Dodaj do ulubionych

OS on Chip; LINUX - BIOS - przyszlosc komputerow

IP: * 26.07.04, 08:22
An interview with Tyan president and CEO,
Symon Chang

Monday 26 July 2004

Tyan Computer Corporation showed its latest Intel and AMD server and
workstation offerings at Computex Taipei, 2004.
An interview with Tyan’s president and CEO, Symon Chang and his special
assistant Steve Hsia about the company’s plan to become one of the premier
white box providers for all levels of systems for the enterprise.

Tyan’s present direction

Q: What is Tyan’s focus, right now?

Chang: Tyan is very active in the workstation OEM/ODM market, with a focus on
channel sales. By highlighting white box offerings in the server market, Tyan
is looking to be a solutions provider for large systems integrators. We may
appear very active in Opteron business, promoting many new Opteron-based
systems; however, our primary business remains Intel-based machines.
If I were to put a number on it, I would say two-thirds of the company’s
business is from Intel-based products.

Q: Where is Tyan going with its server offerings?
Are you looking at the high end or low end of the market?
How are you managing to maintain a high-level reputation for your products?

Chang: That is more of a branding issue, and we are not a brand name.
We do not want to compete with IBM, Dell or the other large system
integrators. We want to be their main supplier.
When you open up one of their machines, we want you to see a Tyan board.

Nevertheless, when it comes to large custom-made clusters, we will compete in
that area because the larger labs don’t want to buy from the mainstream
system integrators. As well, IBM does not want to bother selling to the large
labs. Someday, we will make larger offerings for that market, but the large
clustering market is really a small niche.

Q: Currently, companies such as IBM are moving from being hardware providers
only to Systems Integrators. Does Tyan have any plans to enter this market?

Chang: There are two channels in which to sell server offerings; the first is
to the distributors and resellers. For example, customers like Ingram Micro
and Arrow buy barebone systems. We sell them a white box, and they will
install their own software and resell it to the end customer. The second is
to sell directly to the system integrators and the companies providing
clustered supercomputing for budget minded operations. These companies create
value with their mechanical solutions, with power distribution for example.
Many companies, including Linux Networks and RackSaver, supply computers for
the gas and oil industry. We also had a systems integrator who installed a
super cluster for the Los Alamos National Laboratory in the United States.

Q: What do you see as the demand drivers for servers over the next couple of
years? We’ve recently spoken to memory makers who have lamented that demand
for DRAM in the server market has not lived up to expectations. They are
hoping for a pick-up in demand for server memory. As applications such as
VoIP begin to mature, do you see the demand for memory in the server industry
starting to grow exponentially?

Chang: Naturally, as the industry migrates to new operating systems, demand
for additional memory will increase. Right now, under Windows XP, you need to
have 256MB to 512MB of DRAM to operate comfortably and without latency.
With new 64-bit operating systems coming out, though, you’re going to see a
noticeable increase. At a minimum, the requirements will double, and
realistically, will be even higher. When the desktop memory requirements
finally go to 2-4GB of memory (I’m not sure when that will actually happen),
that is when you’ll see the majority of users moving to a 64-bit platform.
At that time, memory requirements will need to to increase, for both the
workstation and server markets. If I were to hazard a guess, I would say that
within two years, 2GB will become the new industry standard for the basic
amount of DRAM required for smooth system operation. Gamers, of course,
always go for larger amounts of memory; for gamers, no amount is too high, as
long as they can afford it.

Hsia: For the server market, the PCI Express bus is a revolutionary change
from the current computing platform. It allows companies like Tyan to be more
flexible in making both servers and workstations.
The first generation can run at 4x, allowing customers to upgrade to 8x or
16x when they are ready.
In addition, when the card manufacturers come to market with a brand new
product, they will have greater flexibility with bus speed, allowing cards to
be upgraded effortlessly. Currently, speed upgrades are totally bus
dependent, so PCI Express is really fresh for our industry, as well as for
the desktop. We have much to look forward to.

