Article Info
Five Misconceptions About the 10G Optical Market
I’ve noticed a common trend during conversations with investors and analysts about the state of the optics market. People seem to be staking their hopes on 10G as the growth driver for the industry. I firmly believe this is true, but people are assuming the gains will be evenly distributed among all players. Here are the common misconceptions:
Video is Driving 10GbE
This is by far the most over-used and misdirected belief and causes me physical pain when I hear it. Yes, video is playing a major role in the expansion of WDM transport networks. Yes, most of these new links are 10G. But much more growth is coming from the datacenter as bladeservers and computing infrastructure transition to 10GbE. Companies like Broadcom are driving down the cost of putting 10GbE in servers, and subsequently will drive demand for 10GbE switching equipment. Other trends such as re-using the 10GbE standard for datacenter storage interconnect amplify this trend.
At best, 100k new 10GbE connections will be deployed in the core network in 2007. Compare that with the 400k or so modules Cisco alone will sell in the same timeframe. In 2008 the gap grows much larger.
But everybody understands YouTube, so video remains the lip-synching poster boy.
10G is One Market
Few bother to drill down into the innards of module types and specifications to understand product mix, and are quick to dismiss the alphabet soup of module types and laser reach specifications. This is a huge mistake as the majority of unit volume gain in the next 5 years will come in a narrow segment of product and reach types.
The 10G optical market is really several different markets with little or no relationship to each other in terms of cost, market demand, competitive pressure, and growth. Little things, like the tendency for enterprises to deploy OM-3 fiber have an enormous impact on product mix.
Incumbent 10G Suppliers Will Benefit
Some existing suppliers will do well, but the mistaken view that 10G is one market drives investors to believe all companies supplying this market will benefit. To borrow a much abused term, the distribution of growth in the 10G market is ‘lumpy’.
Opnext (OPXT) makes great hay about the importance of the 10G transition to their business and justifiably so given they are top supplier to Cisco for all types of 10G modules. While it is impressive that Opnext (as well as Intel (INTC)) supply the bulk of the volume today, it isn’t clear that they have adapted their supply chain for the demands of tomorrow.
My belief is virtually all unit volume growth will be in the low-cost, short-reach segments of the market, a market characterized by brutal price competition and 20-30% gross margins. Vertically integrated companies will be positioned best when volume ramps. Opnext is a somewhat vertically integrated manufacturer of long wavelength modules but heavily outsources both components and manufacturing of short reach modules.
It would not surprise me to see Opnext acquire a VCSEL manufacturer itself to remain competitive as the short reach market moves to higher volumes. I also believe this was the primary justification for JDS Uniphase’s (JDSU) acquisition of Picolight. (see “Why JDSU Bought Picolight“).
Such an acquisition still leaves both companies with a cost structure unsuitable to high volume commodity production.
I believe Finisar (FNSR) is the best example of a company well adapted to such a commodity environment as they are by far the most vertically integrated manufacturer of these types of components.
It will be interesting to see how all companies adjust in the volume shift in the 10G market. Perhaps Opnext decides to focus exclusively on long wavelength products, like they already have at 2.5G.
SFP+ is a key growth driver
SFP+ is a new low-cost 10GbE form factor and is a hot buzzword. It will indeed have a major impact on the market and is the key battleground in the competition between optics and copper interconnect. Industry insider consensus is SFP+ will not be ready for prime time in 2008 and I agree.
This is one trend where I admit I could be wrong as Cisco (CSCO) could change the whole equation quite abruptly. Cisco itself indicates only a small fraction of the modules it buys in 2008 will be SFP+, but doubts linger in my mind.
As readers know, I believe Cisco derives enormous profits from reselling optical modules. The transition to SFP+ is critical to Cisco for two reasons.
- Lower cost modules mean lower cost 10GbE ports. Driving down 10GbE port costs is the most important vector for driving 10GbE adoption. Cisco must leverage their hegemony in 1GbE into 10GbE and not allow others to gain share. If they can drive costs better than anyone else they will capture the market share.
