As we sat down for an interview in the GTC Expo last week, it was hard not to notice that Sam Werner, the General Manager of IBM Storage, was excited. “Did you hear Jensen talk about Nestlé?” he asked. “The work we had done that was highlighted, and the 30x cost performance improvement?”
Indeed, 15 minutes into his GTC keynote, Nvidia CEO Jensen Huang showcased how IBM’s customer, Nestlé, had adopted Nvidia GPUs to speed up its order-to-cash data mart, which is used to coordinate every order, delivery, and invoice across the food manufacturer’s global operations. Instead of running the 15-minute routine to update the data mart several times a day, thanks to GPUs, Nestlé was able to lower the data mart query time to three minutes, while dropping its costs by 83%, which translates to the 30x cost performance figure cited by Werner.
Nestle’s IBM watsonx data mart is GPU-accelerated thanks to cuDF
But the way that Nestlé implemented GPUs is important. Nestlé uses the IBM Storage Scale System (formerly Spectrum Scale) to house its enterprise data. Thanks to the work that IBM has done to bridge the open source software it’s adopted, including the Velox library and the Presto SQL query engine, with Nvidia’s CuDF, Nestlé was easily able to bring the sheer horsepower of Nvidia GPUs to bear on structured data sitting in its watsonx.data lakehouse.
“It was Scale underneath, which is my storage solution,” Werner told HPCwire. “We help bring all that data to the GPUs with our high-performance storage layer.”
It’s easy to forget how important storage is to AI, whether it’s running the latest large language model (LLM) or running traditional machine learning algorithms. Storage is also critical to running large SQL analytics workloads, as Huang talked about in his keynote with Nestlé.
No matter what workload is proposed, Werner is confident that IBM’s storage solutions will have something to offer. But Werner is particularly excited about the potential for his Storage Scale unit to sell lots of storage in support of AI initiatives, and in fact to lead the industry.
Sam Werner is the General Manager of IBM Storage
Here are seven aspects of IBM’s current Storage Scale offerings that some customers may not know about:
- Content-aware storage: “We automatically vectorize the data,” Werner said. “We have a capability called content-aware storage. So we take that data, we watch it as it changes. And every time it changes, we run it through an AI data pipeline and we vectorize the data. The way other people would normally do it is you’d copy your data to a server, you’d run it on this AI model and vectorize the data, but you lose the connection to the source. So now if the source data changes, there’s no way to know that in the vector database.”
- Storage virtualization: “We don’t make you replatform,” Werner said. “We actually will bring it into our namespace, and then we provide GPU Direct connection into the GPUs. So we give you super high performance, but allow you to leave your data where it lives…. It can be sitting in the cloud. It can be sitting on prem, it can be in an old NetApp NFS file, or it can be anywhere. I’ll put it in my namespace, vectorize it, and make it available without impacting your applications.”
- Multi-protocol support: “It’s S3, it’s NFS, it’s SMB, it’s HDFS,” Werner said. “We’re about to publish our S3 RDMA, which will be the best performance in the industry. We’ve been working closely with Nvidia on that…GPFS technology gives us a huge advantage in the AI space because it’s one of the only true cluster file systems out there that gives you linear scale and performance, no matter how big your cluster gets.” (Now that IBM owns Confluent, you might see Kafka as another supported protocol in the storage stack.)
- IBM Fusion Storage: “I have all this great technology,” Werner said. “I’m putting it all together as a single platform, which we call Fusion. With Fusion, I actually can deliver it as a hyper-converged infrastructure, which includes the servers, storage, the networking, and the GPU nodes, all in an integrated rack running OpenShift. We run bare metal OpenShift on it, and then you can just start building your AI applications and it has Scale in it.”
- FlashCore Modules: “We have FlashCore Modules, which goes back to our Texas Memory Systems acquisition,” Werner said. “We took that big, solid-state technology and we ultimately boiled it down to an industry-standard form factor drive. Just in February we announced our latest generation FlashCore Module 5. What we do is we have an FPGA in there with ARM processor cores, so we’re able to offload a lot of stuff that happens in the storage down to the drive itself. We call it computational storage. And within this drive I do compression of the data. I just added deduplication so we can get five to one data reduction in these drives. And I use QLC memory rather than TLC memory, but I get TLC performance and durability.”
-
IBM Fusion Storage provides a unified solution for modern storage (Image courtesy of IBM)
NAND Shortage: “There’s a NAND shortage?” Werner joked. “It’s actually a really great advantage for us. I have great relationships with the vendors that I buy memory from, and I’ve already secured memory supply in advance. So I’m in a very good position from a supply perspective. Are my costs going up a bit? Of course, memory prices are going up, but we’re able to manage that at a significantly lower rate of increase than what you’re seeing happen with industry standard drives. So I’m in a really good position. I have plenty of supply and my costs are much better controlled than industry standard.”
- Tape for AI: “I have seen demand already go up 2x to 3x and it’s still increasing,” Werner said. “I am working very closely with my ecosystem to significantly increase the amount of capacity we have, which we’re able to do no problem. Think about an enterprise that’s trying to tap into all their data forever. You don’t ever have to delete any of it. You have an S3 API right here, move your data there, I keep track of all the data, make it searchable, and you can retrieve it back. Tape has really high throughput. We have lots of drives. We can stream the data. It’s only challenge is if you’re doing random reads from all over the place, then yes. But if we have a catalog that makes it searchable, we can make that data really easy to retrieve for your AI when you want it. And we can also make it easy to save all the tokens that you’ve created forever, so you can go back to it. You don’t have to recreate those tokens later.”
Like every storage vendor on the planet, IBM is working with Nvidia on its new STX platform, which uses BlueField-4 DPUs to implement Nvidia’s Context Memory Storage (CMX) architecture to break through the GPU memory wall for large-scale AI inference. The more work that AI users can push down into the storage layer, the less work there is for the GPU, and the greater the throughput can be.
(Nick-N-A/Shutterstock)
“AI is triple-digit growth for me in this space. It’s a huge growth driver,” Werner said. “AI has finally given customers the opportunity to unlock the value of this enormous amount of data they’ve been sitting on…Now they’re able to unlock that value and I’m able to help them do that. So I’m able to show a really good ROI and people will be able to do that.”
IBM may not be the first storage vendor that comes to mind when IT decision-makers sit down to hash out an AI strategy. But Werner says there’s no other storage vendor can match IBM in offering such a wide spectrum of storage capabilities. As the AI boom continues, Werner is confident of the hand that Big Blue is holding, not to mention a few wildcards that play in its favor, thanks to its experience and its legacy of supporting the IT endeavors of the biggest organizations on the planet for more than 100 years.
“The biggest banks in the world run on my stuff because they know my stuff is secure and I can do this,” Werner said. “A lot of these others don’t have that experience. And at some point people are going to wake up and go, holy crap, my AI application needs to be highly available. It needs DR, it needs all of this because this is mission-critical.”
This article first appeared on HPCwire.
The post Seven Reasons IBM’s Storage Chief Is Bullish on AI appeared first on AIwire.
