Season 2, Episode 4 – Interview with Yagnesh Ashara, Director of Enterprise Solutions at Super Micro

Steve Sidwell:

On this episode, we interview Yagnesh Ashara, the Director of Enterprise Solutions at Supermicro. During our interview, Yagnesh discusses how the company has grown over the years. How Supermicro is staying innovative in this space, GPUs, and so much more. Brace yourself. You’re now entering The Tech Bench Podcast. Seriously, this is one of the best we’ve done, so you should really listen to it.

(theme music plays).

Steve Sidwell:

We’re here today with Yagnesh Ashara. He’s the Director of Enterprise Solutions at Supermicro. Super excited to be with you today. I’m sure that, basically everybody on the planet knows who Supermicro is, but for anybody who doesn’t, can you tell me a little bit more about your company? I know you guys make servers and all stuff, but tell me more. What is Supermicro? What do you guys do?

Yagnesh Ashara:

 Thank you for having me here, Steve. Everybody here, it’s a pleasure to be here. At Supermicro, we were founded by Charles Liang. We are a Silicon Valley startup company founded in 1993. Ever since the inception, we have seen growth, since ’93. Quarter over quarter. In phase one of our growth, we were basically a motherboard design company where we were designing dual socket Intel motherboards, and our market was mainly the OEM channel market where we sold a lot of this motherboards. In phase two of the growth, what we have seen is a lot of customers were interested in a more complete disk servers from Supermicro, so we started designing our own Chessies, our power supplies, fans, and other components that go along with the servers.

Phase three came along in around 2013, 14 timeframe, where now customers are more interested in having a complete solution from Supermicro, including the software, the firmware, and all other things that go around and in the server. Supermicro started working with a lot of channel partners, a lot of software providers, vendors, and we created our own solutions based, mostly industry standard open technology, or open source providers. We have participated in also remote management consortium, and things of that like. Supermicro, we are, again, it’s a Silicon Valley company. We have operations in three different parts of the globe. The headquarters is Silicon Valley. We have a facility in Netherland. We have facility in Taiwan, and logistics are basically for Asia Pacific. We ship everything from Taiwan, for Europe and everything from Netherland and for US, North America, South America, everything ships from San Jose, the facility.

Steve Sidwell:

Yagnesh, now that you’ve gone from selling motherboards to OEMs, developing your own servers, and now your server software solutions. You’re obviously in the market space with HP, Dell, Cisco. I was just curious, what are your biggest tip for your interiors compared to those larger competitors?

Yagnesh Ashara:

There are few, and I’ll start with one of the biggest one, Supermicro Innovates. You have been to one of those conferences where, to show our customers, our partners, Supermicro is always on the leading bleeding edge of the technology. When we develop, people are still thinking about it and then we have the products ready. We are at the forefront of the technology. Secondly, that’s a first to market or time to market or whatever you call. We have a breadth of a product. We have probably more than a thousand active skews, and what it does for, for customers, it gives them something that’s readily available. They don’t have to settle for small, medium, large. They can actually get extra small, small small medium, medium large, large extra large, extra extra large.

They don’t have to really be told what they are getting, they can choose and pick. On top of that, we provide more customization flexibility, for example, from where, logo changes, bios setting changes. Some of those things, they’re small, but when you do this at a scale. Let’s say you have 10,000 different servers, and you want your servers to be look likes X, Y, and Z from the factory. We are able to do that. We are able to customize that from factory, and it makes a lot of difference when we are talking about skills. That’s another thing. Mostly we worked with the industry standards, open standards, like I mentioned earlier. IPMI 2.0, Redfish, Plaza, the SSI standard. It’s all compatible with what’s available in the market. You don’t have to worry about plugging into something, and that part doesn’t work. That’s how we do it.

Steve Sidwell:

You were just talking about, how Supermicro is really skilled at being right out there, right in the front of technology, right on the leading edge. I know that, like Intel has just released the second generation scalable processors. Two things, one, can you tell us more about those chips, and then also, what are some of the performance improvements of the servers that you guys have just released that are compatible with those?

Yagnesh Ashara:

The second generation of scalable processor codenamed cascade Lake processor, and we have released those with Intel. Some of the listeners who are not aware of, actually we provide a general availability before intel launches their products. We call that early access. This is in partnership with Intel, and in some cases also there are components that go along with it. We also work with other component vendors as well.

