NVIDIA Headquarters
Digital storage and memory maintain the data used to supply AI training and to store the models used in inference. During the NVIDIA GPU Technology Conference (GTC) next week, several storage and memory companies will display their latest products and services to support increasing workloads to allow AI. This article will talk about some of these developments from Kioxia and Pliops pre-events. We will have more to say about digital storage and memory developments at the start of the GTC.
Data centers often use several types of digital storage technology to compromise performance needs and storage costs. For example, for artificial intelligence (AI), the training of fast -solving disks (SSD) is often used to feed data with memory, generally dynamic memory with random access (DRAM) which supports rapid access to the data required for effective and effective use of GPUs.
In recent months, most large SSD companies have announced that large capacity SSD QLC SSDs that are intended to provide warmer and hot data storage, will be used a certain secondary storage of the hard drive, especially when storage density in a rack is a major factor for the choice of secondary storage. Kioxia has now joined the ranks of SSD companies offering large capacity QLC SSDs.
Kioxia announced its LC9 series of 122.88 to NVME SSD, targeted for AI applications, see the image below. This SSD is in a 2.5 -inch form factor and is built with the 8 of the companyth Generation 3D QLC 2TERABIT (TB) DIE using CMOS directly linked to the table (CBA) to increase the density on the memory matrix. The reader has a PCIe 5.0 interface and a double -port capacity for greater tolerance to defects or for connectivity to several computer systems. It can provide up to 128 gigatransfers per second.
Kioxia LC9 SSD
The company claims that high -capacity readers are essential for parts of the AI workload, in particular for large -language models (LLM), training and storage of large data sets and for the rapid retiaval of information for inference and model setting.
This new SSD can also be used with recently announced Koxia AISAQ technology which can improve increased generation of scalable recovery (RAG) by storing vector databases on SSDs rather than a more expensive dram.
Pliops, a supplier of semiconductor storage and accelerator, announced a strategic collaboration with the VLLM production battery developed at LMCACHE LAB at the University of Chicago. This battery aims to considerably improve the performance of the large language model (LMM).
Pliops provides shared storage and effective unloading of the VLLM cache, while LMCACHE LAB provides an evolutionary framework for the execution of several instances. The combined solution can recover from the failing instances, taking advantage of the KV storage backend of Pliops. Collaboration has a level of petacto memory under HBM memory for GPU calculation applications. Using the disintegrated intelligent storage, the calculated KV covers are preserved and recovered effectively, considerably accelerating VLLM inference.
Data centers need storage and memory to provide the data necessary for AI training and inference. Kioxia will display their SSD PCIe 5.0 with high capacity and Pliops will display their shared storage used to improve LLM inference performance.