affordable products for servers and resulting in a scarcity of these particular variants in the market.
In case P2P support is essential for your work, there are some alternatives available, like the RTX 6000 Ada. However, the cost of this card is quite high, while a cheaper choice could be the older RTX A6000, which costs $4,650 and supports NVLink and P2P. The NVIDIA A-series, formerly known as Quadro, offers some advantages, such as the use of certified drivers for some professional applications and better compatibility.