Forgot your password?
typodupeerror
Silicon Graphics

SGI launches R16000 352

Posted by CowboyNeal
from the more-power-damnit dept.
nkrgovic writes " SGI has just launched a new CPU - the long expected R16000. The new CPU works on 700MHz, has 4MB secondary cache and more goodies. For now the new CPU is only used in SGI's Fuel workstations, but we should expect to see it pretty soon in SGI's Origin servers as well. With new high density compute nodes this should make the Origin's the fastest supercomputing server per square foot."
This discussion has been archived. No new comments can be posted.

SGI launches R16000

Comments Filter:
  • by Anonymous Coward on Friday December 27, 2002 @03:28AM (#4964948)
    So fast, it helped me get first post.
  • It runs IRIX? (Score:3, Interesting)

    by Russellkhan (570824) on Friday December 27, 2002 @03:32AM (#4964961)
    I'm confused. I thought SGI was dropping support on IRIX. Why are they releasing new Irix boxen?
    • While SGI supporting linux on IA32 and IA64 platforms, they're still developing Irix for their MIPS platform. No reason not to, it's still a pretty robust OS.
    • Re:It runs IRIX? (Score:4, Informative)

      by bmajik (96670) <matt@mattevans.org> on Friday December 27, 2002 @03:42AM (#4965008) Homepage Journal
      Nope.

      SGI will continue to make investments in IRIX and MIPS until it makes sense to move all of their products and customers to Linux on IA64, and that may not happen until theres something better than Linux+IA64 out :) So who knows when MIPS IRIX will officially go away.

      Linux isn't there yet for the bread and butter SGI customers. Neither is IA64.

    • Some of our customers' highest-end workstations are SGIs. (There are also a couple Sun Blade 2000s and a LOT of IBM RS/6000s.) SGI will continue selling IRIX and the machiens it runs on until they no longer make money doing so.

      - A.P.
    • Re:It runs IRIX? (Score:3, Informative)

      by blakespot (213991)
      Cause on IRIX you can run that super-cool file manager from Jurassic Park. [sgi.com] Why would you want to run that super-slow piece of dog shit Konqueror or Nautilus anyway?

      You can do it with OS X [uiuc.edu]...but in a swimming pool.

      blakespot
  • Uh, can I have that in libraries of congress, please? (Or at least cubic foot of server space / "per 1U rackspace").
  • by Anonymous Coward on Friday December 27, 2002 @03:53AM (#4965043)
    The SGI processors shouldn't be viewed as general
    purpose processor like the P4 or Athlon. For
    specific floating point intensive problems, they
    can be quite effective. What is annoying is that
    they are usually 2 or more generations behind in
    manufacturing process capability. So the lines
    and heat dissipation in the 3GHz P4 are much more
    advanced than the R16000.

    Also, SGI has an annoying tendency to use
    proprietary ASIC's in the their memory which
    make their entire system much more expensive
    than it need be. Some of this is because
    their design cycle is so long that when SGI
    committed to a architecture, the performance
    just wasn't there.

    Given these constraints, it is hard to see
    how SGI could market "cost-sensitive" systems.

    • Also, SGI has an annoying tendency to use
      proprietary ASIC's in the their memory


      If you're refering to the ccNUMA-style systems, it's not just an MMU - it's a whole different architecture for the system. They don't have a bus - they have a switch between core components as the central feature of the system.
    • Also, SGI has an annoying tendency to use proprietary ASIC's in the their memory which make their entire system much more expensive than it need be.

      Only in the same way that other manufacturers have an annoying tendency not to use Crossbow, which makes their entire systems much slower than they need to be.

      Given these constraints, it is hard to see how SGI could market "cost-sensitive" systems.

      It's all relative to what you want to do. Sure, SGI won't make a general purpose desktop (altho' they once tried to, with the Indy). But they are competitive for the markets they sell into, which require extremely high memory bandwidth and fast precise rendering. Sacrificing precision for rendering speed like a gamer's PC does isn't an option for CAD or medical imaging.

