Seagate: If you're adding platters, you've hit a density wall

Jason Feist, Seagate SVP for products and markets, says hard disk drive suppliers adding platters are compensating for not being able to increase areal density.

We contacted Seagate to find out its views on Western Digital's  ecent announcements, including higher platter counts, High Bandwidth Drive (HBD) technology enabling simultaneous reads and writes across multiple platters, and dual-pivot designs that further boost I/O bandwidth. Jason briefed us in an interview call.

Blocks & Files: How does Seagate generally view the WD announcements?

Jason Feist: We're most excited right now about the fact that all of our customers are amplifying and reinforcing that data is valuable, and that storage is a critical component to their architectures. So I think that opens the door for all of the component providers, DRAM and flash and hard drives, to now innovate against that, which is fantastic. So we each have our strengths and we each have our core competencies. And the beauty now is that the demand far outpaces the supply. And so there will be ample room for all of us to operate, which is great. And all of us need to execute swiftly to ensure we can deliver to our customers' needs on supply, which is great. Seagate has chosen a path of areal density growth, which has been very well received.

Our customers qualified Mozaic 3 (3 TB/platter HAMR technology) and now are in the process of qualifying  Mozaic 4 (4TB/platter HAMR tech) already. Those quals are accelerating and moving faster. And as you've heard in our quarterly earnings, all the cloud accounts are basically done. We're moving forward now. 

And the beauty of moving into Mozaic 4 is we also open that technology up to all the other routes to market. So now, from a manufacturing scale point of view, we have a design that we can make large mass capacity drives for our hyperscaler customers inclusive of both CMR and SMR (Shingled Magnetic Recording), and we can take that same technology and use it on lower platter count drives for other use cases, NAS and VIA and nearline enterprise very, very efficiently. That is the one big, big advantage that we believe differentiates and keeps the strategies between the different HDD vendors separated. Adding platters to a drive to increase capacity doesn't also help those other use cases.

Because if you're a 16 TB drive or a 24 TB drive or an 8 TB drive, more platters is not what you need. You need ongoing supply, and you need ongoing cost optimizations, and you need ongoing manufacturing efficiency, which is why areal density in the hard drive industry has always been what we do. And when we don't have areal density, we add platters. In that sense, we're really happy. Customers are happy with the product scaling in production with millions of Mozaic drives already.

Blocks & Files: How about the performance aspect?

Jason Feist: This one's a little different. I'm pretty excited about this one in the sense that Seagate already knows how to do this. We've talked to you in the past. We've shipped millions of dual actuator drives [MACH.2] and we've shipped over 50 exabytes of storage on performance hard drives. We know how customers use it and we know what they did and we know why they came back to us and said, keep going with Mozaic, keep building mass capacity drives. 

The main feedback that they've all given us is that they're optimizing their performance with intelligent data tiering and software within the cloud architecture. They flow the data through small amounts of flash, and then they ask us to optimise the hard drive and focus on capacity, so that the AI workload can saturate the GPU, and then ultimately keep the data for long-term retention for reasoning and logic and everything else.

So we've already gone through that exercise. We learned a ton by going to manufacturing and building those millions of drives at high performance. And we realized they don't need them now and they won't for the foreseeable future.

For the JBODs that they've designed, the systems that they've architected, a single actuator hard drive is sufficient. It meets their needs and scales for them quite well all the way out beyond 50 TB. So that's our focus. 

Between now and then, our focus is to get them up that capacity ladder as fast as we can. And if and when they need something other than that, we already know how to do it. The design is done. We've already manufactured it. We can just pull it back and integrate it. We don't have to retalk about it and republish it because we've already done it.

Blocks & Files: The MACH.2 drives, as I understand it, are a single pivot drive, but logically split into two separate drives with two halves of the read heads on the pivot operating independently. Will that provide as much of a bandwidth increase as Western Digital's dual pivot drives?

Jason Feist: It did. When we delivered that drive to our customers, we had multiple customers in production. You could get 2X performance growth. The whole path of parallelism, both in the announcement from our competitors, and what we had previously put in production, the core architecture says the host, the hyperscaler, is aware that they have access to two heads at the same time.

So you take the bandwidth of one head and the bandwidth of another head, whether they're on opposite ends of the disc or whether they're on opposite discs, gives you the same outcome, 2X bandwidth. The reality is customers then need to do a little bit of host-level work to rearchitect their software to get that bandwidth, and they did some of that work with us in the past. 

