Data Protection & AI Summit | Data Protection & AI Summit Panel

Clips
News
More from Data Protection & AI Summit

Christophe Bertrand

Principal Analyst

SiliconANGLE & theCUBE

Mitch Seigle

CMO

Spectra Logic

Vinny Choinski

Founder & Principal Analyst

Triptych Info

David Feller

CTO

Spectra Logic

Tape storage returns to the spotlight as AI reshapes data priorities

As artificial intelligence continues its dizzying ascent, the unsung hero of the digital age is making a comeback: Tape storage. Once dismissed as outdated, tape storage is now being hailed as a strategic asset for the exponential growth driven by AI.But what, in particular, is spurring this full-circle moment? And how is tape storage set to bolster AI’s position as a critical business enabler?“We’re almost coming full circle at this point, and I think that’s why we should be paying attention to tape,” said Vinny Choinski (pictured, right), founder and principal analyst at Triptych Info LLC

play_circle_outline Exploring the Role of Tape Technology in AI: Insights from Data Expert Vinny Choinski and Panelists

play_circle_outline Revolutionizing IT: The Vital Role of Tape in AI Support and Modern Workflow Management

play_circle_outline Maximizing AI Efficiency: The Critical Role of Tape in Long-Term Data Retention and Cyber Resilience

play_circle_outline Relationship between big data, HPC, and AI, emphasizing tape's scalability benefits.

play_circle_outline The overall consensus: tape technology is essential for effective AI data protection.

play_circle_outline David Feller highlights Spectra Logic’s role in data protection and storage for AI.

play_circle_outline Discussion of geographical boundaries affecting data management and compliance regulations.

Info
Transcript

Data Protection & AI Summit Panel

Christophe Bertrand

Principal Analyst SiliconANGLE & theCUBE

Mitch Seigle

CMO Spectra Logic

Vinny Choinski

Founder & Principal Analyst Triptych Info

David Feller

CTO Spectra Logic

In this Data Protection + AI Summit session, Vinny Choinski, founder and principal analyst at Triptych Info, joins Mitch Seigle, chief marketing officer at Spectra Logic, for a wide-ranging conversation with theCUBE Research’s Christophe Bertrand on why modern AI pipelines cannot scale without tape. Choinski draws on three decades in data protection to explain how object-aware tape libraries bring cold data back into high-performance workflows so teams can train models on complete, unsummarized datasets. Seigle highlights Spectra Logic’s sustainability metric... Read more

explore Keep Exploring

What is Vinny Choinski's background and expertise? add

What is the role and evolution of tape in IT systems and its significance in current workflows? add

What are the perspectives on the role of tape technology in relation to AI data processing and data growth? add

What is the relationship between big data, high-performance computing (HPC), and artificial intelligence (AI)? add

What are the challenges related to power consumption and infrastructure for data centers in the context of AI development? add

What is the impact of using spinning discs versus tape on power utilization in data centers? add

What are the solutions for managing and protecting large data sets, particularly in relation to compliance and geographical boundaries on the data? add

bolt Powered by CUBE AI

Data Protection & AI Summit Panel

search

Christophe Bertrand

>> Hello, everyone. Welcome back to the Data Protection and AI Summit. My name is Christophe Bertrand, Principal Analyst at theCUBE Research. Today we have great panelists, and we're going to be talking about tape in this segment. I'd like to welcome our panelists joining us from the Boston area, Vinny Choinski. Vinny?

Vinny Choinski

>> How you doing, Christophe? Nice to be here today.

Christophe Bertrand

>> Tell us about yourself.

Vinny Choinski

>> Sure. So I am the founder and principal analyst with a company called Triptych Info, focusing on really creating technical content for the data management and data protection space. I started this company a couple of years ago after I took a little sabbatical, and it grew naturally when some of my old customers asked me to help them with some content. So I've been focused on data protection for over 30 years at this point.

Christophe Bertrand

>> Thank you, Vinny. David Feller joining us from the beautiful mountains of Colorado, it looks like.

>> Yeah, thank you for having me. We're out of Boulder, Colorado and I'm the CTO of Spectra Logic. We're a 45-year-old company in the data protection and storage industry, and we're going to talk to you today about how that fits in the AI world.

Christophe Bertrand

>> Absolutely. And joining me in studio, Mitch Siegle. Mitch?

Mitch Seigle

>> Well, good to be with you here in the studio. Christophe. I'm Mitch Seigle, of course, CTO, CTO, gee. David, I'm after your job, chief marketing officer here at Spectra Logic, and obviously in need of a little coffee.

Christophe Bertrand

>> All right. Well, we're going to talk about tape today, but we're going to have some fun because it's actually a fascinating topic, but we've known each other for a long time here. And as you can see, although the technology seems like it's been around for a long time, which it has, you could argue it's mature, but it turns out, and I'll make a statement here that I'll make again, if you're thinking about AI, you'd better be thinking about tape. So why would I say this? Vinny, let's start with you. Tell me why would I say this?

Vinny Choinski

>> Sure. So I hate to date myself, but I started with tape when mainframe systems were booting from tape. It wasn't just data protection, the tape was actually used to support the OS, make it run efficiently. There was the introduction of open systems where I think tape kind of got relegated to data protection only. And I think we're almost coming full circle at this point, Christophe, I think that's why we should be paying attention to tape. People are starting to try to figure out how to get some of the efficiencies of tape into a lot of their IT workflows at this point, right? And I think that AI would be top of priority for most companies. I think they should not overlook it, their AI initiatives when it comes to tape technology.

Christophe Bertrand

>> Well, David, same question for you. We're going to go around the table. Why would I say something like this? Why tape and AI?

>> So it's really interesting. There are a number of different usage for storage in the AI world, and the big conferences everybody talks about how do I keep those GPUs loaded up as dramatically as I can because it's all about the cost of the GPUs and the data centers. But the next step beyond that is what do I do over the next 30 days, in the next six months, the next 10 years? And there's plenty of applications if you're a self-driving car manufacturer, or you're in the space of dynamically creating video or audio or even pictures, you need to retrain your models. You need to make sure that you keep those models long term. And you might even have the government breathing down your neck saying, "You got to retain that stuff for 10 or 20 years for liability sake." And tape in particular is the low cost. And let's be very clear, the low power leader, that very much complements what's going on with the rise of power in the GPU space. So it's the perfect yin and yang to the AI systems that are actually processing the data.

Christophe Bertrand

>> So, Mitch, what's your take?

Mitch Seigle

>> Well, of course I have to agree with David. He is after all the CTO, but he makes some key points around the sustainability angle of it. I think there's a resilience angle to it. That tape is probably intrinsically the most cyber resilient technology with built-in true air gap capabilities and mutability, that's a very important thing. Sovereignty, of course, as the value of the data goes up, the need to protect it and maintain control over it is important. But I think the big angle here is that the scale of data growth being driven by AI will be exponential if not logarithmic. And so we're going to see a very sharp spike in growth and the resolution at which we're going to want to retain all of that data to allow AI to actually do things that human beings can't in analyzing all data to discern patterns and trends that a human might not be able to see is of critical importance. And the only technology that can offer that scale into the multi-exabyte range today is tape technology.

Christophe Bertrand

>> It's a fascinating topic, and you talked about a lot of data, brings up memories of big data and well, let's talk about that actually. So what's the relationship between big data, HPC, high-performance computing and AI? And, Vinny, I'd love your take on that.

Vinny Choinski

>> Yeah. With big data I think tape offers the ability to bring in bigger data sets, right? It's always been leveraged for big data sets, right? So I can bring in a big data set, a complete data set. Maybe I want a long history. I don't have to kind of look at the data, cull it before I bring it into my AI, and I can run my AI prompts and my AI algorithms against the full data set if I want to. And then I can bring it into higher performance, higher cost disk, but then I can easily push it back, right? So I think the ability of tape to be near line in an automated fashion, and one of the biggest things I've been excited about is the ability to have object and file system in tape where I don't have to necessarily use it as an archive. Before even companies were doing this, I think people were doing the DIY, do-it-yourself. They were creating project folders, backing them up, deleting them off primary storage when they needed to work on them again, they called them back. Now we have technology from companies like yours that allow you to automate that.