 Coming back to your point, the cascade Lake processors in general, they provide more core counts or frequency compared to the previous generation. For the same amount of money that you were spending before, now you are getting either more core counts, or you are having getting a higher frequency for the particular CPUs. That’s one of the key advantage. Another advantage that we have seen, is on the memory frequency. It used to be 26 66 megahertz, and now with the 29 33 megahertz memory support.

The application which are memory bound, they get tremendous, it’s about 10% increase in your bandwidth. When you talk about the application that can really use it, they get tremendous bandwidth out of it.

Steve Sidwell:

That’s fantastic.

Increased Server Capacity  

Yagnesh Ashara:

When we add one more variable in the equation, which is Apache pass DIMMS, or what Intel calls a DC persistent memory modules. This memory modules are a flash that runs at a memory speed. You are now running flash, but you are running rather on the memory bus. Any application now, let’s say you have an in memory database that you are running before, and you want it to, probably you had to click create a cluster of, let’s say, 10 different servers. Those days are going away, because now you are able to put up to 4.5 terabyte of memory. It’s in terms of flash, but on the memory bus, in single server. All those limitations of talking between the different servers, they are going away.

Steve Sidwell:

That’s all operating a bus speed?

Yagnesh Ashara:

It is operating at a memory bus speed. DC persistent memory can operate in two different modes. It operates in a memory mode, or also in the storage mode. For somebody who wants to use those memories as a block storage, there is another application which can utilize that and use it as a block storage.

Steve Sidwell:

DC persistent meaning basically, just to make sure that I understand right, is that as long as it’s getting in the same way as regular memory, as regular Ram, it’s persistent, as long as it gets power. Is that right?

Yagnesh Ashara:

It’s persistent even after you lost the power. It’s a nonvolatile memory, in fact.

Steve Sidwell:

That’s fantastic because that just changes so many things, where you were going from having racks of equipment. You’re now down to half a rack of equipment, or a couple of servers, if you want to multi-serve redundancy or whatever, just to be able to access the data that much more quickly, and without relying on going across a bus between multiple servers.

Yagnesh Ashara:

Exactly

Steve Sidwell:

Whereas before if you needed like InfiniBand or something like that, to get speedy access to that data, now you have it right in.

Yagnesh Ashara:

[crosstalk 00:11:01]RDMA or something before, and now you don’t have to rely on those yeah. Up to certain limit for sure. Because then if your data set is still larger than that, you still have to think about what it’s going on.

Steve Sidwell:

Oh, sure. For any kind of midsize data or data, it’s four terabytes. It’s still nothing, but it’s a tremendous amount even just for caching.

Yagnesh Ashara:

Exactly.

Steve Sidwell:

Who benefits most from this increased speed and capacity? Is it a specific vertical, or is it specific people within an organization?

Yagnesh Ashara:

I would say database applications benefit the most out of the DC persistent memory. At Supermicro, we used a DC persistent memory and we have come up with our own innovation. We have now designed a server that can take up to 12 of this DC persistent memory in a server. On top of that, I’m not sure if you guys are familiar with the PCI express storage or NBME?

Steve Sidwell:

sure.

Yagnesh Ashara:

NBME storage is a faster than Sarah, and when you compare the price, you are not paying premium anymore. It’s pretty much in the same line as the SETA pricing. Our new servers, what we call ruler servers. Intel came up with a ruler technology about 2016, 17, and we have implemented. Again, we were the first to put that server out in the market. Now there is a generation, two of the same servers. We have the servers that can, in a one new form factor, it can accommodate up to 32 or 36 MVME drives. Each driver can carry, anywhere from four terabytes all the way up to 32 terabyte of data. Well right now, 16 terabyte is already in production and 32 terabyte is in discussion or it’s coming out, probably by end of this year, early next year. When you add all this together, having a lot of persistent data close to your CPU, having NBME storage running at a BCI express speed, you can use it for almost any database application, any low-latency application that you can think of. Some of our financial customers are, their media customers are using the NBME servers for the same.

Steve Sidwell:

Sure. If you know enough how to program a big database, you know enough how to program the database, so it can be at least a little bit intelligent about what it’s feeding into its caching areas or that kind of thing, so that whatever you’re pulling, you’re pretty much pulling from the fast parts.