Now, the gating factor to future performance gains is the interface. So all the drives in a cloud architecture are SATA. If you go pull all the statistics of all the shipments, of all the nearline drives, predominantly what has been deployed into these facilities is SATA. So [with] anything more than two heads per surface or two heads operating at the same time, the SATA interface becomes the bottleneck.

And cloud customers today work very closely with the host bus adapters, the companies like Broadcom and Microchip to optimise both the architecture and the cost of the SATA architecture for that connectivity. And some of them have even gone so far as to look at other custom silicon to try to do that. And they still go back and use these cost-optimized architectures from Broadcom and Microchip that are very stable, very mature, and have been deployed for years and years and years. So that really becomes the bottleneck.

Blocks & Files: Would you have the option of moving up to a faster interface if it was required?

Jason Feist: Of course. We actually at GTC last year demonstrated an NVMe hard drive. Seagate has been leading the ecosystem development for that. We felt many years ago that it was important for the software developers in cloud as they develop features for NVMe SSDs, would it be easier for them to unify the software stack and have both NVMe SSDs and NVMe hard drives together? 

Because it was very obvious when we had that conversation, that the effort to redesign and re-architect and rewrite the entire software stack that they've built for years and years and years on top of SATA, that the additional small amount of performance gain that you would get from going to NVMe in a hard drive did not pay back for the massive amount of work of rewriting that software stack. 

So therefore they continue to manage those two things as independent tiers, and then they use their ability to orchestrate data and move data around to manage the balance of performance and storage together.

Blocks & Files: Would it be fair for me to walk away from this thinking Western Digital has some glamorous ideas such as up to 14 platters, such as dual pivot drives, and high bandwidth drives delivering a combined 4X possibly more increase in bandwidth. They're glamorous, but they're unrealistic in your view because they don't reflect the hyperscalers' practices of already tiering between Flash and disc, and already needing to remain with the SATA drive interface, because going to NVME would not give them the performance gain that they need.

Jason Feist: I'm almost a hundred percent aligned with what you said. Two subtle modifications, Chris. Capacity per drive is important to hyperscalers. And so I think this is one where Seagate has proven we can get to those capacities through areal density and our competitor is lagging in areal density, and so therefore they're adding platters. 

Both companies have to innovate to provide supply in this environment where demand is very high. So it's not bad one way or the other. It's just a cost, a manufacturing scale and a capability. If you have the capability to drive areal density, you'll do that. Rewind the last 20 years of all three companies; Toshiba, Seagate, and WD. We all drove areal density until we didn't have it, and then we added platters. 

And so that's OK. This is where I think it shouldn't be thought of as an us-versus-them or and-or strategy. The value of data is super important. These hyperscalers are growing faster than the hard drive industry has ever seen before. We need an and. The world is going to need Mozaic technology and the world is going to need WD and Toshiba with more platters to be able to build drives. That's OK. And so we don't want to say one is bad and one is good.

They're both OK. Now, we know that the industry always would take areal density if it has it. And so this is where I think our commitment and our investments and everything else give us a strategic advantage.

Our customers are accelerating and qualifying those drives very quickly, and they are aligned to our roadmap, which we're really excited about. So in that sense, I think it's an and. All the hard drive suppliers are needed moving forward. And I think you're going to see what's going to separate us is our ability to execute and our ability to continue to be predictable. This is where Seagate, I think, is super focused on the integrity of our message. We say what we're going to do and we do it. No change. We've said 20 TB in 2020. We've said three TBs per disc [platter]. We've said four TB per disc [platter]. We've said five TBs per disc [platter]. We just do it so the customers know what they're going to get and they can plan their business and we're there for them.

Blocks & Files: You learn from your hyperscaler customers?

Jason Feist: We build what our customers want and ask for, and then we test it with them. And that's, I think, the story we learned with the MACH.2 dual actuator before. They asked for something, we built it quickly and we scaled it quickly, and then we tested it with them, and we learned why that didn't become the predominant storage in cloud, and it stayed with single actuator mass capacity drives.

The other thing to watch out for too is in times of supply demand across all the components being so challenged, there are dynamics in the balance of flash and disc and DRAM and how performance is managed. And so you've got to know where those things are headed long-term and you don't want to design something that ultimately is only a short window and will likely equilibrate in the future. And that's why we're so carefully planning our business with our customers and making sure we know what you need today, we know what you need for the next three years, and therefore this is what we're going to work on and keep all of our resources focused on that.

And as in any industry, when supply and demand are imbalanced, the most important thing is to have those deep and honest conversations with one another [suppliers and customers] to make sure they get what they really need. And that's where we're focused and prioritizing.