Christophe Bertrand

>> Thank you, Vinny. So, David, what would you add to what Vinny said? I mean, it makes a lot of sense to me, but there's actually I think a lot more.

>> Yeah, and a couple of key points that he made. One is this workflow isn't new. Spectra's been servicing the high-performance computing market for literally decades. All the big national labs, weather, service agencies all over Europe and the US in particular and Asia. And it's a simple workflow. You want to simulate some sort of data, pull back large data sets, create the simulation and then save that off for a long period of time. So just like an AI workflow, there's a powerful compute. In the past it was big super compute clusters based on literally CPUs. You had big cache, so essentially really fast disk and you have long-term storage. It's essentially the same workflow. The biggest differences are I need ready access to that data, I need more intelligent storage and I need industry standard interfaces. So the whole advent of the public sector cloud, AWS, pioneering S3 for instance, and giving access to that to all the rest of us in the world makes it very, very simple to access that long-term storage and be able to recover something that you saved 20 years ago that maybe a grad student was working on. That same workflow is in place now. It's just highly compressed. I want to do that every 30 days and load up my GPUs, and I need to create vectorized data in a way that can be accessed very efficiently. So it's the same workflow, it's a bit more aggressive, but it's all made possible by the standardized interfaces, for instance.

Christophe Bertrand

>> So, Mitch, HPC and AI, maybe you have a couple of customer examples you could give us because I know Spectra has amazing customers, something we talked about recently actually in this summit. Can you tell us more?

Mitch Seigle

>> Sure. Well, I think to kind of amplify some of what David and Vinny are talking about, the modernization of tape has really been the key to linking tape very strongly to where AI is and where AI is going. And I think then the external trend, which will be regulation driven by governments, will require that data to be retained literally at the atomic level. We have many, many customers, certainly in the high-performance computing space, national labs are a great example of that. One of the most visible on a global basis is CERN, the home of the Large Hadron Collider in Switzerland. And so we think about the scale of data being driven in the high-performance computing environments today as allowing us as an industry in partnership with our customers to solve these scalability problems, these retention issues, efficiency concerns in a way that is directly applicable to AI. I look at AI as from a pure infrastructure perspective as being a super set of what we have always been doing over the last 25 years or more in the high-performance computing space. And so great examples. Imagine today where some of these use cases are tracking satellite imagery, celestial data, and so on, and then adding an AI layer on top of that to analyze that in ways that human beings and current technology is unable to do, to discover new galaxies, to discover new planets. Only recently there was news that there may be yet an unseen planet in our own solar system that is now just becoming noticed because of our ability to analyze the data, think about the power of AI unleashed on that, but think about the amount of storage that it takes to retain the data to be able to process that. And so they go hand in hand, no data, there's nothing to reason about. And so if you don't preserve the data, if you can't scale the data, and if you don't store it atomically down to the smallest finite element rather than summary data as we've seen often, for example in the big data or data warehousing world, summarization goes against the grain of AI being able to analyze things with a precision point. And so I think that's the key thing to watch for.

Christophe Bertrand

>> Right. Actually, I'd like to continue with my next question with you, Mitch, because you just mentioned large amounts of data. So why is tape likely the only medium that's going to be able to scale in the age of AI? Why is that?

Mitch Seigle

>> Well, I think you've got some very pragmatic aspects. One of the things that I think most people are keenly aware of is that we're running out of power today. And when I say, "Running out of power," it's that the demand for power to feed AI is actually outpacing the growth of new power coming online. In certain urban areas, there may be five-year waits to actually begin construction of new data centers due to the lack of power infrastructure. There are also certainly floor space concerns. If you've got an urban data center, you may be out of room in which you have no more room to install equipment. And so one of the important things that tape can do for you in the context of AI is actually consolidate cold data, which can be as much as 80% of all data, and actually install that on a tape tier, which is highly dense, highly energy efficient, and therefore return power to your budget, if you will, to be able to deploy more GPU infrastructure to enhance your ability to compute. And so I think power is the great equalizer. Those who have power and are taking steps to actively manage it, will have a competitive advantage over those who are not actively managing their power by moving cold data off the tape.

