   |
|||
|
|||
|
:
News
: Archives
|
|
|
|
|
|
|||
|
|
|||
|
|
|||
|
October 10, 2005 News of Interest - Josh A couple of AMD type announcements have popped up today, and one of them is pretty big for them. Supermicro is now officially offering a wide variety of Opteron based platforms. These include barebone servers as well as standalone motherboards. Supermicro has been an Intel only platform provider for ages, and last year they introduced their first Opteron based products. The only catch was that they didn't advertise about them. Now that AMD has the lawsuit against Intel, many of these "Intel only" partners appear to have a lot more freedom in how they operate when it comes to advertising about their AMD based products. I still vividly remember the first Comdex after the Athlon was released, and not one major motherboard provider publicly showed their Athlon based motherboards. One told me that Intel people were walking around and they feared their company might suffer from "allocation problems" if they were showing AMD based solutions. They still had the boards available for private screening, but you had to be shown to a back room and ask for it specifically. The other interesting release is that of the Tyan SLI based motherboard. With NVIDIA really pushing their SLI products to the mainstream, Tyan has introduced a lower end SLI motherboard that can double as a workstation or server product. I have finally finished testing on Tyan's nForce 4 Ultra product, and I must admit that I am very impressed with it. Not only was it just rock solid as well as speedy, but it turned out to be a surprising overclocker (usually Tyan and overclocking are not seen together). A very high quality product, but it is very pricey for a mainstream board. There will be plenty more to look at this week, so keep tuned! October 6, 2005 X1000 Hangover - Josh If you are like me, you have overdosed on reviews and previews on the X1x00 series of products from ATI. Overall I am impressed by this release, as the technology is quite compelling and very competitive. They do bring some nice features to the table, such as AA support in HDR rendering, as well as a high quality anisotropic filtering setting (though it is not enabled by default). The issue that everyone is pounding right now is availability. Yes, there were products promised to be available upon release, but it looks to be at least a week before actual boards become available. Once they do become available, I wonder how long supplies will last and if ATI and its partners can keep up with demand. There is no question that the high end cards compete with what NVIDIA has (though pricing could be a big sticking point at this time). The X1800 XL competes well with the 7800 GT, and the monster X1800 XT competes quite nicely with the 7800 GTX (and it has 512 MB on board to boot). The thing NVIDIA of course has in its favor is a solid multi-GPU system in place, and now that SLI motherboards can be had for far less than $150, they are becoming favorites in retail and OEM circles alike (extra flexibility for not a whole lot of scratch). Something else to consider is that NVIDIA is going to have a 512 MB 7800 part very soon. I have received no official confirmation of this, but it is pretty widely known that it will be showing up sooner rather than later. NVIDIA has had a lot of flexibility with their product releases due to ATI not being able to hit their roadmap. From what I gather, NVIDIA explored a lot of ideas behind higher clocked 7800 parts that would utilize two slot coolers (of which the Leadtek 7800 GTX Extreme uses), as well as possibly having an external power supply (think Voodoo 5 6000). When NVIDIA learned of ATI's problems with the R520, and what kind of performance the initial samples had, they kept the 7800 GTX as the high end offering and kept cooling and power to a minimum. This turned out to be the correct path for NVIDIA as they sold a lot of these products for a very nice price (not to mention the sweet margins NV and its partners saw with every board sold). NV and its partners made some huge inroads with OEM's due to the single slot 7800 GTX, not to mention its conservative power and heat envelope. So what can we expect from this 512 MB card? My guess is that it will indeed adopt the two slot cooler that was already developed for the "Ultra" version. Not only that, but I think that NV has probably been putting aside faster binned G70 chips to use with this version. So, these cards will most likely be clocked a lot more aggressively than the standard 7800 GTX, and I would assume that we could see products hitting 470 MHz and above. These will be high priced products, but it is in NVIDIA's best interest to have a faster core speed with this part than the standard GTX. My guess is that the core will run at 500 MHz even, but quite a bit of this will depend on what kind of power the 512 MB modules will consume. We will not see an external power supply for this product, but it could have quite a bit more power draw than many are hoping. Still, NVIDIA is not looking to relinquish their performance crown, and the 512 MB 7800 will be a pricey card to own. The one negative that I saw with the X1x00 release was that of the X1600 series performance. With 12 pixel pipelines I was expecting this part to walk all over the 6600 GT and 6800 cards. That was not the case. In most situations it appeared as though the X1600 XT was only slightly faster than the 6600 GT, and rarely could it overcome the regular 6800. Hopefully ATI will really work on the drivers for this architecture. The X1600 XT features 8 ROPS, 12 pixel pipelines, and 8 vertex shaders all running at 590 MHz. With those kinds of numbers, one would expect this product to in fact walk all over the 6800 GT! This is not the case. I am not sure why the performance is as disappointing as it is, but perhaps it just can't get over the 128 bit memory bus that it is shackled with. Bandwidth really is king here, and both the 6800's feature a 256 bit bus and memory that runs as fast (in the case of the 6800) if not faster (the 6800 GT runs at 500 MHz DDR). Looking at this, it is not surprising that NVIDIA has not released new products in the midrange and budget sectors. Currently the sub $300 6800 GT is really the card of choice for people wishing to run SM 3.0 games and HDR, and the 6800 and 6600 series really flesh out the area below $240. I must admit, I am very curious as to what NV will do with its upcoming G70 based midrange products. Hopefully these will not feature 128 bit memory busses running slow memory. One thing to consider though is that these X1300 and X1600 products for the low end and midrange do an excellent job of replacing the X300/X600/X700 series of products! October 5, 2005 Radeon X1000 Day - Josh ATI has finally taken off the wraps of their X1300, X1600, and X1800 products. The reviews I have seen so far are generally very positive towards these products, as they are quite competitive with NVIDIA's latest offerings. The X1800 XT usually comes out ahead of the 7800 GTX in Direct3D applications, but of course falls behind when it comes to OpenGL/Doom 3. ATI brings a lot of good stuff to the table, and it truly is a next generation part. It does have features above that of NVIDIA's latest, and it appears that it can handle SM 3.0 complex shaders quite a bit better. The R520 is supposed to be 321 million transistors strong, but of course only has 16 pixel pipelines. ATI has done a lot to the architecture though, and the increase in clockspeed also helps to offset the pipeline advantage NVIDIA has with the 7800 series. The X1800 XL is around the speed of the 7800 GT, but there will be a pretty hefty price difference once product finally hits the street. Many are expecting around a $100 difference between the two, and that might turn off more than a few customers. Still, it is good to have competition out there. The X1800 XT is much more in line with the 7800 GTX. For starters at $549 it features 512 MB of fast GDDR-3, while the 7800 GTX only has 256 MB. 7800 GTX's are already below $499, and that includes some of the overclocked versions which will provide better competition to the X1800 XT. Apparently NVIDIA is going to release a 512 MB version of the 7800 GTX in short order, most likely we will see that card released before we see general availability of the X1800 XT. ATI is promising availability of the X1800 XL and the X1300's immediately, but so far no retailer has put up any prices or products as of yet. Hopefully ATI isn't fibbing us on this one, and we will see some cards available by the end of this week. The X1800 XT is a different story altogether. If you had read my CrossFire editorial, I had pointed out that ATI had made a big R520 order from TSMC before they found the soft ground issue. This big order would come out with slower clocked chips. These chips will be integrated into the X1800 XL line. After the fix was found, ATI raced to get the design into manufacturing, but apparently it wasn't until late July or so, and only now is ATI seeing good chips coming out of the fab and being packaged. It will take a while to build up stores of this product as well as get it into partner's hands. ATI is expecting X1800 XT availability sometime mid next month due to this situation. On the other side, NVIDIA decided to try to throw a wrench into this release by delivering the Forceware 80 series of drivers (81.84 is available as of now). These do improve some performance aspects of the 7800 series as well as new SLI features and functionality. Few reviews so far have used the 81.xx series of drivers, but they do show a marked increase in some applications. Much to NVIDIA's favor is that they offered this driver to end users the same day they released benchmarks, unlike the now infamous Detonator release with the Radeon 8500 where the driver wasn't released until about two months after reviewers had them. NVIDIA is also improving their PureVideo support in a variety of applications, which is a huge plus for those watching DVD's and videos on their computer. October 4, 2005 Beating the Dead Horse - ATI's CrossFire - Josh Scott, Kyle, Ryan, Charlie, and others around the industry have written editorials on the supposed availability of the CrossFire products from ATI. While it is true that ATI PR did in fact tell these journalists that CrossFire would be available to the market shortly after launch, and there are in fact no products available, I think all parties involved are looking at this the wrong way. ATI PR made a mistake, but these journalists haven't dug deep enough into the reasons behind PR's behavior. I try to look at both situations and have tried to understand why ATI made the decisions they did. Here is a quote:
You can read the entire editorial here. September 30, 2005 Final Thoughts on CrossFire - Josh Now that the dust has settled on this subject, and we have all had a lot of time to really consider the implications and possibilities of ATI's CrossFire, I thought I would summarize my thoughts and concerns about this product. Some folks like Sander think CrossFire is a failure, but I do not necessarily share that opinion. While I do not think CrossFire is a success in its current configuration, I do think that it is a step in the right direction. We really should think of this iteration of CrossFire as a technology demonstration. As a demonstration, I think it did its job. Yes, it currently has the nasty limitations that everyone has talked about (resolution/refresh, no supertiling with 3 quad cards, application settings determined by Catalyst AI, etc.), but there is something that many are probably missing here. I do not think that ATI expects this to be a great selling product in its current form. I doubt that there will be many CrossFire master cards on the market for the X8x0 generation of cards. ATI will be selling plenty of the CrossFire enabled motherboards, but most consumers will simply pass the CrossFire master cards by. If they want a multi-card system, they would be silly at this point to not use NVIDIA's SLI (which is quite mature, works well, and is starting to be supported natively by gaming apps). So, ATI has laid the groundwork for CrossFire with their X8x0 series of cards, and again I think that everyone involved in the project has considered this a demonstration. Yes, ATI will sell more than a handful of CrossFire master cards for this generation, but it seems to me that ATI is really just revving up the technology for the next generation of cards. Next week we should see the wraps taken off the R5x0 series of cards, and we will be getting more information on how those work, as well as the future iterations of CrossFire. My guess (and hope) is that the next gen of CrossFire cards will be a lot more robust than what we see now. While the DVI output on the slave card is a definite weakness, I think I have shown mathematically below that it does not necessarily have to be the patsy for the entire operation. I think the main weakness for the current CrossFire boards is the use of the Xilinx FPGA which they use as the compositor chip. This chip has no significant amount of memory inside it, and what is there cannot be used to perform buffer activities. If the compositing chip had at least 32 MB of memory, either onboard or external, it could help to overcome many of the issues and limitations that we currently see on CrossFire. One would think that ATI would be creating their own custom ASIC for this instead of using an off-the-shelf FPGA that might not fit all their needs. So, pronouncing CrossFire dead and too late at this point is somewhat pointless, as another generation of the technology is about to be released. Things will improve, and if ATI does in fact have a custom ASIC doing the compositing work, then we could expect to see a good performance increase as well as some more usable features. As it is, due to the lack of a frame buffer, Supertiling and SuperAA are actually done through the PCI-E bus (and not the dongle) and these are actually composited on the master card's GPU (and only pass through the compositing FPGA). I guess we shall soon see! EPoX Brings Happiness to Socket 754 Owners The real difference between a single channel Athlon 64 and a dual channel chip is actually quite small. Considering this, as well as the many buyers of Socket 754 chips, EPoX is putting out a Socket 754 SLI board. I personally think that this is a great idea and they will sell plenty of these boards. I know of many people who bought their 754 processors for $300 and more (the S754 3700+ was a very expensive chip when it was first released), and they want to keep their investment going as long as possible. A S754 3700+ is only about 5% slower overall than a S939 4000+, so it makes sense to keep this processor running as long as possible considering the money spent on it. Wanting to upgrade to PCI-E or possibly a SLI setup is a good option for many, and it is great that EPoX sees the market there for such a product. So, for those not wanting to give up their previously pricey 754 chips, take a look at this monster here. September 29, 2005 NVIDIA GeForce Go 7800 GTX - Josh It seems that NVIDIA just can't resist releasing new products these days. With the overwhelming success of their 7800 desktop series, NVIDIA is shipping their slightly lower power mobile 7800. The first product is the