Wednesday, December 19, 2007
Interview with Russell Stern, Solarflare
Russell Stern is the President and CEO of Solarflare Communications (www.solarflare.com), a venture backed firm developing 10G Ethernet network interface cards and silicon. The firm is one of the only firms with 10GBASE-T PHY technology--which allows 10G Ethernet over standard, twisted pair copper, instead of optical connections--and Stern says it is the only one working in production. Stern has had a wealth of experience in the industry--he is the former CEO and President of JNI Corporation, was General Manager and COO of Quantum Corp, and also was an executive at Western Digital. To get some insight into where 10 Gigabit Ethernet adoption is, what the market looks like, and where Solarflare is headed, we spoke with Stern last week. He spoke with socalTECH's Ben Kuo.
Tell us a little bit about Solarflare?
Russell Stern: We started as a PHY company, and developed the first, 10GBASE-T PHY, and still the only one that works, despite what you might read. In April of 2006, we purchased Level 5 Communications, which gave us an extremely strategic piece of silicon and software. Level 5 was developing a 10G, Ethernet controller that is phenomenally low power -- only 1.9 watts, full line rate, and has a very unique virtualized interface, which enables us to do virtualization acceleration on all platforms -- including VMware and Xen. That piece of silicon is part of a core that will enable us to integrate a PHY and a controller together, to create the first 10G-based LOM (LAN on Motherboard) chip, which we will deliver in 2009.
You were in the business of PHY chips, why did you buy Level 5, in the controller business?
Russell Stern: What may help to explain that is to talk about JNI. When I got to JNI, we had three divisions -- Fibre Channel, iSCSI, and Infiniband divisions. We were actually doing reasonably well in Infiniband. However, in that business, we were buying IBM and Mellanox chips, and putting them on a board, and selling HCA (Host Channel Adapters) -- but there was no value add. We were buying chips, and buying the board, and the software was mostly open source -- so there was no way for us to capture value. So, we decided that business was going to stay small, and was likely to continue to be small. My general manager and I argued a bit, and then we fundamentally agreed--iSCSI is going to take over storage networking, Fibre Channel is going to go away, and it's all going to go to Ethernet. So, we decided to look at what was out there in Ethernet.
At that time, there was only Alacritech which had 1G iSCSI technology--which everyone believed at that time required a TCP/IP offload engine in hardware. There was a ton of venture capital money going into companies to do 1G TCP/IP offload hardware. However, those boards cost $1400, and I didn't understand why a customer would pay $1400 for a 1G Ethernet thing, when you could get a 2G, going on to 4G, Fibre Channel card for $500. We had a lot of cash, so we went to talk to some of these companies about 10G offload technology, where they were, and what they knew about it, and when they would have it. We went to a bunch of companies and asked--why do you need this technology? They told us -- the processor is not fast enough. I asked them--how about when they are fast enough to do this? Isn't there a lot of risk trying to do this in hardware? They would just kind of stare at you. To tell you the truth, every time I pushed on them, their answers came back soft. None of them really understood networking, networking software, and the dynamics of the business. I figured out that best case, there was this short gap of only one or two years where processors would not be fast enough. Plus, what they were building was incredibly high power--mostly in the 8-14 watt range.
So, we started building 10G at Solarflare, and we knew that we needed to get NICs into market. We started calling on the NIC players, and they were all building very, heavy powered NICs, near the limit for PCI-E--without a PHY on it yet. Then, we came across the Level 5 guys, talked about what they were doing, and it was tremendously exciting. These guys are out of the AT&T research labs, out of a computer laboratory in Cambridge University. They are true network engineers, and understood the software, the hardware, and the architecture, and how to make it work faster, and understood virtualization and the trends in the business.
We went out to raise money in late 2004 and 2005, and we met with Bandel Carano, of Oak Investment Partners, which also had an investment in Level 5. When he decided to fund us, he told us it would really make sense to put the companies together and integrate the technology. I told him--I have my own issues, let me work through that and we'll talk another time--that lasted about six months, and we decided it was time to put the companies together.