      At the high end, things like CPU power/square foot really do matter, and SGI are competitive there too. Hopefully, the company will be able to recover from the Belluzo regime.
    • What is annoying is that they are usually 2 or more generations behind in manufacturing process capability.

      This was true in the days of the R10000 and R12000. However, things began to change with the R12000A.

      The R16000 uses a 0.13 micron process utilizing copper interconnects. It is indeed buzzword compliant.

      SGI's R1x000 series is designed in-house these days and is fabbed for them by NEC.
      • The R16000 uses a 0.13 micron process utilizing copper interconnects. It is indeed buzzword compliant.

        Scratch that. As someone else already pointed out, the R16000 uses a 0.11 micron process. I was thinking of the R14000A.
    • Anonymous Wanker wrote:

      > SGI has an annoying tendency to use
      proprietary ASIC's in the their memory

      Little Linux Boy, can you say bandwidth? Check out the memory bandwidth specs on a years old O2 vs. a brand new Intel/AMD POS. Then look at the bandwidth of a new SGI workstation vs. promised Intel/AMD POS'.

  • by Anonymous Coward on Friday December 27, 2002 @03:55AM (#4965051)
    Didn't somebody from Google recently say that server density isn't really the issue. The real issue is how much computing power you're getting for a given watt of power?
    • Didn't somebody from Google recently say that server density isn't really the issue. The real issue is how much computing power you're getting for a given watt of power?

      The real issue is both. SGI provides 16 CPUs with a nutty high performance interconnect in just 4U of rack space. Each of the current CPUs consumes something along the lines of 12 watts. There are also the various chipset ASICs, the RAM, drives etc...

      It does add up, so don't stick it in your bedroom closet... but it beats the pants off a P4 (or even P3) blade system for performance-per-watt. Especially when you factor in I/O.
  • Mhz Muppets (Score:4, Informative)

    by Tomah4wk (553503) <tb100&doc,ic,ac,uk> on Friday December 27, 2002 @03:56AM (#4965055) Homepage
    Ok, ive read at least 3 posts on this so i feel i must explaing Mhz. Many idiots will say 'i have a 2 ghz p4 so mine is faster'. WRONG. For example, lets assume a p4 takes about 15 clock cycles for an average instruction (thats made up by the way), this means it can do 2/15 = 0.13G instructions per second. Now lets assume the mips cpu takes 1 cycle per instruction (again, made up figure). This means it can do 0.7 G instructions per second. Alot faster you see...
    Disclaimer: I know ive ignored how much work can be done in an instruction, pipelineing and other features, but im sick of all this idiotic posts that think mhz is anything but a meaningless indication of processor speed, like a bogomip :)
    • Great. You ignore pipelining and so come up with a completely irrelevant post.

      No modern processor operates without a pipeline. It may take 20 cycles for an instruction to complete on a P4, but the P4 will issue another 20 instructions (more, in fact, because the P4 is superscalar) in the meantime.

      Feel free to discuss actual relevant points, like issue width, number of functional units, instruction latency and dependencies, cache misses, branch mis-predictions, bus bandwidth, cache and memory latency...
      • Relevant: The R16000 is also superscalar, don't forget.

        Also, the cache size on the R16000 is over 10 times greater than that of the newest Athlon XP, the memory access speed is faster, and IIRC, the bus bandwidth is greater. Kudos though on disspelling bullshit on the pro-MIPS side, though there's a lot more in favor of P4s and such.

        One thing I did find out from the AMD website (yes, i *DO* do my homework before arguing) is that the AthlonXPs have some kind of instruction pre-fetch "lookahead buffer". That sounds pretty cool, and would probably see the Athlon XP to victory in terms of instruction latency and dependencies. Or at least the former, the buffer *is* just a buffer, not a dependency checker.
    • You are incorrect.

      When we say the p4 has a 15 stage pipeline, that means that it takes 15 clock cycles for an instruction to go from start to finish.

      This does not mean we complete an instruction every 15 cycles. It's like an aseembly line with 15 stations. When instruction #1 reaches station number 2, station number 1 is free to start on the next instruction.

      So, for a single pipeline, we complete an instruction every cycle. The p4 can issue up to 6 internal operations, so the peak rate of a P4 2ghz is 12gigaflops.

      In practice however, we rarely see average performance anywhere near the peak. This is because not all instructions calculate in the same amount of time (divisions for example). And we can't always find enough work to fill the pipelines. Or we wait on memory to fill the cache.

      In any case, the R16000 was originally going to be a dual core chip multiprocessor, comparable to the Power4. It seems SGI abandoned this and instead did a shrink and cache enlargement on the R14k.

      The mips line are well designed to their economic situation and workload. Any statement along the lines of "my athlon roxor's it" or "my risc ownzers your lame intel" is overly simplistic. If you know how to use a computer, one would hope you could think a bit better than that.

      It's sad that many people in this thread have no idea what they're talking about. It's frightening that they still feel led to post this misinformation.
  • SGI is dying (Score:5, Interesting)

    by Tester (591) <olivier.crete@ocret e . ca> on Friday December 27, 2002 @04:19AM (#4965099) Homepage
    I worked all summer in an all-SGI shop.. And I call tell you how far behind they are. The place where I work is specialized in HPC, so when they started in 1992, SGI was probably a pretty good choice, but now for workstation, I wouldnt say its overkill, I would actually say that its underkill. We made a benchmark comparing an SGI Origin and a linux Ahtlon cluster, the athlon needed only two nodes to beat the origin and with all 16 nodes where about 10 times faster... SGIs are just overpriced, for 99.999% (that's 5 nines) PCs can do the job and even do it better and especially do it much cheaper. So their workstation market is being destroyed from under them.

    On the other end, their HPC (super-computers) is being attacked from above. On that sector, price is not really a problem, its just pure performance. And there too they are being beaten, SGI just does not have the research power that
    NEC or IBM can have. So they are starting to be pretty much behind, so they become not only more expensive (which does not really matter), but more importantly much slower...

    Also on the workstation market, their desktop SUCKS, its just a pain to use. They are still stuck in the pre-win95 era... It might have been good compared to win3.1 or twm, but it just is not in the same world as GNOME, KDE, WinXP or MacOSX.

    Also, their other strengh where there graphics board, they invented modern 3D hardware. And for a long time the roadmap for the PC 3d hardware was simple, they just had to do what SGI already had, but we have now passed a point where the PC hardware has actually more features then the SGI stuff. The only difference now between the pro and game markets are the amount of ram/cache and those "pro" cards exist on PCs. They do cost $ 2000-3000, but they are nowhere near the cost of the SGI workstation that includes them...

    SGI has no future. They have been losing money for years. I have been thinking for quite a while that they where a good target for an acquisition, but now that MSFT has bought much of their patents. It might be cheaper to wait for them to go bankrupt and to pick up the pieces. They where in a fast playing game and they have gotten slow.....
    • Re:SGI is dying (Score:2, Interesting)

      by scheveningen (305408)
      The only difference now between the pro and game markets are the amount of ram/cache and those "pro" cards exist on PCs.

      riiiight, unless you think E&S and Quantum3D are selling regular pc's, enlighten me on:
      - memory bandwith
      - dynamic resolution
      - genlocking
      - multi channel displays
      - hard real-time update rates
      - calligraphic lights

      Of course I won't choose SGI every time I need some graphics horse-power. But if you need to get a really big job done in real time, PCs don't cut it yet.
      • riiiight, unless you think E&S and Quantum3D are selling regular pc's, enlighten me on:
        - memory bandwith
        - dynamic resolution
        - genlocking
        - multi channel displays
        - hard real-time update rates
        - calligraphic lights


        I got most of that, but could you please explain 'calligraphic lights'?
    • BSD is dying too!

      Thanks for impressive comment as usual, comparing $3000 Wintel PC's with SGI etc.

      Hey, look they don't have SSE/MMX too! 700Mhz even!

      argh...

    • Re:SGI is dying (Score:2, Insightful)

      by Shinobi (19308)
      "We made a benchmark comparing an SGI Origin and a linux Ahtlon cluster, the athlon needed only two nodes to beat the origin and with all 16 nodes where about 10 times faster... SGIs are just overpriced, for 99.999% (that's 5 nines) PCs can do the job and even do it better and especially do it much cheaper."

      Then you're running small tasks that require little memory, little I/O and don't use much cache, and a substandard compiler. I've got a particle simulation going right now, the Origin 300 with 2 R14000A@600Mhz and 2MB L2 cache and 4GB RAM, using MipsPro compiler, that I have access to outperforms the dual Xeon 1.9GHz with 512kB L2 cache using both VS and Intel's own compiler. The difference in time is measured in days. It's the same thing with a cluster of athlons(And if you run a task where the task isn't easily parallellized, and need to keep in synch with the others, a node crash might ruin a lot of work and force you to start over)
    • Maybe so, but much of the graphics technology we take for granted was pioneered by SGI. Their engineers and technology built 3DFX and Nvidia and revolutionized the PC industry. Silicon Graphics can still easily build systems to run circles around PCs for real-time graphics. Their buses, memory and design are usually far more efficient. We'll forget they ever made the O2, for now. But they could never compete with the PC price/performance ratio. Cheap PCs have mass bandwidth, but are really inefficient, buggy and hard to scale (GB+ clusters and the right software might strongarm even this obstacle). Although I bet it is extremely difficult to outclass SGI's supercomputers. Perhaps IBM or NEC have a chance, but any cluster of PCs would be laughed at today.
    • > for 99.999% (that's 5 nines) PCs can do the job
      > and even do it better and especially do it much
      > cheaper.

      I don't buy it. With ix86 PCs, it's not just the software that's crap compared to legitimate enterprise solutions, but the hardware too. Linux is nifty and all, but it only improves the software side. The hardware is still shit.

      I've used ix86 boxes from most every builder... from solidly well-built IBM machines, to crap boxes built by dell from commodity parts. Not a one of them has achieved five nines. Remember, that's only five and a quarter minutes of downtime PER YEAR. With most OSs, if you reboot two or three times, that eats up all of your downtime right there, assuming NO other problems.

      ix86 boxes just are NOT up to the "five nines" standard. OTOH, I've seen more than a few Sparc, SGI, and RS6000s that can do it.

      Remember... just because you CAN do something on the cheap with crap hardware doesent mean that you should. And it doesn't mean that enterprise hardware doesn't have its place.

      cya,
      john
      • Right, except that you're still oversimplifying.
        First, five 9s does not make the mistakes that three 9s makes.
        Second, five 9s recovers smoothly from the mistakes that three 9s makes.
        Third, five 9s does not let errors go by unnoticed like three 9s does.
        Fourth, five 9s has a much more critical sense of what constitutes an error.
        Fifth, when something does break, it has to be fixed. That counts.

        Misquote from Dijkstra: "A baby crawling and a jet plane from JFK to LAX are both means of transportation".
        Probably stays up 100%. That's what? About -1 or -2 9s, methinks.

  • by wayne606 (211893) on Friday December 27, 2002 @04:27AM (#4965120)
    Instead of everybody saying "GHz doesn't matter, dummy" why doesn't somebody quote some real benchmarks? I poked around on the web a bit and all the benchmarks I can find either (1) are out of date, or (2) show Alpha, Intel and AMD blowing everybody else out of the water.

    In my experience SGI's are slow but are extremely scalable. With IA32-based machines you'd be lucky to get 4 CPU's sharing memory, unlike the 64+ you get from SGI. Very good for scientific codes but not so hot for applications that are either not parallelizable at all, or embarassingly parallelizable such as Seti@Home or ray-tracing a feature film.
  • SPEC (Score:4, Informative)

    by Kiev() (592438) on Friday December 27, 2002 @04:47AM (#4965151)
    Before evryone assumes that this thing is fast here some numbers to keep in mind:

    OK there are no numbers for 16K but here the numbers for 600Mhz 14K
    SPECint2000 500
    SPECfp2000 529
    For comparison

    UltraSPARC III Cu 1.015GHz
    SPECint2000 576
    SPECfp2000 775

    AMD XP 2800
    SPECint2000 913
    SPECfp2000 843

    INTEL P4 2.8
    SPECint2000 1040
    SPECfp2000 1048
    • A valid point. Although this chip is really designed for SMP architectures....and most of the applications that tend to run on SGI take good advanatage of that.

      You can't build large node MP AMD or Intel machines, period. So it's something of a moot point.

      -psy
    • The R16000 has many significant architectural and memory-related improvements over the R14000A. However, you are correct in that it's not the speed demon that some folks are making it out to be.

      But... keep in mind that it consumes far less than 20 watts of energy (and thus gives off little heat) and will eventually find itself packed in with other CPUs into Origin servers/supercomputers. The CPU bricks for the O3900, for example, have 16 CPUs in just 4U of rack space.

      SGI's ccNUMA MIPS/IRIX machines are typically used for tasks that are severely I/O bound, that is, their strong point is chugging thru massive amounts of data where raw per-node CPU power is important, but not the largest factor. Somewhat like a mainframe, but with less redundancy and more CPU power.
    • This is definitely not a flattering benchmark for the R14k, earning roughly half the scores of the Pentium 4 2.8 GHz, but when you remind yourself that the R14k is running at only 1/4 the clock speed of the P4, it reinforces the point that clock speed isn't everything.
  • competition (Score:2, Funny)

    by nEoN nOoDlE (27594)
    Looks like China has some serious competition on the chip front. SGI is already reaching a whole 700MHz! The Dragon better catch up soon is China wants to stay in business.
    • That's the great thing about socialism! You don't have to have a sound business model to suceed. If you can't pay your own way, you'll get bailed out.

      That's so very different than the capitolist companies like the US Airline industry and software companies such as Mandrake, where they make moey based solely on their commerical success... *cough*
  • by Emil Brink (69213) on Friday December 27, 2002 @05:07AM (#4965186) Homepage

    "[...] the 48-bit RGBA provides the highest level of precision available on any desktop system today"

    Oh? Quick, everyone with Radeon 9700 PRO graphics boards in your PCs, make sure you have them in tower cases, or something!

    For reference, the ATI specs page states:

    Pixel shaders up to 160 instructions with 128-bit floating point precision

    I guess SGI might refer to actual output precision, i.e. the RAMDAC D/A-converters... In that case, it seems they still have the edge, since the ATI boards only have 10 bits per component. Still, I think that's of lesser value than the actual precision image operations are performed at.
    • That's 16-bits per colour in the RGB plane, for a total of 48-bit colour. Most PC cards do 24bit (or 32bit with alpha/Z plane).

      Not the width of the GPU!!!!

      You're confusing apples and oranges.

      -psy
      • and iirc you windows can't officially have higher than 32bit(24bit+8bit alpha/wasted/whatever) currently.

        the matrox card that does 48(? or was it 42), is actually using a dirty hack of some sort to get that depth on windows..
        • It's 11 bits for blue, 11 bits for green, and 10 bits for red. People don't like red as much as the other colors.
          • well, during my time coding gfx effects, from my dos days with vesa modes(2.0 vbe was sweet)...
            i have never enountered such mode in actual use.

            the 8 bits got wasted in every mode available on any card(that had any sense to use)..
    • "...output precision... of lesser value than the actual precision image operations are performed at."

      Not true if you're doing real imaging work. How about that fancy LCD monitor you've been eyeballing (or just picked up)? Noticed any of the color problems, especially with dark shades? No?

      Then you aren't doing graphics work that needs the display accuracy of an SGI or equivalent.

    • "[...] the 48-bit RGBA provides the highest level of precision available on any desktop system today"
      Oh? Quick, everyone with Radeon 9700 PRO graphics boards in your PCs, make sure you have them in tower cases, or something! For reference, the ATI specs page states:
      Pixel shaders up to 160 instructions with 128-bit floating point precision
      You're comparing apples and oranges here. SGIs can handle 48bpp (12 bit per channel RGBA) integer colours all the way through the pipeline. The ATI card you mention, as well as nVidia's new chip, can use 128 bit floating point numbers for some internal calculations, but that's about it. The only point of this is that you have less round-off error if you run a complicated shader program, plus you get a bigger number to put on the box :) Like you say, ATI can also do 10:10:10:2 RGBA integer mode, but that's pretty useless for most people looking for this sort of thing.

      Bottom line, if you need high precision integer colours, you still need an SGI. Of course, there's not many people who do, and someone will probably be doing it on the PC in a couple of years, so it's looking pretty grim for SGI as that's one of their few remaining technical advantages in the graphics workstation market.

      • I think you're underestimating ATI's and nVidia's latest generations. If you go and read this technical document [nvidia.com] (PDF, ~4 MB) on the nVidia GeForce FX, you'll see that it does, in fact, support full 128-bit floating point all the way up to, and including, the frame buffer. I'm not certain why 12 bits of integer should be better than 32 bits of floating point, per component, maybe you could educate me a bit there? Also, I don't know what the actual analog output quality of the FX chipset is, maybe SGI has a (small) lead there, still.
    • OMG It happened!

      Finally someone compared a desktop (not pro, I know some ati pro level cards as matrox ones) Gfx card, optimized for gaming... With... SGI...

      Yes, folks this is history. /me dies
  • by qwijibrumm (559350) on Friday December 27, 2002 @05:15AM (#4965207)
    The R16000 has Out-of-order instruction execution? Sweet! So what was SGI's plan when they made this?
    1.???
    2.Profit!
    3.Build new processor.
  • BUT (Score:3, Informative)

    by katalyst (618126) on Friday December 27, 2002 @06:39AM (#4965344) Homepage
    SGI is still in trouble. I love the company, their concepts, their hot rod machines and the supercool names they come out with..... BUT they are in trouble. And Linux is one of the primary reasons for SGI getting into trouble. A large number of design studios seem to have succumbed to the temptation of a cheaper yet stable machine (i.e Linux Boxes). As some other slashdotters pointed out, these guys are using Macs for the artwork and Linux boxes for the actual bull work. I wonder if SGI can reconquer their old customers and charm even more people.
  • It's not the run-single-task-as-fast-as-possible. Intel/AMD rule that part, with Intel up a slight lead.

    It's not the very parallell computing, like movie rendering. Clusters, usually Linux clusters do much better compared to cost.

    It's not most kinds of servers, that are usually IO bound and it's the disks, controller, NIC and mobo (backplane) that make the server. Few of those need more than dual MP cpus to do well.

    I know roughly where the SGIs still shine. But how many really have those specific needs? Not many that I can think of.

    Kjella
  • "SGI has just launched a new CPU - the long expected R16000."

    Isn't it MIPS that make the CPUs?
    (This is not sarcasm, I really wanna know.)

  • die size (Score:3, Informative)

    by vlad_petric (94134) on Friday December 27, 2002 @11:22AM (#4966206) Homepage
    4MB L2 cache => *huge* die => low yield => huge cost. Yeap, it's that simple.
  • "700MHz, has 4MB cache"

    Hey, according to intel, this processor, at 700MHz, is about 4 years old, and has no hope of competing with intel's True MHz processors! :-P

I cannot conceive that anybody will require multiplications at the rate of 40,000 or even 4,000 per hour ... -- F. H. Wales (1936)

Working...