Go 7800 GTX, which is nearly identical to the desktop version, except it is clocked down at 400 MHz. As you may remember, the Go 6800 Ultra was actually comprised of 12 pixel pipelines running at 450 MHz (12 PS/5 VS/12 ROPS). This time the Go part is identical to the desktop. The Go 7800 GTX features a full 24 PS/8 VS/16 ROPS. Only 30 MHz (400 MHz core) and slower memory (550 MHz GDDR-3) differentiate this product from the desktop high end. The best part of this release is that product is now available from a variety of OEM's and ODM's. NVIDIA gave out some benchmark results, and I would say that they are well in line with what consumers will see. The Go 7800 GTX outstrips the older 6800 Ultra by as little as 30% up to a pretty hefty 80%, depending on the application. In the past few years ATI has really surged in the mobile market with plenty of really solid chips, but NVIDIA has picked up the gauntlet and really focused on putting out very solid products for this expanding market. Apparently the power management system is quite flexible, applying to not only core and memory speeds, but by how many PCI-E lanes are in active use. The core can scale down to 16 MHz at max, and it can dynamically adjust from 16X PCI-E lanes to 1X. Overall, the enhanced memory scheme allows the Go 7800 GTX to keep the same power and thermal envelope as the previous Go 6800 Ultra. ATI is set to release their next generation products at the beginning of October, but they most likely will not be introducing a new mobile product at that time. ATI's X600 and X700 mobile series are doing very well, but their X800 mobiles are being pushed out by the NVIDIA products in that space. Hopefully soon we will see ATI heat up the competition in the high end mobile arena again! September 26, 2005 ATI Releases CrossFire - Josh The reviews are finally out, and some really interesting things are confirmed and learned. The 1600x1200 @ 60 Hz limitation is there for CRT's, as well as widescreen 1920x1200 @ 52 Hz. This is a heck of a limitation for owners of higher end LCD's and CRT's. There has obviously been a lot of discussion about this recently, and there has been a lot of finger pointing about the culprit. In my opinion, after looking over this information and these recent results, the main issue here is that the compositing chip does not have any onboard frame buffer that it has access to. So basically ATI has to clock the compositing chip at a certain rate to be able to handle the frame inputs from the two different cards and successfully display them as a coherent frame/scene. Another interesting aspect is that the Super AA modes are actually sent over the PCI-E bus to the master card, as compared to over the DVI link in the other rendering modes. This again looks to be because of the buffer limitation on the compositing chip. The Super AA'd frames need to be stored, combined, and then displayed. It seems that the ROPs on the master card are doing this work, and to do it effectively it needs to have an onboard frame buffer. One last limitation is that the 12 pipe cards cannot utilize the SuperTiling rendering mode. That being said... the performance of the CrossFire solution is quite good, and it keeps up with NVIDIA's SLI in most cases (especially when dealing with the 6800 series of cards in SLI). Once the 7800 series is used though, these tend to work a bit faster than ATI's solution. In terms of maturity, SLI has CrossFire beat hands down. Still, it is good that ATI has a solution that is at least competitive in many areas to SLI. Things will only improve from here on out with CrossFire. Overall I am impressed by the performance of CrossFire, but some of the results are a bit odd. Super AA of course takes a hit much like NVIDIA's SLI-AA, mainly due to the increased traffic over the PCI-E bus and the extra load the master card's ROPs have to handle. ATI promises that it will use the compositing chip for this functionality in the future, but have no definite time frame in which this is to be implemented. Obviously CrossFire is not perfect, but it does work. Hopefully with the R5x0 series of cards we will not see the same resolution/refresh limitations that this first iteration of CrossFire shows. There will be people buying these setups, and the flexibility of using different/older cards (as long as they are PCI-E) is certainly nice for people who already own X800/X850 cards. Something else to consider is that there is no physical reason why CrossFire shouldn't work on SLI boards, but it is a matter of licensing. While it is not a stellar release for ATI, it is definitely a solid one. The implementation is good, the performance is good, but users will be faced with the limitations the product does bring to the table. For users who typically game at resolutions under 1600x1200 and have a X8x0 series card, then CrossFire could be a very good decision for them. My feeling overall is that users should wait for the next generation chips and CrossFire setups before they spend money on such a solution. Here are a few reviews for your reading pleasure: One thing that I am somewhat curious about is that many of these