Q: From a corporate perspective, can you say something about how plans have
changed over the last few years, and how you see them developing in the next
several years? Could you tell us something about Tyan’s relationship with
Mitac, as well as Tyan’s IPO, and how you will be funding expansion and

Chang: Well, we have really worked well with departments within Mitac, on a
variety of OEM projects, utilizing Mitac facilities for the manufacturing.
We moved our headquarters to Taiwan because our manufacturing is here.
Being here allows us to better consolidate our development. Taiwan is our
financial and logistics center. Our research and development in Taiwan and
Shanghai will now be our main centers, although we will continue to maintain
our advanced development centers in the United States.

Q: From a funding point of view, are you looking to attract funds?

Chang: No, we’re doing fine on the financial end of the business, and our IPO
is mainly for the employees.

Future computing and software

The conversation turned to Dr. Chang’s views of the future direction of
computing over the next few years and how operating systems can affect demand.

Q: Do you see 64-bit computing as the future for servers? How about operating
system support? In which direction is Tyan heading?

Chang: I think so. The primary reason for this is that most modern servers
require a large amount of memory, and if they need to go over 4GB, a 64-bit
solution is really the only way to address that issue. The Linux business in
64-bit is pretty far along, but Windows will take a little time. Primarily,
64-bit is already the de facto standard in the Linux RISC PC business, and
probably in some selective workstation markets, but not in the generic
workstation market.

Q: Do you think that PCI Express will drive demand?

Hsia: I think that a lot of things will go hand-in-hand. For example, in the
next year or two when Microsoft’s new Windows OS, Longhorn, is officially
announced, you’ll definitely see a big spike in the adoption of PCI Express
based products. Between now and then, it’s really just a waiting period. 2006
is really going to be a big year for computing, and we will see some dramatic
changes in the way computers and systems are made.

Q: Do you think, as we get closer to Longhorn, that sales will drop off, with
people growing tired of the wait?

Hsia: I don’t believe so! Now, in the HPC (high-performance computing)
market, most of our customers are using Linux, and as Linux’s clustering
capabilities mature and improve, we believe the uptake will stay steady.

Q: Do you think Microsoft’s 64-bit OS will come out on time?

Chang: I hope so. There are delays, but I believe it will. Interestingly
enough, a couple of significant things have happened this year; for example,
Intel’s Xeon processor with 64-bit extensions is a reaction to the unexpected
popularity of AMD’s Opteron, w
Obserwuj wątek
    • Gość: bzykbzyk. OS on Chip; LINUX - BIOS - przyszlosc komputerow IP: * 26.07.04, 08:25
      Q: Do you think Microsoft’s 64-bit OS will come out on time?

      Chang: I hope so. There are delays, but I believe it will. Interestingly
      enough, a couple of significant things have happened this year; for example,
      Intel’s Xeon processor with 64-bit extensions is a reaction to the unexpected
      popularity of AMD’s Opteron, which put Intel under pressure to provide a
      similar solution for the OEM market. If Intel had not reacted, it would have
      lost out. Their response was to come out with a 64-bit CPU that is not optimal,
      but at least they have it, and I would compare that with what Microsoft is
      doing now in the realm of the 64-bit operating system.

      Q: You said that you expect a big spike in 2006; however, research groups such
      as IDC and Gartner have said that 2004 will be a bumper year for the desktop
      market, with new PC purchases partly driven by replacement of previous Y2K-
      upgrade machines. Where do you see the differentiation for the server market?

      Hsia: It really depends on how you look at the playing field; maybe Symon could
      better explain how the market will treat servers.

      Chang: Well, I’m really only watching the 64-bit solutions. When Microsoft
      officially launches its 64-bit operating system, I expect to see more demand
      for Tyan’s 64-bit platforms. If Windows does not come out with a viable 64-bit
      OS soon, they will lose out as Linux takes more market share, at least in the
      server business. The fact that Linux already has a viable 64-bit OS and is
      adding new 64-bit programs daily is a major reason for its current adoption in
      the server market. So it’s to Microsoft’s advantage to come out with its 64-bit
      operating system as soon as possible, to prevent Linux from running away with
      the server market

      Q: Is Tyan now working with the Linux community to optimize its products for
      use with the Linux kernel?

