Podcast Archive - StorageReview.com - Podcast #154: Oregon State’s Chris Sullivan on 122TB SSDs at Sea, Spotted Owls, and AI as a Data Problem

Episode Date: October 12, 2026

Brian sat down with Chris Sullivan, Director of Research and Academic Computing in Oregon State University’s College of Earth, Ocean, and Atmospheric Sciences, for a live podcast at Yotta 2026 in Las Vegas. This is an audio-only episode, recorded on the show floor. Chris has been a steady collaborator on our edge AI work, from The post Podcast #154: Oregon State’s Chris Sullivan on 122TB SSDs at Sea, Spotted Owls, and AI as a Data Problem appeared first on StorageReview.com.

Have a seat, Chris Sullivan. Thank you. Brian, how are you? So we did this at Supercompute 2026, or wait a minute, 25 in St. Louis. And as we look out at the crowd here, this is a very busy Yada event. Lots of people here. Yeah, this is a beautiful infrastructure conference. And that's what we need to talk about.

Everybody's having conversations about infrastructure and how they're being used. And we want to show people how to use infrastructure in a really positive one. Well, you have a very positive message, and you guys are doing something a little different than most of this audience in that your stuff isn't necessarily ephemeral. You're collecting lots of data over several instruments from ships in the ocean to little edge things in the forest. But you want that data and you want it forever because you never know when inference or the rest of AI is going to catch up to help you leverage that data. Yeah, people need to realize that data is the most important part of science. They say, I'm a data scientist, all scientists or data scientists, okay? In the end, data is how we drive science.

It doesn't matter if you're in the lab or you're talking about compute. All of it generates data and all of it has to be stored. And universities and researchers have traditionally been very good at curating, managing, and helping prove that that data is valuable. And I think that we're moving into a time where we really want good data stewards in the future. and universities can play a large part in that. I certainly can, but I saw a video the other day with 10 AI doxins that were rolling over in a field, and that used AI.

Is that the same thing that you're doing? No. So let's talk about some of the projects that we have going. So we're using data centers to help us understand the world around us and help us live better lives on this planet, for example. And our mission at Oregon State is prosperity widely shared. And so in the end, that's what we're really trying to do. I work, for example, with the Forest Service, the United States Forest Service, around a project that monitors endangered species, specifically the spotted owl. And spotted owl is one of those umbrella species that help us monitor the forest and understand how they're going change.

And in the end, we don't want them to collapse or go into monodonance. We need to have a lot of diversity. I know that's a bad word, but a lot of diversity in the forest to keep and maintain the forest the way that they are. Forest are very important to Oregon. We farm trees. and because of that, we need to make sure that they're, you know, well-managed and well-taking care of. And it's ideally always there.

Yeah, and so the project that I'm talking about actually enables us to farm those trees, and we communicate that information to industry. So, you know, lumber industry can actually cut down trees without taking down habitat. We work with industry on this. It's not something that we're trying to fight in any way. We're partnered with them. They don't want to cut down trees and kill animals. People need to realize the lumber industry cares about the animals in the forest.

And so they really look to us to help them with this. And so we wrote an algorithm. I wrote this algorithm with the Forest Service where we can find all the animals in the forest using sound. And so, you can't find all of them. No, it's just true. The one I published this, the convolutional neural net I published this spring does 135 different species. Wow. It is not every one of them, but it is a lot of them.

And we can even pick up when planes are flying over, when it's raining. So some of the other conditions, the odd one is when a human walks through the forest. forest, the rest of the animals go quiet. It's a human. In Ohio, we've had reports of Squashwatch. Have you found that? I'm not. It's only because we find it. We may. It would definitely be in the

Pacific Northwest. I think that there are rare species, and I'm going to need a little bit more training data to pull them out. That's the only problem? That's my only problem. Yeah. We have about 4,000 recorders between San Francisco and Canada currently. We call them Autonomous

recording units or ARUs. And those are actively recording the forest to where I can pick out those species. I can even tell you which direction they're heading using Doppler shift. And I can even come down to male and female, you know, in terms of, you know, things like that. This is important because what we're trying to find in this

case is mating pairs so that we can identify that they're mating and that they're actually doing that. And we have statistical models to show when that's true. So... Well, that's on land. In the sea, though, you were telling me recently about seals and trying to get them. Was it seals trying to get them back up north?

No. Was it otters? It was something that they're trying to get back up north. Yeah, so I'm monitoring all of those. I have the crabs, the seals, the mud, and the whales. Okay. So, yes, all of those. We need to realize that the temperatures and the waters are changing,

and it changes the way life works in the ocean. And to help us monitor that, I've helped write an algorithm to monitor plankton. Right. Plankton is one of the most important things because all of those species you talked about are actually leaning on plankton for food and all of those pieces. In the end, the food web is built partly out of plankton. It produces 17% of the food that we eat per capita. And it produces 50% of the oxygen you're breathing and 25% of the carbon scene.