Is your business now both Ethernet controllers and PHYs--are you still in the PHY business?
Russell Stern: PHYs are huge. It's a combination of the two. The 1G Ethernet market is owned by Broadcom and Marvell. In server land, in the motherboard space, Broadcom has 100% of the market. Intel has 0% of it. Intel gets some LAN on Motherboard for desktops and notebooks, but they have 0% of the server. Intel doesn't have a 10GBASE-T PHY, they only have a controller. There are only two companies that have all of the necessary IP to make LAN-on-Motherboard for 10G--Broadcom and Solarflare. And, Broadcom does not have a 10GBASE-T PHY, they're significantly behind us, more than two years. The 10G controller market is owned by Broadcom and Intel, who are the leaders. But, only Solarflare has the ability to integrate a controller and a PHY--which is up and running, working, and shipping--and we'll be the only company to ship integrated silicon in 2009 or 2010.
So Broadcom and Intel are your major competitors?
Russell Stern: Yes, but at the same time, there's a level of co-opetition as well. Our PHY works with Broadcom and Marvell. In the controller market there are lots of players. However, there's no way for the controller companies to develop a PHY, which is a $100M proposition. Who's going to raise that and go develop one now? We still have the only working PHY in the industry. We have a tremendous advantage with the architecture of Level 5, and our two year lead on a PHY. We were shipping our first standard PHY in August of 2006, and it's now December of 2007 and we're still the only one out there.
The 10G Ethernet market has had some issues getting some traction, what's your view on the industry and where adoption is now?
Russell Stern: Solarflare has always had the view that people have wanted 10G, but the lack of copper has inhibited the market. I've heard stories--that there's no killer app yet, that people are not using 1G pipes--but frankly, that's not the case. Anyone coming in with multiple 1G NICs obviously needs more than 1G. The problem is, the way that data centers are built today, they are built using copper. They're built using copper for a very specific reason. If you look at the way they're configured, there are patch panels where they wire servers up, and at the top of the rack there's a switch, another patch panel, that goes over a longer run, over to some core switches or storage for distribution, and yet another patch panel. They do that because it's impractical to wire everything point-to-point. You wire through a patch panel, so that the data remains flexible. But, there's no analog in the optical or CX4 world (Editor's note: 10GBASE-CX4, 802.3ak, uses copper cabling similar to Infiniband to transmit Ethernet data). In CX4 world, there was a limit of 15 meters, and you can't run through a patch panel. Optical still remains expensive. And even if it wasn't, you still don't have a way to build a data center using it. You can use it from core switch to core switch, or for an uplink for a couple of runs, but you can't really expect to build a full data center that way. So, that has installed adoption of 10G, because 10GBASE-T did not happen. That's changed. In our mind, the way to get adoption is to hit $500 per port. Right now, optical is $3000 per port. We enable that $500 a port.
Playing back to JNI -- servers, guess what--they are now 4-core, 8-core, and have plenty of bandwidth. We've got lots of gigahertz processors to do the work for me, and can handle 10G with no sweat. Add one more thing on top of that--the killer app, which is virtualization. You can take a hundred different applications and load them all up on one box. That bugger is network bound. It's a killer, you can't put enough adapters in--if the box is big enough, you need 10G.
In addition, there's all sorts of rich content -- for example, we've got an office in Cambridge, England, and in the last two years Internet traffic has increased 100 times. 70% of that is video. We're pushing lots more content, much richer content, there's virtualization, and storage convergence. 10G is here, it's simple and easy to use, and it's now at a low enough cost, from an economic point, to get it integrated and at a commodity level.
You mention virtualization -- it sounds from a lot of players that this is really having an effect on the market. Have you seen this?
Russell Stern: Yes, yes, yes, and yes. It's one of the hottest pulls on our product. When we go and show them our technology working, accelerating in our virtualized environment, people are just besides themselves--on how simple we've made it and how easy it is. You see it when you talk to guys like Dell--they just pre-announced a VMware ESX platform. We're seeing it in droves. Every single OEM we talk to goes into gory detail on how they support virtualization. By far, we think we've got the best story there.