      Chang: We do a couple of different things, but not so much in the driver field.
      At the moment, the company is working in the HPC arena developing diskless
      clusters in large installations, by compressing a bootable version of Linux
      onto Flash. That is a big job for us, to port all of our systems to a Linux
      BIOS. Tyan has been working for some time at developing a Linux BIOS for its
      motherboards. The company will soon switch all of its boards to having only
      Linux in the BIOS, for reliability and for quick boot and reboot capabilities.

      Q: What are the reasons for Tyan moving to a Linux BIOS? Is it for a faster
      boot time when recovering from power outages?

      Chang: No, mainly, the switch to the Linux BIOS is a US government requirement.
      We’re doing this to comply with government lab regulations for the improved
      security and reliability that a Linux BIOS affords. And it really isn’t about
      power outage recovery, but more for getting rid of the hard drive for boot-up
      purposes. Use of a Linux BIOS is in response to the requirements of the
      diskless systems that the government labs are using. For example, when you have
      a thousand nodes, and a thousand drives at each node, every time you reboot the
      system you run the risk of failure. What we’re working on is trying to reduce
      the number of failures. This is becoming a requirement for most of the large
      government installations we handle. Performance is one issue, but it takes a
      back seat to reliability. When we test and verify our motherboards, we
      initially use Unix and Linux to put them through their paces because only Unix
      and Linux have the capability to load the boards properly with verifiable

      Q: In the area of diskless workstations, does Tyan work with Wind River’s
      VxWorks at all?

      Chang: We have embedded customers who use many different operating systems, but
      we are spending more time working with Linux. Most of the company’s effort is
      spent on server management. The main reason why Infiniband has not taken off is
      the complexity of the software. Embedded Infiniband controllers are definitely
      in the company’s future. The only competition for Infiniband is PCI Express.
      Generically, Infiniband and PCI Express come from the same origins.

      Hardware development

      Q: Has the launch of the 64-bit Xeon helped sales of Intel platforms?

      Chang: No, but at least it has helped to slow down the penetration of the
      Opteron in the server market by giving those in the commercial server business
      another choice.
      I do not really expect it to help the overall sales of Xeon platforms.

      Now, there are a few OEMs pushing AMD platforms in the commercial enterprise
      market, but mostly, I think you will see Intel defend its share of the
      commercial market vigorously. With the 64-bit extensions, they have the tools
      to do that.

      Q: Can you talk about the heat dissipation problem with the Xeon Nocona?

      Chang: Nocona’s heat dissipation is a problem, now it is running at 130W.
      In contrast, the Opteron is running at 95W, and this is an advantage in its
      adoption. The power requirement and related heat dissipation problem are two
      reasons for the migration to a wholly new design in CPU architecture.
      One example is Intel moving to a dual core in its next generation processors.

      Q: Do you think heat problems are affecting the reliability of production

      Hsia: Well, heat problems do lengthen the testing and verification procedures,
      and increase time-to-market, not just for us, but also for manufacturers. But I
      don’t believe they adversely affect the end product in any way.

      Q: We’ve seen the four-way server offerings from Tyan, including the one
      launched last year, and this year’s version launched here at Computex.
      Does Tyan have any plans for building larger platforms?
      Does Tyan have an eight-way in the works?

      Chang: We do have an 8-way server board, which is primarily a four-plus-four,
      but we’re not offering that right now with the present chipset. I am aware that
      our competitors are doing that, but we’re going to do that in the PCI Express
      generation. We felt that the current chipset does not have the right I/O
      capabilities and so we might as well wait until the new ones come out.

      Around 2006, when the market moves to AMD’s next generation of chips, you will
      be able to go over 8-way.
      What I mean is that with eight sockets, and dual cores, you then have sixteen
      processors, but with K9, you’ll see it go over that. I think we’ll see a
      significant increase in the amount of crossbar switches in the CPU. I’m not up
      on all the minute details, but you’ll be able to go over 60 processors without
      adding any external crossbar chips. We can do all that within the structure
      that is being currently created. The crossbar bar chip is the standard in the
      mainframe business whether it is for the Xeon, Opteron or other processors.
      There are a couple of versions of the crossbar chip today, but I don’t think
      that anyone is currently using them for anything in the generic market; these
      solutions are really only for the mainframe market. Today’s mainframe market
      with computers from IBM or Sparc will be using up to and over 128 processors,
      with chips such as IBM’s 390 microprocessor. These machines are starting around
      US$1 million.