sites are comparing the CrossFire setups to older 6800 GT/Ultra SLI boards. Once the current costs of the CrossFire boards and the recently released 7800 GT are compared, we can see some price parity there. I can understand the urge to compare the CrossFire boards to products from NVIDIA that were from the same generation, but a person considering buying a CrossFire setup for a certain price would be silly not to take a look at the 7800 GT's in SLI. September 23, 2005 Reviews of Interest - Josh MS takes a good, long look at the Athlon X2 3800+, and he comes to the conclusion that this is one of the best processors a consumer could buy. I am currently testing one of these as we speak, and I can see where MS is coming from! It is a well performing product, and it doesn't heat up your case like other dual core processors (plus it is pretty darn energy efficient). If you don't believe me, take a quick read of MS's review. Anand's kids take one of the first looks at a production GeForce 6100 based board. While this is not a barn burner, it does put in some very good numbers compared to the top of the line ATI integrated chipset. It looks as if NVIDIA did a fine job on this product, and we should start seeing retail/OEM 6100's and 6150's within the next few weeks. Dave B. at Beyond3D takes a good look at ATI's AVIVO technology that they will be introducing with their latest series of upcoming cards (apparently in the first week of October). Some good info there, as well as a good overview of what to expect. HardOCP reviews the Velocity Micro Gamer's Edge 1500. This is a solid gaming system, and it can perform well in a variety of applications. Remember, no "benchmarks" here. TSMC's 90 nm Process I must admit that I am very impressed with what TSMC has done with their 90 nm process. TSMC does not have the R&D resources that the big guys have when it comes to developing smaller process nodes, and they are usually nearly a generation behind the big boys (TSMC started 90 nm production early this summer, while Intel apparently is aggressively ramping their 65 nm production as we speak). Both ATI and NVIDIA have several products being produced on TSMC's 90 nm process as we speak, and NVIDIA was the first to introduce a 90 nm part with their GeForce 6100/6150 integrated products. TSMC is doing a couple of new things with this process, and the first is that they are using 300 mm wafers for all 90 nm production. This allows nearly double the amount of dies over a 200 mm wafer, but at what appears to be a lower overall cost per die. This is a big win for both TSMC and its partners, as more die can be produced for cheaper, allowing TSMC to make more money per chip, all the while the partners will pay slightly less per good die. The second is the use of a Low-K dielectric. In previous nodes Low-K was only available on certain lines. For example there was a 130 nm FSG line, while there also existed a more expensive 130 nm Low-K line. ATI was burned last year by requiring Low-K on many of their products, and TSMC just didn't have the production capabilities to get adequate numbers of parts into ATI's hands. TSMC's 110 nm process did not feature Low-K at all, as it was essentially an budget process based on an optical shrink of 130 nm FSG. TSMC's 90 nm process was apparently quite clean by the beginning of Spring 2005, but unfortunately for ATI their R520 design had some issues. Apparently ATI identified this as a "soft ground" problem, and it affected their RV530 chip as well. From what I gather, ATI was essentially forced to start large scale production of the R520 chip to be able to release a product by early October, and the fix was found sometime in early Summer. By the time the fix was implemented in the design, and a new revision was completed, it was already in July. TSMC's 90 nm process takes approximately 12 to 14 weeks to complete from start to finish, and then you have to add in the time to package the raw dies, and then deliver the final chip to board producers. Apparently the R520 release we expect to see at the beginning of October will be of products using the initial order of chips. The boards reviewed look to be X1800 XL's, which are clocked lower. When ATI finally gets the revised order chips coming in, we will see the higher clocked X1800's hit the market. ATI also appears to be on the verge of introducing new budget and midrange chips, but I don't know when we will expect to see these released. NVIDIA has at least two 90 nm chips in production right now. The aforementioned 6100/6150 chips, as well as what looks to be a midrange chip that should replace the current 6800's as well as push down the 6600's. Many expect these to be the 7600 series of chips. It makes a lot more sense for NVIDIA to release a faster clocked 12 pipeline chip based on the G7x architecture on 90 nm than to have the old (but proven) NV40/NV45 chips make up the $300 and below categories. From my understanding NVIDIA will continue with the 6200's and 6600's at the $149 range and below. These products still look to be competitive with what ATI has coming out, so there is no real reason to introduce new