- Cisco will make significantly more money reselling SFP+ modules than the existing X2 modules. Pulling in this transition just one quarter drops $25m of pure profit right to Cisco’s bottom line.
I am nervous betting against Cisco’s desire to make money and retain hegemony.
Copper (10GBase-T) isn’t a factor this time
Jury is still out on this one. One thing for sure- it shouldn’t be a factor until 2011 or so and faces significant technology headwinds. 1GbE optical component growth stalled in late 2004 as it could not match the per port costs of Gigabit Ethernet over copper. We shall see if this time is different.
Author is long Finisar.
… and I vow to torture a small animal in my office every time I hear or read the words “Video” and “Explosion” used in the same sentence. You, your Marcom group, and your conscience have been forewarned. (Just Kidding.)
* good post
* 10GBase-T will become important because, as you note, the majority of the 10G connections are intra-POP, or data-center, based that can leverage this type of port. If it costs way less it will be used in all places that can live with the distance and cabling limitations. (Same thing happened with GbE in the POP/DataCenter.)
“If it costs way less…”
I think that the cost equation is really interesting here.
I would guess that curretly the BASE T hardware would be about a third of an LRM link. The cu solution, however, is going to burn 5 times the power. So, one fundamental question that I have (and I imagine has been analyzed extensively) is what is the cost per watt of power consumption and heat removal? Any numbers would be appreciated even if they’re back of envelope.
Andrew,
I agree with your analysis. I wanted to give you one more pointer to an interesting piece of “history”:
Four years ago there was an effort to get IEEE802.3 to study an option for a new standard for 2.5Gbps Ethernet (see grouper.ieee.org/groups/802/3/minutes/nov03/1103_CFI_2_5G_report.pdf). The 2.5G option was supposed to be a simple extension of the 1G option, very similar to the version of 10G that is made of 4 parallel links of 2.5G (but only one such link). It was supposed to cover both copper and optical options.
Even today, most of the Ethernet cables in use are CAT5E. Physical limitations make 2.5Gbps the maximum speed that can be supported on the worst case CAT5E over 100m (100m over worst case copper was the target for 100Mbps and 1Gbps as well).
The data presented to IEEE showed that the market will be much slower moving from 1G to 10G, relative to the adaption rate from 10M to 100M to 1G (the prediction was right on). In addition, component manufacturers showed that the cost of a 2.5G port will be almost identical to that of a 1G port. Similarly, a 2.5G optical adapter uses the same technology as (and costs marginally more than) a 1G one.
IEEE802.3 did not want to pursue this option, and continued concentrating solely on 10Gbase-T. The main reason was the the deviation from the 10x track record. I wonder how much longer it will take for the elusive 10Gbase-T to become a reality, and when will it finally extend from data centers to the enterprise space.
M@ – Copper interconnect doesn’t work because it’s advantages of low cost and high density run counter to high power consumption.
Also, don’t forget the added latency of the copper 10G-BaseT spec. Some people make this out to be a big deal. I don’t know if it is or isn’t.
Adi – one thing for sure is the Desktop market has not embraced 1G as a must have. Sure, the technology is built in with new PCs and low-end switches – but I doubt many are pulling out 100M switches in mainstream applications to put in 1GE. It is hard to even think of why 10GE to the desktop would be remotely useful.
I think 1GE copper and 10GE optical will have a long run. I think 100GE optical transport is the next big thing.
For scale out server connectivity within a data center, CX4 copper cabling (essentially the same as what is used with InfiniBand) is by far the cheapest solution and will remain so for some time, and it also has the lowest latency and consumes the lowest power. While it has drawbacks (heavy bulky cable, limited to 15 meters, complex connectors, lack of structured cabling patch panels, and inability to assemble cables on site), recent advancements allow connectivity up to about 45 meters, and for shorter distances permit smaller diamater, more flexible CX4 cable.