      Q: Last year in July, Dell was saying that it had stopped development of its
      eight-way server because the speed at which the smaller machines were
      developing was so much faster than that of the larger architectures, and that
      by clustering the smaller machines, you would get a cheaper more upwardly
      expandable solution.

      Chang: The MPI and SMP protocols are totally different approaches to similar
      computing demands. If you run a database or any commercial application, you
      still cannot compete with a clustered environment with a highly coupled SMP
      • Gość: bzykbzyk. OS on Chip; LINUX - BIOS - przyszlosc komputerow IP: * 26.07.04, 08:28
        Chang: The MPI and SMP protocols are totally different approaches to similar
        computing demands. If you run a database or any commercial application, you
        still cannot compete with a clustered environment with a highly coupled SMP
        environment. It’s like comparing a cluster of two two-way systems with a four-
        way system; the four-way system will be more efficient. The reason is latency.
        Even with Gigabit Ethernet between the two servers, the delay is around 30
        microseconds. So, with every computing cycle being a nanosecond or so, you have
        to wait for thousands of cycles before you can pass the information to the next
        node in the cluster, and that can cause a backlog when processing large amounts
        of information or compiling large programs.

        The products

        Q: Your 1U, 2U and 4U offerings, are they all OEM products?

        Chang: Yes, at the moment Tyan’s Transport 1U, 2U and 4U products are all white-
        box solutions; there is no own-brand in this area yet. I believe that all
        companies are going to have to focus on more specialized offerings that answer
        specific solutions and needs. Tyan’s barebone solutions for Systems Integrators
        are the company’s best way to leverage the R&D capabilities and certifications
        the company already possesses.

        The 3U and 4U offerings are easier to manage with barebone boards, when taking
        into account heat dissipation with some of the newer processors. You can easily
        get a case from Chenbro and set it up yourself. The 1U and 2U offerings are
        another story; they are more difficult to piece together because of some of the
        same issues. Heat is the most difficult problem to deal with in the smaller 1U

        Q: Is Tyan going to get into the blade server market?

        Chang: Tyan’s foray into the blade server market is only in the OEM arena. When
        blade servers see standardization, Tyan will enter the market with its own-
        brand offerings. The blade servers made by Intel, IBM, and HP are quite elegant
        solutions, well put together and clean architectures. The lower-end Sun Blades
        are what the smaller companies can afford, although they’re not as good as the
        more expensive counterparts.

        Two blade offerings now have the possibility of becoming the standard for blade
        server form factors. The one for the high end of the market is from IBM, and
        the one for the low end is from Sun.

        We may be producing a dual-Opteron blade. We are not working with Sun or IBM
        yet, but we are working with some second-tier SIs.

        Right now, every company has its own blade server design. Companies such as
        IBM, Intel, Hitachi and Fujitsu-Siemens all have their own design. HP and Dell
        also both have their own blades, and, as far as I know, HP’s is quite well
        designed, but I’m not sure about the Dell. I’m not too confident they have a
        very good design.

        Q: On the blade server solutions, how is Tyan looking into future development?

        Chang: Well, there is a lot of research in this area. One interesting area is
        the development of optical backplanes, where you converge all the
        HyperTransport signals into optics, and then you do not have to worry about
        distance as long as you have fiber lining all of your machines. This is one of
        the technologies that will be coming in the latter stages of the HyperTransport
        and PCI Express generation of computers.

        Q: What area of the workstation market is Tyan focused on? Does Tyan have any
        boards that are geared more towards the gamer side of the market? Do your
        boards support AGP pro slots?

        Chang: The type of workstation we build focuses on the engineering side of the
        workstation market. We cater to the large OEMs, and they dictate the type of
        platform and form factor that we support. All of our workstation boards now
        have AGP pro slots, and these are 4x- and 8x-capable. However, as all of the
        boards move to PCI Express, the AGP support will follow suit, I believe right
        now we are using Pro 50, and some of the next generation will go to x16 on the
        PCI Express boards.

        Q: As a current user of a Tyan dual-CPU board, running a Linux cluster, I am
        looking forward to the advances in hardware support for clustering. Is it the
        individual or corporate user who is expected to adopt these new products first?

