/i/2003840768.png?fit=%2C&ssl=1)
Nvidia shared more details about the GeForce RTX 3000 video cards during a virtual Editors Day. The GPU designer announced its RTX 3070, 3080 and 3090 on Tuesday evening, but it was still unclear how Nvidia had ‘doubled’ the number of cuda cores.
A big surprise came at Tuesday night’s announcement that Nvidia claimed all of its new RTX graphics cards had double the number of cuda cores rumors previously predicted. Even Nvidia’s board partners were surprised by this. The extensive preview on Tweakers described the possible option that Nvidia has given the fp64 units of the server-oriented Ampere variant a different destination. New information from Nvidia shows that things are different.
Nvidia has indeed removed the vast majority of fp64 compute units from the consumer version of Ampere, as it did at Turing. Such calculations are also sporadic in gaming workloads. What Nvidia did, however, was also made the cluster shaders that can perform int32 calculations suitable for fp32. Each calculation unit in that cluster can therefore perform every clock tick either an fp32 calculation or an int32 calculation.
–
In this, Nvidia finds the justification for doubling the number of cuda-cores it specifies in the specifications: it uses the definition ‘number of fp32 units’ for this. However, half of that should also be available for int32 instructions and there are also many other bottlenecks if you suddenly double the fp32 computing power. So users should not expect a doubling of performance, which the doubled number of cores might suggest. Although the intensity of the various types of instructions varies considerably from game to game, according to Nvidia, questions from Tweakers still elicited a statement about the average performance gain. This would be more than 30 percent ‘across the board’ compared to a design in which the int32 cores as a whole cannot process fp32 instructions, such as Turing.
Nvidia also released an official photo of the RTX 3080 Founders Edition pcb. That is particularly short: the power connection that is roughly in the middle of the entire video card is placed at the end of the irregularly shaped printed circuit board. In any case, a lot of effort was made to get everything to fit, as witnessed, for example, by the diagonally placed capacitors, which also had to find space for ten gddr6x memory chips. Previous short PCB AMD cards, such as the Fury Nano, had onboard HBM memory.
Finally, Nvidia pre-sorted the mediocre availability of the new video cards, at least in the beginning. Only last week the first production GPUs were delivered to the various video card manufacturers. Nvidia indicated that production is now running at full speed, but it expects high demand.
–
/i/2003840726.png?f=imagegallery){kind=link}