Christophe Bertrand

>> So that's fascinating. That's a big topic. And actually, Vinny, I know it's an area you've covered in the past. We worked together in the past and we talked about this quite a bit. So, Vinny, I'd like for you to give us your perspective on this topic.

Vinny Choinski

>> So, yeah, I agree. I think probably the last paper I did prior to starting for the tape consortium was a environmental impact. I did quite a bit of modeling. I actually took a workload and anybody that's in the data center is a concern with space and power cooling. So I tried to model it pretty significantly, pretty detailed, and I compared it at that point in time, probably about a 500 terabyte grade array solution with the same amount of use space that was taken up in the data center compared to the same use space in a tape library. And some of the modeling I did, it equated the data that I have off the top of my head is it equated to about a 6.5 X reduction in CO2, but that CO2 is a calculation from power. So that means it was using a significantly less amount of power to run that tape with the same data profile and keep it running for a 12-month period. This was over a 12-month period. So tape has, besides being cost-effective on the amount of storage you can put on it, you can't overlook the power consumption. It's not drawing power, it's near line when you're not reading from it, right? So there was a really significant impact on the power utilization between the spinning disc and the tape for the same workload.

Christophe Bertrand

>> Thank you, Vinny. So, David, is this true? You're a CTO, so you should know all about that?

>> Absolutely. In fact, some of the latest NVIDIA systems, amazing as they are, can take up to 120 kilowatts per rack. And you start stacking that up and you get into, I'm going to basically take the bulk majority of the power from my local power generation. That creates problems all over the place. Small dropouts can cause tremendous problems in the industry. But again, the perfect counter to that, and a better way to express it is we've got customers that in the past have been floor space constraints. So we go in and say, "Hey, you need whatever, an exabyte of storage, it'll take this many rack units to do it." In the past it was a problem of can I fit that in physically? Now that's not the problem. Now the problem is do I have enough power to run my data center? And in fact, I'm limited by power so there's practically free data center floor space and a number of these big data centers that you need to add an exabyte. Well, you can't add an exabyte of disk cost effectively or power effectively. You certainly can't do it with NVMe and flash storage, but a fraction of that power consumption can be used for a tape system. So I can get an exabyte in a floor space that's available and counter that power draw essentially. So again, it's the yin and yang of power in the data center world.

Christophe Bertrand

>> Well, I'd like to stay with you, David, because I know you've done some work on use cases specifically for tape in the context of AI. So I mean obviously we can see the benefits from a power perspective, a green perspective, space. There are other aspects linked to performance that maybe we should talk about too. So tell me, in your research, in your work, what are the top three or the top five use cases you're actively working on with your customers?

>> Yeah, so this is actually really interesting. I think most people have a misconception when we talk about tape, what's it capable of? And tape actually has the ability to have significantly more performance than anybody is even aware. So there are a number of use cases. The short-term use cases are I'm going to retrain my model every 30 or 45 days and release a new version of my model. Maybe I'm doing a commercial version of something that creates some sort of product deliverable for a customer using AI. I want to collect that information over a 30-day period, 45-day period and then retrain my model and then save that model for long-term retention. So I need a really fast capability. I might need to save a hundred petabytes a month, and I might need to do that really effectively on some combination of disk and tape storage. We've got a lot of customers that need long-term retention because of compliance. And I mentioned at the beginning, self-driving car capability. I don't think that's clear. There are tens of exabytes at multiple different customers that are required to save the raw data from all of their input from the cars and the models that they're creating. And depending on the country you're in, anywhere from 10 to 18 years of forced retention, combination of legal and government regulations. But there's also the desire to retain models long-term just for the sake of what if my model goes wrong somewhere along the lines? what if a year down the path I realized that I've been retraining and that model's been going in the wrong direction. I need to step back six months and do the training again. I need that raw data, I need the models and I need the simulations from it going back a significant period of time. So there's at least three models that are perfect combinations of disk and tape. And again, Spectra has both to offer in one seamless solution. So it's a lot of fun to talk to companies about how you're going to retain this for more than two weeks. That's always the question. If you're going to retain it more than two weeks, there's a play for tape in this.