It's one of those crucial species on this planet. And to do that study with the plankton, you guys are launching several vessels that are going to be at sea, not just on reconnaissance for data, but also processing data at sea. And this is somewhat novel. Yeah, so we have a grant from NSF, National Science Foundation, to build three regional class research vessels. The first one comes out to us this December. And so the last year, I've been working to try and identify what technology we really need to put onto that ship and be able to do what I call semi-real-time processing.

It's important to realize that ship is $150 million, and for me to take it out into the ocean for 10 days, cost me a million dollars between, you know, hardware, labor, and all these things. And there's about 15 experiments happening at any one time. My plankton is only one of those experiments. But that one experiment for the plankton generates 100 terra of captured data in five days.

So you have what we're talking about these days around density of storage, the most confined data center on that ship. You want to process in near real times that when you get back to shore, you've done a lot of that legwork and that the other scientists can pick up and go. That's absolutely true, but we also want to do it while we're looking in the ocean because we don't know where the plankton are. And if I can't get actual plankton on the camera, that's a problem when I get back.

There's no data. Airpoint. You wasted that whole cruise. Right. We would have wasted that cruise. And that has happened where we did not get good data through the transects that we were going through. And I think that that's the most valuable piece here is that we are now dynamically moving the ship around to be able to ensure that we're able to quantify

what plankton are there. Plankton don't swim. They go by currents and stuff like. So we have to kind of guess where they're going to be and then try and sample through that space. So how big is your data center on each one of these three ships? Yeah, I don't know the physical space. Or how many racks? Like what sort of math can we put behind it? I'm not, I'm not entirely sure about that. I've been focused in on the amount of power. Okay. So power is another constraint. What sort of, what are your power challenges on the ship?

Yeah. So we're not going to be able to light up some of the biggest beautiful new pieces of hardware. We're always going to be constrained. And we're not going to have massive power. So we're going to be in the, you know, 30 to 50 kilowatts, stuff like this. And so that's enough for us to run several racks. But as I started to look and test some of this hardware in the last year, like we test some Dell hardware and stuff like this, the XE 7745 with four RTX6, 6,000s in it, we realize very quickly

it's power hungry. And those RTX 6,000s, they do enjoy a lot of power also. And so as I put that piece of equipment in, it does reduce down the other pieces I can do in that data center. Right. So you're not so much physically constrained by rack units. You're really more constrained by power to the ship. Correct.

Or power to the data center. And cooling out at sea is difficult. Right. And so we need to recognize that. So when we realized that we were going to need that GPU power, we instantly transitioned from what we were looking at in some storage capacity to how could I save energy on storage? And this is where we really focused in on what we call large capacity SSDs. There were several groups out there looking and doing large capacity SSDs. And we found the Solidine, you know, D5P5336, 122 terabyte drives worked really good.

but for that actual deployment, we got better performance on their 30-terabyte drives. And we really wanted to see that performance piece because if I'm doing real time trying to move a ship around, I really needed to know stuff fast. And so with the Plankton one, we were actually getting reeds and the 55 gigabytes a second, and we were getting rights in the 15 gigabytes a second,

which was beautiful. And those were the standard U3, you two. You choose. You choose. And with that performance, we were able to do six to eight streams at a time. And I think that the limit really became our disk, because obviously I could have added more GPU. Right. But the disc became, and I only did this with a limited set of disk.

Remember, this is a test that I have to prove to NSF before they give me the money to go and buy the equipment. Well, QLC, though, does give you that nice blend of capacity and especially capacity per watt, because that's for you. what that comes down to. And the more watts you can free up from storage and other parts of the system, the more you can give back to the GPU. That was exactly what we were trying to do. And I think that people need to realize that there is money savings in replacing old spinning disks and things like this and the way we're doing that. Yeah, we recognize that that was a huge savings to us. It also allowed us to land the data once. We were able to put so much space on this

one machine that I didn't need to move the data way. Right. This was a big deal because before we would have disc arrays, we would have all these other pieces surrounding the GPUs. That's all gone. I don't need to have any more than a single piece, and it is able to serve that. Well, and for your researchers, too, it brings simplicity to their workflow, where they can be streaming that data to the storage and then not have to worry about external devices, connecting them, copying it. You only get one shot.

That's one interesting thing about what you do in science. You don't get a lot of retries. No. we were working with a PBS production that just went down to Brazil and we sent them with a bunch of QLC to Brazil in these little raid boxes and talking to the team like we're trying to film rare birds flying by and we get one time the bird flies by and if we can't get that to protect the storage and something that we can haul around the jungle it's the same thing that you're talking about Yeah.