        Hsia: I think what you will see, is a market that will be maturing over the
        next few years. SMB and enterprise customers, who always need to stay on the
        cutting edge, will spearhead the uptake of the newer offerings. It is really
        when the software that runs on these newer platforms matures that you will see
        a spike in the uptake.

        Chang: The architectures are all moving to a serial bus. The I/O bus, USB, a
        higher front-side bus speed, and fully buffered memory will require a serial
        bus, and all of this will make life easier for motherboard makers. All the
        designs will be much simpler; there will be fewer pins and simpler cables.

        Q: As the motherboards move to a serial bus, do you foresee a reduction in the
        form factor size?

        Chang: Well, they will certainly get simpler. And the move to a serial bus will
        have the effect of reducing lead-time for research and development.

        Hsia: The move should allow Tyan to move to market quicker with its products.

        Q: Are your current boards supporting an 800Mhz bus speed? Is the bus speed for
        both AMD and Intel boards the same?

        Chang: The boards for AMD that use HyperTransport technology support an 800Mhz
        bus. Intel does not have a point-to-point bus; however, it does support an
        800Mhz front-side bus. I think that by the end of the year, you will see both
        AMD and Intel boards moving beyond those speeds.

        Q: Where do you see the limitation of the bus speeds on the motherboard?

        Chang: Well, at the moment, with the copper interconnects, I believe we still
        have some room to grow on the board. Off the board, they’ve got copper
        achieving around 10Gbps, and on the board we could even see that speed increase.

        Q: What are the main differences between the Hyper-threading capabilities
        offered by Intel, and the HyperTransport ones offered by AMD? I know there are
        differences, but for some of our readers, I would like to clarify them.

        Chang: HyperTransport is a bus related enhancement, and Hyper-threading is a
        logical division of a single CPU into two virtual CPUs. In my opinion, Hyper-
        threading has not seen wide adoption in the industry, with the exception of the
        gaming market; I do not know how useful it would actually be. What I mean is
        that the software has to be tuned to take advantage of that, by optimizing the
        usage of the process, but as far as I know, most applications aren’t taking
        advantage of it, and I think that with most current applications, it can’t be
        done without a drastic re-write. HyperTransport will have more use in the
        overall computing market.

        Q: Apart from the Intel and Opteron high performance systems, do you have
        anyone working on ARM or Xscale CPUs?

        Chang: No, we do not. We do have special appliances that some of our customers
        have requested, such as encryption devices. For example, we have made a load
        balancing security server, and we have used chips from Cavium for some of these
        solutions. However, that is really more of a custom design for the embedded
        market; we do not go out of our way to build those types of solutions.

        Q: When it comes to high performance networking, do your motherboards use pro
        networking chips? I know that the Tyan Tiger 2466 uses a 3Com 10/100 base T.
        What other network chips are you using?

        Chang: We used 3Com Ethernet chips on earlier motherboards because we did a
        reference design for AMD, and they would not allow us to use Intel network
        chips. That has changed now, and we are using both Intel and Broadcom
        networking chips.

        Q: What is the biggest cluster using Tyan products?
        • Gość: bzykbzyk. OS on Chip; LINUX - BIOS - przyszlosc komputerow IP: * 26.07.04, 08:29
          Q: What is the biggest cluster using Tyan products?

          Chang: The largest supercomputing cluster using Tyan boards has 640 four-way
          nodes. It is located in the Chinese Academy of Science in Bejing. This computer
          has 2500 CPUs and ranks in the super computer Top Ten. It is the largest four-
          way cluster of Opterons anywhere; it is rated at around 10 teraflops.

          To be in the Top Ten of super computers with Opterons is quite an
          accomplishment because, at that level, you’re competing against vector
          processors and Itaniums.
          It’s really a price issue; the fastest systems are usually priced at around
          US$400 million. The Opteron is well suited to the clustering market, and
          utilization of the CPU is heading in that direction. Company budgets can’t
          normally afford the most expensive super computers, but that’s the market in
          which we would like to compete.
Inne wątki na temat:

Nie pamiętasz hasła

lub ?


Nie masz jeszcze konta? Zarejestruj się

Nakarm Pajacyka