Christophe Bertrand

>> So talking about use cases, Mitch, I'd like to return to you now. What could you add to what David said, and maybe we should talk a little bit about some of the basic use cases around just protecting the data itself, cyber resiliency and compliance being part of it, as David was mentioning.

Mitch Seigle

>> Well, I think to add to what David's talking about, certainly the examples of training in large language models are really in the context of Generative AI. There are many more AI use cases. Of course, there's machine learning, vision and visual data, if you will, that are going to be critically important to a lot of national security applications, to robotics applications of all types in society. And the more that these things become accepted in society, the more data that that will drive, right? The more people that are taking a robo taxi, for example, will beget more robo taxis and automation and new forms of that. So I think the scalability is a key use case. I think the ability to retain the data for regulatory purposes, we don't yet know the full extent of regulation, but we're in the beginning stages of this. And as I think we talked about the other day, yesterday on our session, you mentioned that there are 75 laws already on the books, and I think that's just the tip of the iceberg in terms of regulation. I think that the primary use cases will simply be to protect the data, but to preserve it for future optionality. Things that we can't imagine that AI will be able to do in the future where AI itself is actually self-developing and self-learning new capabilities that it can bring to bear on the data. And so I think a few of the things that I would say would be important for folks to think about would be how atomic they want their data to be. What is the finest granularity that they want in terms of preserving their data and the implications for the scale and the cost of doing that? And to anticipate the need to preserve everything before somebody mandates you to do it. The last thing you want to have to do is respond and scramble. I think it's pretty important that having that data available and developing applications to facilitate compliance audits so that humans don't do it would be an important use case as well. So a lot of adjacent use cases to support the infrastructure and primary use cases to provide storage that AI can then subsequently access for reasoning purposes.

Christophe Bertrand

>> And I bet a lot of our viewers haven't thought through what you just said. We're talking exabytes. So talking about use cases, Vinny, I'd like to get your opinion on this topic. I mean, do you agree?

Vinny Choinski

>> Yeah. I mean these solutions are there to handle large data sets, right? I think the combination of tape and disk and the object-level file systems allow you to manage big data sets, right? But then you want to back that up and protect it. And I mean, who wrote the book on immutability and air gaping, right? There's really every data protection company that doesn't leverage tape is trying to figure out a way to have air gaping and immutable copies that have always been inherent in tape. So I think that the use cases make a lot of sense, the big data, making sure it's protected, the big amounts of data set that you have to manage. I guess one of the things that I would ask, you mentioned a little bit about compliance and being so close to the customers. I was curious how geographical boundaries on the data are affecting this. I'd been very interested in some of the compliance regulations in Europe, not being able to move data sets outside of a geography. Are you seeing an impact there and what is Spectra Logic doing to help customers manage that?

Christophe Bertrand

>> Great question. I don't know. Let's start with you, Mitch, and then we'll go to David.

Mitch Seigle

>> Well, Vinny, that's a very prescient question on the topic is that residency of data is critically important. Generally there's a trend to not want to let that data out of your national boundaries and sometimes state boundaries for all sorts of legal reasons as well as to protect the data. One of the things that we're actually doing and we've pioneered is tape as a service. And that's the idea of taking object-based tape, a tape library that is abstracted to support object storage as we've been talking about, and placing that into public cloud infrastructure or into co-location data centers or other private cloud spaces so that you can actually determine and document the residency of the data and make sure that you're taking care of all of that. And so having that resource available in the public cloud, we're rolling that out now in partnership with a company called Geyser Data and making that available directly in the cloud as a service where you're fundamentally storing data on tape on a rental basis where you're renting that capacity without actually having to manage the tape infrastructure or quite frankly, even fully understand the tape infrastructure. So that's kind of a new innovation in the space. We see a big role for that both in the cloud in certain use cases, but ultimately even setting it up as a service that is consumable inside your own enterprise centers to make it very easy for the AI practitioners to protect their own information without requiring an IT layer of storage specialists to manage that for you.