If you go to your ranger stations or the ship or anywhere else, if you can't protect that data quickly, you may not get another shot. Yeah, like I said, when we go out in the forest, all the animals go quiet. We're not going to be, and they're going to hide. That camouflage, you're not going to find them, okay? We have to run large models out at the edge to even find these things, okay? And so, yeah, you need to realize that if I can't sit there and sample without people being around and let it be in nature and catch those rare instances, we're not doing the science that people need us to do. Right. And so, yeah, when we talk about the forest one, we did a completely different deployment

where we worked with super micro on a machine that had liquid-cooled GPUs. Because now I'm in a ranger station. I don't have a data center. I've got to be able to put both processing and storage out into a forest in a ranger station on a 515R plug. I mean, we don't even have a lot of electricity. It doesn't have cooling at all. No AC in those.

No, AC. Oh, God, no. And so in the end, that lives. liquid-cooled A-100 system had U-2 drives locations on it. So again, we went for those solidime P5-336 and stuff like this, and they were the 122s. Because out there, we really wanted a longer iteration of time to be able to sample and to be able to gather that data. That one's basically bringing 250 terra to me every month and a half to two months, and we have to process through that. And again, I was able to put 300 terra onto that workstation, land the data at one place,

work all the data in one place, never have to go anywhere. Well, that's the thing, too, at the edge, it's the operational benefits, whether it's a factory floor or remote site or telco or anything, anywhere where this edge AI is getting pushed further and further out, where you don't have IT staff there, or you don't have maybe anybody there at all, if it's on a telephone pole somewhere. Right. And so being able to operate longer without human, intervention is a pretty big deal, I would think.

It's crucial to our ability to do science. Okay, I mean, like you said, the birds aren't just sitting there talking all the time. And I have to catch them when they talk. Believe it or not, the female spotted owl doesn't like to talk when they're mated with a male. I mean, it's just get real quiet? Yeah, they just become very quiet.

And so it's like, I heard a male. Is that a mating pair? Is that, you know, and so I have to start just, you know, figuring these things out. And we have to use a lot of statistical models. And all data, we have to hold all that data from all that time so we can now take the new data and still make it relevant. Okay, I have to have work that people actually physically did to go on sample and then prove

that that model and those statistical models still work. So we talk about this in the enterprise all the time, and there's an interesting sort of connection to what you do there, is that enterprises never want to give up data because they never know when it's going to be valuable. And AI in the last year and a half, two years has changed that whole spectrum so that all that transactional data that you've ever had, all the sales data, all the business data, all the IoT edge data, now that I've got new models. And instead of just looking at what's happened in the last six months, I can look back to

the last 60 years conceivably and run analysis against that and find all sorts of ways to improve processes, deliver better, whatever. You're doing the same sort of thing with scientific data that you will never be able to recapture from 1984 or whatever year it is. You can't go back and do that. And if that data ever gets lost or corrupted or destroyed, then it's just gone forever. And for what you're trying to do, which has improved the world through science, that's a big deal. That's a problem for future science. As scientists, it's our responsibility to make sure that the work we do is accessible to the next generation. That's why we publish. That's why they're citations. And that's why we push these data

is in the National Archives. It's important that this is done so that way future science can be done. We need to realize that we are constantly retraining models constantly to have them become more and more accurate. Models, and right now we're working towards Transformers and stuff like this. It's important so that we can do

multiple things at a time. And that will change the scope, remove bias, and create better science for everyone. In the end, the world we live in has been built on science. The fact that you have a phone in your pocket and all of these things was built around science. And science helps us drive technology.

So we've talked about this before, you and I, about the ultimate long-form media. Yeah. And you said stone tablets was the best way to pass data forward. I had conversations with people now saying, how would you send data 50 years into the future? And they look at me with no idea. And I tell them it's stone tablet or tape. Because-tape, that's the other one. That's viable then.

Yeah. So tape, we have to realize, I see. still have ship data coming off of tapes from the past, and we are still able to bring it back and use that ship data for modern analysis and compare it to what's happening. We are trying that, like the Ocean Observatory Initiative, which is a grant to Oregon State University where we have a massive $10 million infrastructure to support that, that's a 25-year project. I have to hold that data for 25 years minimum. It'll be held for 50 to 100. And that data is one-minute level interval data coming from the ocean from all these wire walkers and gliders and things.

