GPUs possess a seemingly endless appetite for data, which poses a significant challenge in maintaining their optimal performance. Weka recently unveiled a groundbreaking series of knowledge storage solutions, capable of processing up to 18 million input/output operations per second (IOPS) and serving 720 gigabytes of data per second.
The latest GPUs from NVIDIA can harness knowledge from memory at unprecedented velocities, boasting capabilities to ingest up to 2 terabytes of information per second with the A100 model and an astonishing 3.35 terabytes per second with the H100 variant. With the advent of the latest High-Bandwidth Memory 3 (HBM3) standard, a significant amount of reminiscence bandwidth is necessary to train the largest and most complex giant language models (LLMs), as well as support various scientific applications.
To fully utilize the capabilities of GPUs, it is essential to conserve the PCI buses, which necessitates a robust information storage infrastructure capable of maintaining the saturation level of knowledge. The parents claim to have achieved this milestone with their newly launched WEKApod line of smart home devices, which debuted last week.
The corporation is offering two distinct versions of the WEKApod alongside the Prime and Nitro models. Households typically start with collections of eight rack-based servers and store approximately half a petabyte of data, allowing them to seamlessly scale up to support hundreds of servers and numerous petabytes of information.
The WEKApods Prime lineup leverages cutting-edge PCIe Gen4 technology and high-speed connectivity options, including 200Gb Ethernet or InfiniBand interfaces. The system boasts an impressive starting point of 3.6 million input/output operations per second (IOPS) and 120 gigabytes per second of learning throughput, scaling up to a remarkable 12 million IOPS and 320 gigabytes of learning throughput.
The Nitro lineup leverages cutting-edge PCIe Gen5 technology and high-speed connectivity options, including 400Gb Ethernet or InfiniBand interfaces. The Nitro 150 and 180 models boast an impressive 18 million IOPS, enabling blazing-fast read speeds of up to 720 GB per second and write speeds of 186 GB per second.
According to Colin Gallagher, VP of Product Marketing at WEKA, enterprise AI workloads demand extreme efficiency in both learning and writing knowledge.
“A few opponents have recently claimed to be the top provider of knowledge infrastructure for AI,” Gallagher stated on the WEKA website. “While attempting to drive change, they tend to cherry-pick a specific metric, often focused on knowledge acquisition, while disregarding other crucial indicators.” For modern AI applications, a crucial measure of efficiency is often misleading.
As a consequence of the dynamic interplay between learning and writing in AI knowledge pipelines, the need for continuous adaptation arises when AI workloads shift, according to experts.
“According to Gallagher, the initial process of learning involves absorbing information from diverse sources, storing it in memory, preprocessing it, and then rewriting it.” Throughout the coaching process, it is essential to learn how to adapt and replace model parameters as needed. Meanwhile, checkpoints of various sizes are saved, allowing for a thorough evaluation of progress, and outcome metrics are meticulously documented to facilitate in-depth analysis. After coaching, the mannequin produces outputs ready for further evaluation or utilization.
The WEKA Pods leverage the WekaFS file system, the company’s high-performance, parallel storage solution that supports a broad spectrum of protocols. Nvidia’s GPUDirect Storage (GDS) leverages RDMA technology to significantly boost bandwidth and reduce latency between the server network interface controller and the GPU’s memory.
WekaFS offers comprehensive support for GDS, having been jointly validated by NVIDIA alongside a reference architecture, according to Weka. The WEKApod Nitro is also licensed for Nvidia’s DGX SuperPOD.
Weka’s new home equipment boast a comprehensive suite of enterprise features, including support for multiple protocols (File System, Server Message Block, Amazon S3, POSIX, Google Drive Storage, and Cloud Storage Interface), robust encryption capabilities, seamless backup and restoration functionality, snapshotting options, and advanced data security mechanisms.
To ensure the integrity and security of stored data, it employs a cutting-edge patented distributed knowledge safeguarding mechanism that safeguards against catastrophic knowledge loss caused by potential server failures or outages. The corporation claims that its solution offers the scalability and robustness of erasure coding, yet without the efficiency drawback.
“Generative AI applications and multi-modal retrieval-augmented technology have accelerated adoption in enterprises at an unprecedented rate, prompting a need for cost-effective, high-performance, and versatile knowledge infrastructure solutions that deliver extremely low latency, significantly reduce the cost per token generated, and scale to meet present and future organizational demands as AI initiatives evolve.”
“WEKApod Nitro and WEKApod Prime offer unmatched versatility and choice while delivering exceptional efficiency, vitality, and value, enabling accelerated AI projects to run seamlessly anywhere, anytime.”
Associated Objects:
