Memory capacity and bit width are very important for GPU graphics cards, and it is very expensive to rely entirely on DRAM chips. SSD hard disk is mainly connected to CPU through PCIe, and GPU can't directly access hard disk data, so BaM technology bypasses this restriction!

It is reported that NVIDIA and IBM jointly developed BaM (Large Accelerator Memory) technology, which allows the GPU to directly bypass the CPU limit and directly connect to the SSD hard disk. If you have 2TB of SSD, it is 2TB of "memory".

Similar technical concepts have existed for a long time, and BaM technology jointly developed by NVIDIA and IBM is one of them.

According to their published papers, BaM technology does not depend on traditional virtual address translation, and it is no longer CPU-centric. It directly allows GPU to get data from memory and storage and process it, without the need for CPU kernel to assign tasks.

The goal of BaM technology is to expand the memory capacity of GPU, improve the effective access bandwidth, provide advanced abstraction for GPU, and enable GPU to access the massive data in the extended memory on demand and in fine granularity.

This BaM technology is easily reminiscent of Microsoft's recently released DirectStorage technology and NVIDIA's RTX IO technology, which are similar in nature, and also allow GPU to bypass CPU restrictions and directly access data on SSD, greatly shortening the loading time in the game, and even entering the game in seconds.

However, BaM is currently mainly used in the field of GPU high-performance acceleration, including GPU computing and machine learning, and it is still in the prototype stage, and the specific application time has not yet been determined.