And when you saw the information about the earthquake that UW put out last year about the underwater earthquake, that was the data coming from Oregon State's data center. We support UW, Rutgers, Woods Hole Oceanographic, with that exact same data that we're the ones running it. We're just really good at running tech. That's what we do. That's the Pacific Northwest for you. Yeah, I mean, that's why we have famous people who came from Oregon State.

Speaking of famous people that came from Oregon State, in addition to all these edge initiatives, which are really cool. And I know you're doing a bunch of stuff we just wrote about our concept on enabling students to have access to sparks and high-end local AI systems. Tell us a little bit about what's going on with Wong Complex and where that sits, because that's your big data center initiative. Yeah, I mean, we love Jensen Wong. I mean, he comes from Morgan State.

He and Lori both have degrees from us. I think he's the greatest electrical engineer who's ever existed. And, you know, we work with them to try and figure out ways to fix the planet. He's given us a building and a data center in that building and some NVL 72, some Vera Rubin, that we're going to be using to try and help us understand how the planet is changing and how we can live on this planet in a really positive way. And that's what we use data centers for. And I want people to understand we're not going to fix this planet without data centers.

Okay. And we use Amazon's data centers. and I'm using Azure and I'm using Google Cloud right now. I mean, all of my work with NOAA is all in Google Cloud. So when I hear people complaining about data centers, I think, yes, there's always going to be problems with any technology, but it's our responsibility as society to communicate those things so they can be adjusted. Not to lose these items. Data centers are crucial for us to be able to change the world in a positive way, especially around science.

All right. So you've always said in the past that AI is a data. data problem. As we sit here today and you look at what's going on at this conference and all the energy here, what's your forward-looking take for the next six to 12 months on the next challenge for data centers and AI in general? So I think data, I strongly believe data centers don't need to be gigantic. I think when we start bringing in liquid cooling, we can put a lot of capabilities into smaller spaces and distribute a lot better, creating a lot less impact to the planet. And I think that that's one of the core pieces. I do constantly tell people that

AI is a data problem. And it's kind of an arms race. As the processing gets faster, we add more data. There's never been a moment. There's no stopping the data. There's never been a moment I didn't have more data to throw at those processors. I always have more. And so that's why I say it's a data problem. When I look at the AI is the context of your kitchen. Okay, like your fork, knife, spoon, saute pan, those are your processors, right? But the, the, the, we're refrigerator holds everything you care about. The kitchen's dysfunctional without the refrigerator. Okay? And the food is the thing we're talking about. The food is the data. And I laugh because nobody can tell me what fork, knife and spoon they have, but they can always tell me the manufacturer

of the refrigerator. And that tells me something. It tells me, yes, we care about the food in the kitchen, and we care about the data when it comes to AI. Data informs and is used for processing. I don't need the processing without the data. It's important to realize that. Yeah, and as we look at this audience of where we're going and all the progress, the storage doesn't stop. And for us, as storage guys, it's been really fun to watch the rotation come back into storage as critical infrastructure to every business initiative that's going on out there. Except for the price. Well, I can't do anything about that. I know. But what we can do is get the higher densities, lower the rack footprint, the power consumption, per terabyte, and do a lot of other efficiency

stuff that feeds back to... That's where the savings is. People need to realize that we're back on that little curve going back up because of the fact that demand is our reaching, you know, the ability for supply. Okay. But when science is done, we actually write grants. And when we get those grants, it's usually months after we've submitted the quotes for everything we're trying to do. And so we're watching our grants not be able to fund the technologies that we said we needed at the level we wanted because the change is happening too rapidly. Normally a change occurs, but it happens. happens in an 18-month cycle. Right now, we're in a three to six. It's more like three. Yeah, it's wild, right now. It's wild. And so, we're hoping that calms down because it affects our ability to keep gathering and move forward with data. I am not going to stop gathering data. I just need better ways to do it, and you're right, when I move to the 122-tabite drives, and I have eight of them producing a petabyte, and I get rid of an entire rack of disc, that's where the savings is, and we have to equate everything. Even though I spent a lot to get those

discs, I saved on hardware to put those distance. The thing people need to realize is that we spend a lot of time in science moving data around. So when you talked about the ship, all that data lands, I've got to get it off the ship. I can't just go there and take the machine off the ship. And so I had to run a 100 gig line down the dock, put 100 gig switch at the end of the dock,

and I have to make it so I can plug the switch in 100 gig so we can transfer. And so people need to realize that we want to try and minimize that. Large capacity SSDs help us minimize the moving of data around so that way I don't have to spend my time just doing that and I'm focused in on actually using the data for meaningful things.

Chris, this has been great. As always, we'll get back to the conference and see this podcast studio to the next group, but no one's doing it quite like Oregon State when it comes to edge to data center and cloud and always impressed with the work you're doing.

Thanks for sitting down. Thank you. Thank you.