New technology allows expanding GPU memory with SSDs for use in AI and HPC

A new technology being developed by Panmnseia in South Korea could offer an interesting alternative to GPUs used in AI and HPC. Called CXL-GPU Image, the new technology allows the GPU to take advantage of the PCIe port for memory expansion.

The accelerating growth of the AI ​​segment has resulted in ever-increasing demands on the hardware used in the technology. GPU memory is often one of the weak points in the area, which is why so many companies have turned their efforts to HBM. This type of memory has been the solution to increase the capacity of GPUs in high-performance computing (HPC) and AI.

Panmnesia’s technology, however, offers an alternative that could be more affordable, allowing the creation of future components in the style of an SSD or DRAM, but to expand the GPU’s memory. These could be connected to the PCIe port and accessed by the graphics processing unit with a latency of “double digits in nanoseconds”, according to the manufacturer.

For the technology to work, however, a fully compatible system is required. Unfortunately, it is not possible to install a simple controller on the board or anything like that. Panmnsesia needs to establish partnerships with Nvidia and AMD, for example, so that compatibility can be developed at the GPU level, something that could prove complicated.

HBM memories are here to stay

Panmnesia’s proposed solution is eye-catching, but it is unlikely to threaten the growing HBM memory market. Many companies will likely prefer to buy GPUs with ever-increasing amounts of memory rather than overhauling the entire system functionality to a new method.

It all depends on how much the CXL-GPU Image would actually help a company save on its HPC or AI server in the end. Still, TechSpot points out another sad possibility for the South Korean company: if AMD and/or Nvidia like the solution, they could still try to create their own version instead of paying for its licensing.

Source: TechSpot