NVIDIA and IBM work to communicate GPUs and SSDs directly, without relying on the CPU

NVIDIA, IBM and several members of renowned universities are working on the development of a technology that would allow SSD drives to communicate directly with the GPU. I am sure that more than one of our readers will have thought about DirectStorage technology, an important advance that allows decompressing data coming from the SSD to the GPU, thereby avoiding the bottleneck which represents the classic process of doing that work on the CPU, and yes, in general terms the goal is similar.

However, what NVIDIA and IBM have on their hands is something much bigger, and it is known as BaM, an acronym for “Big Accelerator Memory”. With this technology it is possible to overcome one of the most important problems that derive from the dependency of that relationship between CPU and SSD: excessive CPU-GPU synchronization and an increase in I/O traffic, which reduces bandwidth. available storage unit cash for emerging applications with fine-grained data-dependent access patterns, such as graphing and data analytics, recommendation systems, and neural graph networks.

By allowing direct communication between the GPU and the SSD, it is possible to significantly improve performance, and overcome the traditional problems that derive from this CPU-SSD binomial. This way, the GPU can directly access data present in the memory and the SSD drive, without these having to go through the CPU beforehand, and without depending on said component. It goes without saying that this also entails a significant release of resources at the CPU level.

The implementation of this technology is still in the testing phase, but we have already seen an experimental prototype used in a Linux-based system, where the results have been, in general, very positive. BaM no longer depends, thanks to the software-managed GPU memory cache, on virtual memory address translation, which means that it is not affected by failures typical of serialized events, such as TLB leaks, for example.

In addition to this cache, which works on the GPU cores, this NVIDIA and IBM technology also works with a software library that allows the different processing threads of the GPU can directly request data from NVMe SSDs communicating with them without any intermediary.

NVIDIA and IBM work to connect GPUs to SSDs directly, without relying on the CPU

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *