AMD Announces FirePro W9100

by Ryan Smith on 3/26/2014 1:00 PM EST
Comments Locked

31 Comments

Back to Article

  • Ian Cutress - Wednesday, March 26, 2014 - link

    Time for a new card for that 6x 4K HTPC (!). Seriously though, the GDDR5 chips to make a 16GB card are not commonplace and probably made by one company. That means 16 Gbit chips to make a 512-bit bus. Sapphire just put 8 GB on the R9 290X, so here goes AMD with one better. I wonder if clock speeds have to come down because of that. Either way, 1/2 DP with the larger memory buffer are big pluses. I hope Supermicro or someone has a full Ultra Workstation that I can look at during Computex.
  • ddriver - Wednesday, March 26, 2014 - link

    Clock speed will likely not be as nigh as gaming series, but who cares, with 16 gigs of ram on board, snap 4 of those on a workstation mobo with 64 gigs of ram and you got yourself a mini supercomputer at home. Build a few, get a good 10 gbit switch and you got yourself a computer farm cluster.
  • mczak - Wednesday, March 26, 2014 - link

    You are very wrong about the gddr5 chips required not being common. Nowadays just about everyone produces (and graphic card manufacturers are using, ps4 too even) 4 gbit chips (but 2 gbit chips are still used too on some cards, most notably on 2GB 256bit cards). Since the gddr5 chips are 32bit wide, you require 16 gddr5 chips - 16 4 gbit chips give 8GB of memory. But gddr5 can use a so called clamshell mode, where you use 2 chips in a 2x16bit mode to make a 32bit wide channel. So, for a 16GB card, you simply use 32 4gbit chips in clamshell mode. (nvidia though of course due to only having a 384bit bus, is restricted to 12GB.)
  • mczak - Wednesday, March 26, 2014 - link

    That should have been "require 16 gddr5 chips for a 512bit bus".
  • Kevin G - Wednesday, March 26, 2014 - link

    Couldn't they use two 8 GBit chips on the same bus? Some previous workstation cards did so with GDDR5 to increase capacity.
  • yasamoka - Thursday, April 3, 2014 - link

    You don't have to use 8 chips for a 512-bit = 8x64-bit memory bus. The 7970 uses 12 chips for a 384-bit bus, so this could be using 16 8Gbit chips.
  • extide - Wednesday, March 26, 2014 - link

    This would be interesting, and a bit disappointing if it is true that the W9100 has 1/2 FP64 where as R9 290(x) have 1/8 FP64. Up until now, the consumer Radeon cards were never cut down and always had full performance FP64, compared to the FirePro versions. Unfortunately it seems they are probably following nVidia's lead here, and cutting the FP64 capabilities on their consumer lines, like nVidia does on Geforce vs Quadro/Tesla.
  • coder543 - Wednesday, March 26, 2014 - link

    I don't know if AMD is making Radeon cards 'cut down' in this case. It would be nice to hear the technical side of the matter, but it could be that AMD is simply binning the chips and only a small number can sustain the clock frequencies necessary to have 1/2 FP64, or that they're upgrading the chip design for the W9100 so that it can do 1/2 FP64. It's possible they're simply releasing artificially limited chips like nVidia does, but that isn't the feeling I'm getting here.*

    *Feelings can be wrong.
  • SaberKOG91 - Wednesday, March 26, 2014 - link

    The way I understand it, the floating point units on the consumer versions have a different pipeline width altogether. The pipeline is narrower in favor of higher FP32 throughput, but that makes the double precision take longer to process. This also allows more room for integer units if I'm not mistaken (thus the increased mining capacity of desktop parts).

    On the workstation grade parts, the integer side is reduced and the floating point width is increased in favor of the faster FP64 performance needed for HPC and 3D modelling work-flows. This increased floating point performance, coupled with application certifications, has always made these cards more expensive than the consumer parts with the same core count.
  • twoodrow - Sunday, March 30, 2014 - link

    Where did you come across this information on the architecture?
  • g101 - Wednesday, March 26, 2014 - link

    It's true, there may be a technical/cost reason for the higher ratio in the regular desktop cards... Hard to say for sure.
  • g101 - Wednesday, March 26, 2014 - link

    The difference in ratio has been confirmed. The reason is unclear, though I agree it's most likely the same reasoning that is behind nvidia's decisions (though nvidia's ratio is much higher (which is worse for the consumer)). This appears to be a more moderate but still marketing-driven decision.
  • JDG1980 - Wednesday, March 26, 2014 - link

    Will there be custom AIB versions of this card? We already know from the existing Hawaii cards that that cooler just isn't going to cut it. I doubt that people doing intricate construction on SolidWorks or whatever want their workstation roaring like a jet engine.
  • WithoutWeakness - Wednesday, March 26, 2014 - link

    They also don't want to pay for a card that is going to continually throttle itself down to minimum clock speeds. I really don't like AMD's new strategy of listing the maximum boost speeds of their cards as the GPU's clock speed when their cards are unable to maintain that speed with the stock cooler. AMD will either need a redesigned stock cooler or a third-party cooling solution if this card is going to do anything different than the R9 290X does - ramp up to 95C and then ramp up the fan speed and throttle clock speed to stay within thermal requirements.
  • Flunk - Wednesday, March 26, 2014 - link

    Based on previous experience they'll all be built by Sapphire and AMD-branded. The stock cooler in the image above will be on all cards. The people who buy these either don't care how loud they are or have taken other measures to quiet down the computer (like sound-proof boxes or locating the box in a server room).
  • JDG1980 - Wednesday, March 26, 2014 - link

    These are workstation cards, not server cards. Noise is definitely a factor. A good case with noise dampening can help a little, but not fix the problem completely - and many of these will be in mediocre OEM cases, not a nice Fractal Design R4.

    I don't understand why AMD didn't just ask Sapphire to use their Tri-X cooler on the cards - we know it works very well on Hawaii, far better than the outdated stock blower.
  • NicoleMWilson - Friday, March 28, 2014 - link

    While I do 3930k and 4930k and ECC memory, they opt for off the shelf consumer cards. Some of these firms have contracts with Ford, Chrysler, Nascar and many other companies. The lure for them of course is price. They are comfortable paying $3500 a machine vs $5500 - $8000. I've tried to push several of my clients toward the AMD FirePro W7000 as I can get these cards for around $500 but they still pass. http://num.to/4871-8450-4883
  • birdiemaker227 - Wednesday, March 26, 2014 - link

    Just want to point out to extide that 1/2 rate double precision is much, much larger than 1/8, 1/4 or 1/3. The Radeon cards have been limited on FP64, and for FirePro at 1/3 until now. With this new card from AMD the full potential can be realized at 1/2 rate. It means you get exactly half of the single preicsion rate now, not 1/3 or less.
  • eanazag - Wednesday, March 26, 2014 - link

    The margins are clearly different, so i would expect a beefed up ~$4000 version of the consumer card to have a good cooler. Question is: does it stack up against the Titan Z well? I expect it to be more from a cost perspective.

    I have no use for this today, I can't afford all the 4K displays this card can run. The specs are crazy. I'd hate to be the PCIe bus on the 4 card workstation.
  • HigherState - Wednesday, March 26, 2014 - link

    ......or in the same room!!!! Four cards running full tilt would be ear-splitting.
  • ddriver - Wednesday, March 26, 2014 - link

    I am actually very excited about this product. I like AMDs strategy to offer significantly better DP performance for the buck compared to nvidia. This is my sole purchase decision since all I need the GPUs for is OpenCL and DP data crunching.
  • g101 - Wednesday, March 26, 2014 - link

    Heh, funny how both of our needs are met by this one product.

    I need single precision and actual opencl support like you (rather than the nonsense nvidia offers).

    Luckily, I can get away with the desktop AMD gpus since I don't need ECC or the display port optimizations/driver validation of the firepro line.

    However, the real difference/attraction for me with Firepros are the frame buffers...in this case, 16GB per die would be incredibly useful...Might have to bite the bullet due to that alone!
  • extide - Wednesday, March 26, 2014 - link

    If you can live with 8GB ther are 290X's out there. (Sapphire)
  • sascha - Wednesday, March 26, 2014 - link

    Me too, AMD has great GPUs! I just wish OpenCL would evolve faster and their FirePro line would be cheaper but on the other hand why should they do me the latter favor...
  • MrSpadge - Wednesday, March 26, 2014 - link

    Just use a proper cooler and the chip temperature is nothing more to worry about than for any other 200+ W high end GPU.
  • blanarahul - Wednesday, March 26, 2014 - link

    So, AMD one upped even NVIDIA with FP64 performance. Boy am I impressed.
  • tomhefley - Wednesday, March 26, 2014 - link

    I actually build machines in my area for engineering firms who use Solidworks. None of them opt for these high end cards. They are just too expensive. While I do 3930k and 4930k and ECC memory, they opt for off the shelf consumer cards. Some of these firms have contracts with Ford, Chrysler, Nascar and many other companies. The lure for them of course is price. They are comfortable paying $3500 a machine vs $5500 - $8000. I've tried to push several of my clients toward the AMD FirePro W7000 as I can get these cards for around $500 but they still pass.
  • extide - Wednesday, March 26, 2014 - link

    Yeah Solidworks guys aren't the target for these.
  • g101 - Wednesday, March 26, 2014 - link

    Why are you linking to some game on amazon and incapable of understanding thermal targets?

    Ban that 'EduardoHRami' idiot.

    Anyways, yes it does appear AMD took a marketing note from nvidia with regards to regular desktop Hawaii and the 1/8th double precision.

    Still far more generous than nvidia's ratio, though I would like to see AMD go for the throat and release firepro display port optimization and full double precision across all their cards...That would be a serious shakeup.

    Still, this is one hell of a single-die GPU.
  • ac1dra1n - Thursday, March 27, 2014 - link

    I would be willing to bet that the W9100 will have a 512 bit bus. Is there anything technical that would argue against that?
  • sanaris - Saturday, April 12, 2014 - link

    I am using right now FireStream 9350 (175W) which claims to be 400 gigaFlops. It gives 24 msec time in Lucas-Lehmer FFT.
    For comparison, 2 processors of AMD (300W) are MUCH better because they give 7 msec out of 160 gigaFlops.
    Conclusions: (1) the so called "pro" cards lie to you when they say about Flops, (2) buy real processors which does not fool you with false speed numbers.

Log in

Don't have an account? Sign up now