Yagnesh Ashara:

Exactly.

Steve Sidwell:

I’m curious, obviously the gaming industry is becoming a huge industry. I’m sure that they’re utilizing your service for that, all the streaming that’s out there. Is that changing your business model at all, or how you put out product?

Yagnesh Ashara:

It works well with our model because, like I mentioned earlier, we have a breadth of a product. We have gaming desktops. We have gaming workstations. At the same time, we have GPU servers, which goes in the cloud and then customers are using those for, the streaming GPU desktops as well.

We have different products that fit well with almost any business demand or any needs for gaming. Recently, if you follow our followed our CES 2019 PR, we launched a motherboard, C9Z390 GPW. With liquid cooling, we were able to overclock it up to 7.5 gigahertz. It’s stable at that frequency. We have different products, and for different markets, we have a different variety of this product that can satisfy those needs. The components that we use are server grade components. It’s not desktop great components. This is where the reliability of the product comes in. Your gamers are able to utilize it for longer periods of time. Typically that’s not the case. As soon as they see something new, they would want to go for it.

Steve Sidwell:

But at the same time, they’re willing to buy the good stuff because, forbid it going down when they wanted to. It’s always worth it to have that extra quality.

Customized Server Solutions

Speaker 3:

You’ve mentioned, obviously your products and services come in a variety of flavors, and I’m sure you work with many clients globally. Do clients come to you and have you tailor solutions directly to them? If so, do you have a success story that you could share with us?

Yagnesh Ashara:

Sure. We have both the models where customers want something off the shelf, and they choose it, pick it and off you go. We have other clients who want customization, and we do offer those customizations. Whether it’s a logo, basil change, color change, different things. Different customer, different requirements. One of the nice success story that I would like to mention is about, Rutgers university. That’s right here. New Jersey. There was an RFP to build a large supercomputer that satisfies about 500 teraflops of computational power. We were invited as a part of the RFP process, and we worked through the process, and with our twin servers. Now, let me give you a little more introduction about twin servers. I mentioned about innovation. In 2006 and seven, around that timeframe, Supermicro co-designed with Intel, the twin servers. It’s basically like the name indicates. It’s a single chassis, but in this chassis you have two servers. Two fully functional dual socket servers.

We came up with a twin architecture in 2006, 2007, and over the years, we have modified it to be more useful for different applications. We have a product called big twin. We have a product called twin pro. It’s a two yule, four nodes where we have redundant power supply. Then we came up with a fat one, it’s a four U eight nodes where we have redundant power supplies. At the same time, we have front hot swap nodes for easy maintenance. With all the different products, now they are basically optimized for HPC. Any customer like Rutgers or, who wants a highly dense compute. The servers are very useful. We used a fat print servers, and we were able to achieve the density that nobody else had at that time with a modular data center. They had a requirement of putting all the servers in a modular data center. We had to provide a schooling solution as well. Rather than using a standard air cooling, which is traditionally, everybody else uses, we have used liquid cooled doors. We put it behind the servers, which absorb all the heat that’s generated with this servers, and it absorbs, and then it cools the water and then it goes back into the system again.

Speaker 3:

That’s awesome.

Yagnesh Ashara:

Yeah. The density that we had per rack, was 56 servers in a 42 year rack. Typically you put a venue, 42 servers basically. That fills up a whole rack. But now with the twin products, we were able to put 56 servers on top of that. We had three switches also only the path fabric switch. We had a 10 gig switch for management, so on and so forth.

Steve Sidwell:

That’s really cool.

Speaker 3:

There’s been lots of news recently about AMD, really trying harder to get more and more into the data center space and gain market share over Intel. Do you see clients ask you more about AMD, or are people buying more AMD equipment?

Yagnesh Ashara:

AMD is coming in market with the second generation of their APIC processor, codenamed Rome. It’s around mid to end of 2019. What we have seen, is there is a lot of ask for the roam processor from the customers. With the 64 course of compute power, 128 PCI lanes, even with hyper-threading, you can have 128 logical course. It’s going to be on seven nanometer technology architectures. With all this, yes, they have generated lots and lots of traffic within the industry. There is definitely a lot more asked about AMD than it was before. Just like with Intel, we are also close partners with AMD as well. We will have AMD servers available. Well, the first generation is already available right now, and the second generation servers are available right at the launch.

Steve Sidwell:

That’s awesome. It sounds like it really works great because, for you guys, you get to utilize the technology right, when it’s coming out through these partnerships. But at the same time for those companies, for the chip manufacturers, they’re able to get their products into the marketplace through a single, scalable solution where they know they’re working with a vendor, and they get a lot of stuff out there, and get feedback on it, I would imagine. It sounds like a pretty good symbiotic relationship between Intel and AMD Supermicro.

Yagnesh Ashara:

Definitely. Earlier I mentioned about early access program. It’s kind of a pilot program that a customer has early access to the technology. Same program is going to be run on AMD as well. Some customers will have early access to the AMD CPU and servers. They will be running it on their apps, and see what the performance improvements are. That way everybody is able to understand when the product is launched, what they are going to get, and be able to ready right at the launch to roll out a full production speed.

Industry Trends

Speaker 3:

You mentioned your clients coming to you and asking you about those AMD processors. Are there any other trends that you see in your industry based on customer requests, like more people moving to the MVME servers, MicroBlades, anything else in the industry?

Yagnesh Ashara:

Definitely. NBME server is the new storage trend. Well, I wouldn’t say it’s new anymore. We have been selling NBME based servers for the last three years or more. We were the first one to come up with a hot plug NBME, that word. We have been actually really successful in that market. If you look at some of the products, we have, up to 48 NBME servers, I’m sorry, 48 NBME driveways in a two year box. It just gives you tremendous density for your storage apps. We have some regular purpose as well, but in general, any Supermicro service that’s shipping now we’ll have one or the other form of NBME drives that it will support. Putting this with the blade architecture, or combining it with a blade architecture.

 When we talk about high performance computing, again, we are trying to utilize everything that’s out there and then putting it in a way that customer can use it. MicroBlade architecture is now also supporting the NBME, our regular MicroCloud architecture. Now, there is a difference between MicroCloud and MicroBlade, but eventually they are there to support customer who wants to have a physical isolation of the servers at the same time, easy maintenance, and they don’t want to deal with a lot of virtual machines. That’s where the MicroBlade architecture is coming in useful solves. Some of the hosting providers, if they are hosting servers for a different end customers, they would sometimes have a hard requirement of not having a virtual machine.

That’s when our, MicroCloud or MicroBlade architecture will be really useful. If you follow Intel’s a recent release, we actually are providing the MicroBlades. We are selling into Intel, and they help fill up their data centers with, more than I would say, 70,000 of this MicroBlades in other data centers. Their application is a little bit different but, it can work in different verticals.

Speaker 3:

If I’m say, excuse me, so let’s just say, I’m an Ad tech company, and I’m going to move to that type of architecture. What’s a typical timeline when people start planning to the full implementation of that? Is there a standard?

Yagnesh Ashara:

I don’t believe I understand your question.

Speaker 3:

If I’m going to move to say  start, if I’m a….

Yagnesh Ashara:

You want to evaluate the pro product?

Speaker 3:

Correct. If I’m an ad tech company and I want to start evaluating a product, and I may want to move to that product. What’s the typical timeline for that, to take years, months?

Yagnesh Ashara:

Typically, it can take anywhere from a two to three weeks to few months, depending on the client requirement. If the client requirements are basically having a certification done at, let’s say lower level, they want to certify the integration, how this servers integrate well into their operations. They can check for the bios IPMI. They can make sure that their tools are able to work with our tools. It’s a seamless process where if they need to, in future, operate the firmware, they need upgrade the bios. It can be done within a few seconds to a few minutes. Doing this kind of evaluation can take, probably around two to three weeks, all the way up to probably two or three months, depending on what’s your workflow like. We have seen both. We have customers who have done it in two weeks. Some customers even took three months.

Speaker 3:

It’s a lot quicker than I thought.

Steve:

That really is. I know that you’ve been with Supermicro for quite a while. Can you tell me, so I’ve been a liquid tech for a while, and the technology that we’ve seen come through here in that time has just changed phenomenally. But I would see it coming from the back end. Where I’m going with this is, what are the biggest changes you really feel like you’ve seen while working for Supermicro.

Yagnesh Ashara:

One of the biggest change that I noticed back in around 2006 2007 was, Intel brought the memory controller in the CPU. What that did, it opened up a lot of real estate for other components on the board, or it shrunk the overall components on the motherboard. That’s what helped us to enable products like twins, or big twins or fat twins. That’s a really cool change. Don’t forget AMD was doing it before, but Intel followed it, and with them a memory architecture, we were able to enable products like this. Products like make twin. Another product is, like you already know this, it’s a PCI express flash. That’s changing everything that’s happening right now. Traditionally we have been using the spindles for how many years, and now that’s changing.

Everything is coming at a speed of PCI express and flash. Speeds are going up. If you are following NBME over fabric, that is going to make some big differences as well in the storage market storage. Anywhere, whenever we are talking about a software defined storage, and going forward, NBME is going to definitely make a lot of difference there.

Steve Sidwell:

But certainly, it’s just, to me it’s revolutionized the way that you can think about writing applications, because you don’t have to worry so much about what that latency is going to be when it has to go and pull in more data from the disc, when it has to get rid of the data, and what it has to write back. The way that you think about these things is, a lot of the time is like, okay, so users doing this and we’re going to start pulling in data that we think is going to happen. That we think it’s going to be important. We’re going to be writing this stuff back to caching it back to 10 files or whatever. Then hopefully we can just execute to commands to turn that into real data later. Hopefully they won’t change it and we have to do it over again.

               But now, that stuff’s all going away.

Yagnesh Ashara:

Exactly.

Steve Sidwell:

When we started moving to higher level programming, away from just writing everything into assembly, now this is really changing. Now I can start looking at things in a way that is certainly not as economical, but I don’t need to be economical. There’s no point in it.

Yagnesh Ashara:

For the media companies, it gives them a caching layer that’s available at really good price point. When, when they do see, if you talk about CDNs, they have to have this available at really fast speed.

Steve Sidwell:

Absolutely.

Yagnesh Ashara:

They have it.

Steve Sidwell:

Prior to that, the only things that we would see coming through here are, they to have these huge disks that were just tremendous amounts of Ram with battery backup on a scuzzy bus. Which sounds like a terrible idea, but that’s all there is. That’s the only high-speed storage there was. But now all that stuff is just like, Oh yeaH that’s so unnecessary.

Remote Server Management

Yagnesh Ashara:

Its changing. Another thing which I’ve seen, it happened recently is a Redfish, introduction of Redfish. It’s the open standard for remote management of the servers. We are actually an active member of the consortium, and with other partners, and also competitors of ours. This is going to be changing how you manage your servers remotely, how you scale it up, how you flashed the firmwares, how you change the bios settings. So on so forth. It’s going to make it very easy. It doesn’t matter where your servers come from. Everybody will be complying to same set of commands for remote management. That is going to make the CIS admins, or SIS operations life really easy.

Steve Sidwell:

Does that include access to, so not only is it bios, but does that include access to the upper level of the lower level command set? I’m breaking away from the podcast, I don’t know what that would be called and Supermicro’s world, but it would be like ILO.

Yagnesh Ashara:

ILO.

Steve Sidwell:

Something like that.

Yagnesh Ashara:

Sure. Remote management. Yes, the BMC aspect of it, or baseboard management controller or IPMI, whatever you call it. It’s going to, there are commands to make changes into the ILO types, or the IPMI configuration of the board. It will help you to change your active directory setting, your time settings, or anything that’s there on your IPMI device. Yeah. Eventually, when the standard matures, they also are going to facilitate the flash of the rate controllers, or network controllers, so on, so forth. When the specification is mature enough, we will see really easy maintenance for servers, from any vendor. SIS admins are going to be benefited the most out of it.

Steve Sidwell:

It sounds to me like that’d be something that, as in the same way, as there are multiple levels of cloud. Where you can have one machine with a hypervisor, or multiple virtual machines running on it, or you can have just actual cloud services where you’re just buying a service and you don’t know whether it’s running on one machine or 40. You start to look at the actual hardware in the same way, where you can have an application that manages all these servers for you, and just keeps them running.

Speaker 3:

You nailed.

Steve Sidwell:

It’s just an automated data center solution.

Yagnesh Ashara:

Absolutely.