But the biggest roadblock to broad 10GE server connectivity hasn’t been addressed here. The incumbent 10GE switches have too few ports, and cost way to much per port for anybody to connect masses of servers with 10GE. For example, a 2nd tier incumbent supplier showed a (non-functional) switch with 128 10GE ports, at a list price of > $5000 per port (not counting optics), and a rated power budget of about 80 watts per port. If, for example, you wanted to connect 256 servers with any-to-any non-blocking throughput, you would need a two-stage distribution layer, where three switch ports are needed per server connection, since 1/2 of the ports on each switch connected to the servers need to connect to a second stage in order to provide non-blocking throughput. Very few data centers are going to spend $15K per 10GE server connection to connect $5K rack mount servers together.
Why is this so? The port counts are limited because all the incumbent data center switches use single-stage central crossbar switching, whose capacity is limited by the I/O bandwidth per crossbar chip. While the maximum bandwidth per chip (i.e. the maximum feasible number of data pins per chip multiplied by the maximum frequency of the signal) is improving over time, at roughly 2X every 4 years, it is way lower than Moore’s Law. The cost, complexity, and power consumption are high because they have lots of separate chips for buffering, QoS (i.e. bandwidth priority rationing), packet inspection, and so on.
InfiniBand allow multi-stage fabrics, but comes with the challenge that if you want good performance, all the applications need to modified to change from standard TCP sockets APIs to InfiniBand-specific I/O calls.
There is yet one other problem in large server deployments that is just now beginning to be recognized. It has showed up in GE-connected server farms, and InfiniBand HPC clusters. The problem is congestion, which manifests as higher than expected latency in HPC clusters, and lost data during bandwidth spikes in data center server farms. The problem is that all the existing InfiniBand, Layer 2 Ethernet, and Layer 3 routing protocols are static. Imagine driving from the North side of Chicago to the South side and having to stay in the same freeway lane the whole way. Bottlenecks are inevitable, and, just like traffic jams are worse in major cities than in small towns, the bigger the server farm, the worse the problem.
The biggest technology headwind for 10GBASE-T transceivers is the enormous power consumption required to meet conventional 100m Ethernet reaches – currently 11-14Watts per port with first generation going to 7-9Watts per port in 18-24 months using 65nm technology.
SFP is like SouthWest Airlines. It reduces optics to the bare minimum functionality of E-O-E, & raises an important question – are today’s optical PHY standards over-spec’d? Can Cisco afford to throw out the demarc between analog and digital – “Just toss those raw electrons over the wall and we’ll take it from there”?
No, I don’t see Cisco taking any hasty Plug-n-Pray risks. It’ll take a few more years around the track at Monza before SFP gets FAA approval. Right now SFP is just a stick to drive optics vendors down some more.
The correct question for optics vendors is how many optical XFP’s will ship in the next 3 years. There is some solace here from CX-4 sharing the same form factor.
The comment script clobbers the ‘ ‘ signs. SFP above should read SFP-Plus.
i thought this was interesting…
perhaps the general mass uptake might not be as far away as people think?…
http://www.eetimes.com/news/latest/showArticle.jhtml?articleID=199600035
”
EE Times: Latest News
Infiniband proponent rolls out 10G Ethernet device
Rick Merritt
(05/21/2007 12:00 AM EDT)
SAN JOSE, Calif. — Mellanox Technologies, the sole provider of merchant Infiniband silicon, is jumping on a growing bandwagon of companies fielding chips for 10 Gbit/second Ethernet.
Overall the Mellanox part falls in the middle of a pack of a half dozen competitors, in line or ahead in some respects, but behind in others.
“We think Infiniband is a huge growth market, but we want to grow even faster than that market,” said Dan Tuchler, director of product management for Mellanox.
“There’s no doubt 10G Ethernet will be a huge market too, and certain things are coming together to suggest that it’s about to take off. So we want to layer one growth market on top of another,” Tuchler added.
