(Davarti CGI/Shutterstock)
AI is still in its infancy, but looking at trajectories such as ChatGPT, Pure Storage makes it very clear: Since you need to store and serve large amounts of data to train AI models on GPUs, you will almost certainly need large amounts of fast data. Generate an all-flash array.
Large Language Models (LLMs) have given the world a taste of what’s possible with AI, but there’s still a lot more to do, says Shawn Rosemarin, vice president of research and development at Pure Storage.
“The challenge here is that the vast majority of companies are trying to gather information from data sets that are not available on the open internet, some of which are highly sensitive, highly secure, and highly restrictive,” he said. That’s it,” says Rosemayne. Data Nami. “And you have to train on all that data yourself to be really useful.”
AI models like the ChatGPT function have given us a kind of inference engine. This is great, says Rosemalin. Pre-trained conversion models like ChatGPT gave us another avenue, without the need for humans to absorb large amounts of data to be able to understand it and ask questions about it. .
The next step, he said, would be to apply the same technique to private company data such as radiation therapy records, transaction records and oil reserves. That requires massive increases in storage and computing.
“It puts a lot of strain on storage. The tapes that hold most of this aren’t fast enough to be parallelized. The hard drives aren’t fast enough to be parallelized,” says Rosemalin. say. “Customers clearly recognize that storage is the bottleneck to getting the most out of their GPUs. , requires a large amount of storage.”
Companies that originally considered flash as a performance storage tier may need to rethink their approach and move to flash as their primary data store, he said. The flash array will continue to feed the training data to the GPU and be ready to handle all other data tasks required to train the AI model.
“You have to think of this training concept as very data intensive. You have to get very large datasets. Literally, it means labeled, ideally labeled information,” says Rosemalin. “And then you can feed it to these GPUs and train models.”
Not only do large data sets require more storage, but training LLM on large data requires more performance and more IOPs. All of this points to a future where ultra-fast flash arrays will become the standard for AI model training.
“More parameters means we need more IOPs, because we have more IOs per second to be able to actually train those models,” he says. “The GPU consumes as much data as I put in, so performance becomes essential. And most of the time, getting enough storage for the GPU is actually a big problem. And , there is parallelization of all these data services, potentially thousands of GPUs are all short of storage, they all want to give storage in a very short time, and others No one wants to wait for it to be completed.”
Naturally, Rosemarin believes Pure Storage has an internal trajectory that can meet this pressing demand for fast storage for AI training. He points to the fact that the company manufactures its own disks, or Direct Flash Modules (DFMs), from raw his NAND that it procures from suppliers, which he says gives Pure Storage more control. says. He noted that the company is developing its own operating system, Purity, which also gives it more control.
Pure Storage also leads in terms of capacity, says Rosemarin. Pure Storage’s roadmap says he needs DFM at 300 TB by 2025, while other flash providers’ roadmaps only have him up to 60 TB, Rosemarin said.
Pure Storage works with some of the world’s largest AI companies, including Facebook’s parent company Meta, and supplies storage for Meta AI’s Research Super Cluster (AI RSC), one of the world’s largest AI supercomputers. doing. Pure worked with Nvidia to come up with an AI-Ready Infrastructure (AIRI) solution built on the Nvidia DGX BasePOD reference architecture for AI and including the latest FlashBlade//S storage.
At this week’s Pure//Accelerate 2023 user conference, Pure Storage made several announcements, including new features to its FlashArray//X and FlashArray//C R4 models, as well as ransomware protection for Evergreen//One storage. I did. Providing -as-a-service.
According to Pure, FlashArray//C R4 models offer up to 40% better performance, 80% better memory speed, and 30% better inline compression. The FlashArray//C line will include a 75TB QLC DFM, X product, and the FlashArray//X line will ship with a 36TB TLC DFM, the company said.
Meanwhile, a new service level agreement (SLA) for Evergreen//One storage services offers customers certain assurances after a ransomware attack. Specifically, the company says it will ship clean storage arrays the day after the attack at the latest and will work with customers to finalize recovery plans within 48 hours.
Related products:
Why object storage is the answer to AI’s biggest challenges
LLM is Dinosaur-Killing Meteor for Old BI, Says ThoughtSpot CEO
Like ChatGPT? Haven’t seen anything yet
