NVIDIA’s GTC conference showcased what AI models can achieve. Memory startup MemVerge claims that its software can maximize GPU utilization. This could open new doors in training AI models.
How will AI training change with MemVerge?
In its presentation at the GTC conference, MemVerge demonstrated how it can accelerate language model training using Compute Express Link (CXL) technology. Working with Micron, the company tested its memory tiering software, Memory Machine X, on various AI chips from NVIDIA.
The company’s software was tested on a Supermicro server rig equipped with an NVIDIA A10 GPU, Micron DDR5 memory and CZ120 CXL. The test compared GPU+DRAM with MemVerge’s GPU, CPU and CXL data layering software.
Thanks to Memory Machine X’s management of memory traffic, there was a speed increase of up to 50 percent compared to standard methods. GPU utilization also increased by half to over 91 percent.
Charles Fan, CEO of MemVerge, said it is possible to scale AI models in a cost-effective way. Saying that the company “feeds GPUs with data” with its own software, the CEO underlined that they maximize efficiency.
While the importance of the chip in training AI models is undisputed, various software can also play a critical role in this training. MemVerge aims to work with NVIDIA and Micron to optimize this hardware.
{{user}} {{datetime}}
{{text}}