Accelerating the move to 10G at the server, switches for 10G Ethernet are coming down in price, and 10GBase-T copper links are at least on the distant horizon as affordable, he said.
In addition, the latest quad-core processors are doing a good job of handling part of the 10G processing tasks while virtualization technologies are ratcheting up the amount of I/O data centers need, he added.
The Mellanox chip is supporting per priority pause, a congestion management feature, as one step toward distinguishing its Ethernet chip.
The feature is currently under discussion in an Ethernet IEEE standards group.
“We wanted to get this feature out there and help shake it out in the market,” said Tuchler. “We think its one of the biggest and most significant pieces of congestion management,” he added. ”
…..
“The 21×21 mm chip integrates CX4 and XFP drivers and consumes as little as 12W on a CX4 link. ”
…..
“Likewise Mellanox is running standard Windows and Linux Ethernet drivers, but not the OFA stack.
“In the future, we are looking at running some of the OFA components,” he said.
Mellanox has implemented the Receive Side Scaling capability Microsoft is putting into its Ethernet software stack in Windows.
The capability lets the chip associate particular Ethernet flows, MAC addresses or virtual LANs with individual cores in a multi-core CPU.
Mellanox also offers the feature in Linux and the VMware virtualization software. ”
…..
“By the end of the year, Mellanox will release a variant of its ConnectX EN that can handle either 20 Gbit Infiniband or two10 Gbit Ethernet links.
Early next year the company will ship an adapter that can handle 40 Gbit Infiniband links via four 10 Gbit serdes.
Mellanox developed its Ethernet technology in house. Adding an Ethernet framer and MAC to the existing serdes and other I/O components on an Infiniband part was not that difficult, Tuchler claimed.
“We have shipped more Infiniband chips than all the 10G Ethernet chips combined,” he said. “
i dont have a sub so can confirm the source
http://www.the451group.com/report_view/report_view.php?entity_id=48272
“Broadcom finally expands NetXtreme line to focus on unified 10GigE NIC
The company has entered the 10GigE NIC fray with its first offering specifically designed for use as a converged or unified platform, bringing together networking storage and other functions in a single card.
MIS/ Impact Report, 16 May 2007
Greg Quick”
theres also the mass SAN storage markets that will be looking at all this too, perhaps even the dinky NAS ;) too.
opps, didnt see that link in the story , sorry ;) still a valid point about SAN i think helping drive the trend…
Andrew,
Interesting post. I wondered if you had any thoughts on the impact of 10GBase-LRM and how the slow rollout of products to address this legacy fiber market impacts the short reach market you refer to?
Hi Andrew,
Nice post…insightful. Thanks. :-) I wonder what your thoughts were in relation to the SFP as it’s seen some decent interest of late for 8g storage applications.
Thanks!
Andrew,
I know video thing is driving you nuts.. but that’s where more and more of the “bits and bytes” of IP packets are coming from, and flow thru the data centers, etc. So while it’s “hyped”, it has the ring of truth..
Face it, we’re evolving with our means of communication from text, to voice, to ..Video. there is no escaping..
Not really driving me nuts. I do believe that Video is driving traffic growth. I just don’t like the way it is universally applied to every business segment as justification for beyond trend growth. I am more disturbed about how little quantification there is for something so broadly believed.
I smart guy I know compares the Video Traffic ‘explosion’ hype to the “traffic doubling every 9 months” hype to 1999. I don’t know that the claims are false, I just haven’t seen enough data to make me believe they are true.
In a more global sense, more business is moving on-line, or more GDP is moving on line. To replace the brick-and-morta business model, the winning business is towards a friendlier but simulated “real world” experience. “video” in a broader sense is a re-created human experience for an excelerated, and increasing degree of human interactions/connectivities. At the end of the day, “digitization” of a hyper connected human experience is driving the bandwidth demand, and exponential “bits and bytes” growth. “Video” is just a catch phrase for this transformation.
Trackbacks / Pingbacks