Christophe Bertrand

>> Yeah, it's a great solution because I think what you're doing is you're really democratizing the use of tape at scale for organizations that may not have the skill set, may not know how to set things up because it's a different technology they may not be aware of. The truth is, you're going to need tape. Like it or not if you're doing AI, you're going to be doing AI at scale. That's the name of the game, and you're going to be using tape. So what I'd like to do is maybe ask for your closing thoughts. Each of you. I'm going to start with you, David. What advice, and what should people do next? What closing thoughts do you have on all of these great topics we've just discussed?

>> Absolutely. So a couple of key things, and as a storage provider, we're very interested in making sure that everybody understands what the role of on-premises storage is in your particular organization. And as you get to scale beyond just a few petabytes, having that data at your facility or in an MSP close using a tape as a service is a much better choice economically from a data serenity and an ownership perspective than using, for instance, a true public cloud offering. It becomes a battle of marketing of how much does it really cost to save things into the public cloud? And what we found is the ability that you get is almost never what you expected. And so that's a key point of what does the cost look like? What does the power consumption look like for your particular use case, and how much do you really own that data? It's really important that everybody understands that data protection is more than just air gap, and it's more than just having the right backups. It's making sure that you own the data that you can afford to keep it, and that it is available to you in decades in the future if you need it. So we're all about making sure that data protection has a much broader definition for customers than they're used to.

Christophe Bertrand

>> That's a great perspective, and I do agree, actually, we should definitely expand this definition. So definitely discussion I've had many times with Vinny. So, Vinny, what are your closing thoughts on the topic?

Vinny Choinski

>> Yeah, so I'm thinking that customers, I think there's probably a lot of customers that are doing some of their AI on-prem. And a lot of the smaller customers can move smaller data sets in and out of the cloud and run their applications. But financial firms that have built their own analytics and things like that are probably doing a lot of their AI operations on their own infrastructure in their own data centers so that they have control over the data. I'd be interested in following the trend of what percentage of AI is happening in the cloud. Is there any kind of move to doing it on-prem? I got to believe that geographical boundaries and actually physically moving data that is very large. Tape's always been portable. How do you do that when these data sets get to that size? So I'm kind of following that trend to see if things are changing where and when or how the AI processes and prompts get done on the data.

Christophe Bertrand

>> Well, thank you, Vinny. And, Mitch, we'll turn over to you for your words of wisdom and closing thoughts.

Mitch Seigle

>> Well, I would remind people or encourage people to take a very long-term view. AI is really, it's the end of the beginning if you want to coin it that way. So today, where there's a race to deploy all of the infrastructure and to start to extract value from it, I do think you need to think decades into the future and start to establish now your policies on retention. Everything begins and ends with the data. AI is just the way that we're going to process that data. And so like anything else, you need to start and define what is the granularity or the atomicity of the data that you want to preserve. And scale it from there into what are the data protection implications, the scalability implications, the cost, the sustainability. And so establishing strong policy upfront because you never get a second opportunity to recover data that you didn't preserve to begin with. And so if you make the mistake upfront of not preserving something, the realization of what that costs you may not be apparent for decades to come, and you don't want to subject yourself to that if you can do that work upfront now.

Christophe Bertrand

>> It makes a lot of sense to me. And again, we covered so many topics. I'd like to really thank you, the three of you. So, Vinny, David, Mitch, it was great. Great panel. I think we'll keep chatting about a lot of these topics. Again, I'll reiterate, there is no data protection in AI without tape. There is no AI without tape, you better get on it quickly. So thank you so much everyone for watching us. My name is Christophe Bertrand, Principal Analyst here at theCUBE Research. Thank you for your time.