Future of the Data Center Industry

Steve Sidwell:

That is so cool. Where do you see the data center industry going? As I know that certainly,  a lot of the services that we buy, and companies that we work with are moving into cloud and whatever. But all of those companies need all that runs on your servers too. Where do you see the industry going with that as far as public cloud, but then also as far as the private cloud, we’re talking about earlier, with the MicroBlades and NBME and all that.

Yagnesh Ashara:

I don’t know if you are aware, but a soft layer, which is a public cloud from IBM. They are one of our biggest customers. It’s as a public information, it’s on our website. We sell a lot of those servers to public cloud providers. At the same time, we have customers who are, they are pharma company, or customers who have their data sensitive, and they don’t want to have their data out there yet, unless they’re more conservative. For them also, we sell into them, and they have their own big farms of the servers. For us, it’s a win-win situation where we are selling into all different providers, whether it’s private companies or public cloud provider companies. We work very closely with the software vendors to make sure, in many times actually they also come to us and they in fact start using our servers before we know that they are using our servers. Eventually the partnership forms, and before we know it, our servers are being shipped to a company, and they are being used in quantity of thousands to 10 thousands with open source technology that’s available from our partners. Yeah, it’s a interesting place to be in, and going forward., I’m sure there’ll be other changes that’s coming. We’ll have to see.

Steve Sidwell:

Yeah. It’s an awesome time to be alive, honestly. When it comes to this tech it’s, there’s just so much going on right now. The growth has just changed so much in the last 10 years, even five years, where things are just opening up as we’re moving away from spindles. Everything is getting a little bit cleaner, a little bit faster, a little bit more organized.

Yagnesh Ashara:

The storage demand, if you see, I’ve seen graphs there, and all the data that’s being generated in last couple of years, I think that equates to all the data that was generated right before the inception of the computers until that point. We have the cell phones, now handy. We have cameras. We take videos, photos. All this data that’s generated, it has to stay somewhere. Facebook’s of the world, Twitter, all this data goes somewhere, and it’s staying in the server.

Steve Sidwell:

As we know, it goes somewhere and then it stays there forever. Having that growth be, able to be exponential. That’s just the key right there because whereas we used to say like, a little double, a little double, but we’re not talking about doubling anymore. We’re talking about powers of 10. We’re talking about GPU servers before. I know that Supermicro has been working on this for quite a while now, and has believed some stuff. Tell me more about what you guys do with GPU servers.

Yagnesh Ashara:

Just like any other product that we do, we do it in multiples of different skews. We have a GPU servers that can take all the way from one GPU all the way up to 20 different GPGPUs in one server. Now with AI and machine learning, being key buzzwords for next few years, uh, people are, so if you take Uber and Lyft, they do a lot of predictive analysis. Say if you talk about financial companies here in New York, they do a lot of predictive analysis. What GPU servers do, they cut this simulation time, or prediction time for them by, it gets them by large fractions of hours. That drives all this growth in the GPU market. Previously,  some of this GPU market growth was largely driven by Bitcoin mining that you have known and seen. That’s changing as the cost of a mining a Bitcoin goes up. People don’t see that to be so much lucrative anymore. The other other side, the AI and machine learning, when you talk about those, they are the key major driving factor for the GPU market. In the GPU market, especially, also if you talk about th automated driving. All these big automotive companies, whether that’s a BMW, Uber itself, GM, Volvo, they’re all coming up with their own version of automation in fully automated driving. Ultimately this will help a lot of growth in the GPU market.

The way we see it is, there are basically two different product lines here. One that helps them to train the model of a particular data set, and one model that is going to therefore influencing, so making a decision live in the product. We have seen, and we have servers that satisfy both those needs.

Steve Sidwell:

Thank you so much for coming on The Tech Bench Podcast today. I really appreciate it. I know we all do. James thank you too, for being here. This was a lot of fun. I really enjoyed speaking with you and just learning more about Supermicro and your career there, and just everything that’s going on in the market space right now.

Yagnesh Ashara:

It’s definitely a pleasure being here. Steve James, thank you for your time.

Steve Sidwell:

Thank you for joining us for another episode of the tech bench podcast. If you enjoyed this episode, please make sure to subscribe and follow us on Instagram, Twitter, and Facebook at the LTTB podcast. If you have any comments, questions, or show ideas, please feel free to email us@thetechbenchatliquidtechnology.net. For show notes, visit liquid technology.net/tech bench.