chips for this category. TSMC certainly appears to have done a bang-up job on their 90 nm Low-K process, and the graphics guys are certainly happy about it. September 22, 2005 Catching Up and NVIDIA's Latest Motherboard Products - Josh The motherboard on my main computer died, so I had to finish a couple of other projects before I could really replace it with the MSI K8N Neo4 Platinum/SLI. So far I am really liking this board a lot more than the Chaintech VNF4/Ultra that died (for obvious reasons). I have rarely heard of any problems with these boards from Chaintech, and BFG uses the same design (though some different components) for their nForce 4 Ultra motherboard. I received this board some months ago, and once things had calmed down this summer I started testing on it. I kept getting back odd results, some strange instability, but nothing I could really pin down. I decided to really hammer on the board by installing it on my main machine and seeing what it would do. It died. That's what it did. Again, this is not a common problem I hear about these Chaintech boards, but it is always unfortunate when a reviewer gets a bad board as everyone hears about it. NVIDIA has finally released their latest integrated motherboard chipsets to the market. These are based on GeForce 6 technology, but are produced at 90 nm. Instead of calling these products "chipsets" NVIDIA is making sure people separate them into two distinct products. The graphics portion will be called the GeForce 6100 or GeForce 6150, while the typical southbridge portions will be controlled by the nForce 430 and 410 series. The 6150 is the top end product aimed at the HTPC sector. It features a slightly higher clockspeed (475 MHz compared to the standard 425 MHz) but more importantly it can handle HD tv-out up to 1080p and has a built in TV-encoder. The 6100 only handles standard resolutions and does not feature the encoder. The nForce 430 sports Gigabit-Ethernet, RAID 5, 4 SATA 3 Gbps ports, and Active Armor. The 410 has a lesser feature set, but still covers the basics well (no Gig-E, Active Armor, RAID 5, and only has 2 SATA ports). NVIDIA also introduces Azalia audio for this platform, which is a big step for the AMD crowd. NVIDIA does not expect these products to be big retail sellers by themselves, but rather will cater to the OEM and white box crowd. Featureset and performance should be very good, and most definitely will be on par with what ATI has to offer at this time. NVIDIA is looking to take away a big portion of ATI's integrated marketshare with these products, and the ability to mix and match the chips gives a lot of flexibility. Another very positive aspect of this product is that NVIDIA has fully functioning southbridges, which is something ATI cannot fully brag about. It really is about time that NVIDIA stepped up to the plate in the integrated space. Their last integrated product was the nForce 2 video, which is based on the GeForce MX core. Not exactly a stellar performer anymore. Full DX9 functionality, as well as decent performance (NVIDIA states it should play most mainstream games at 1024x768 with solid framerates), and with the top end chip a whole host of good video out functionality should make this a popular product with OEM's. NVIDIA is introducing this product right in time for integration into the Holiday season buying cycle. By Christmas we should expect to see many boxes of inexpensive computers stocking the shelves of Wal-Mart featuring these chipsets. I think we might even see the rise of the value Home Theater PC as a option from the big guys like HP. Those might actually sell fairly well in an environment like Best Buy, Circuit City, or Wal-Mart. ATI Releases Catalyst 5.9 If there is one area where ATI really excels at, it is their driver support. The Catalyst Crew really has it together, and I can rarely find any complaints with how they do business there. These Catalysts not only fix a few problems, but also add video shader enhancements to de-interlacing. This should provide a nice visual quality increase in video applications (WMV, DVD playback, etc.). No word on if you have to use ATI's DVD decoder or software to get this to work properly (or if it works fine on 3rd party applications like WinDVD or PowerDVD). Again, strong work to the Catalyst folks on releasing another solid product. You can pick them up here. September 18, 2005 An Update on ATI's CrossFire Limitations - Josh On Friday I wrote about the Sil 1161 receiver chip that some had thought would hold the CrossFire Edition cards to reproducing video at a max of 1600x1200 @ 60 Hz for CRT's and 1920x1200 @ 52 Hz for LCD's that support that resolution. Since it is essentially a data link between the two cards, on the surface there appears to be no real reason why this chip would cause the resolution and refresh rate limits that we are starting to hear about. From what I am able to deduce, the Sil 1161 receiver chip is simply a red herring in this equation. It really should not make much of a difference, as it is essentially just passing data from one card to the other. That being said... there is a problem with CrossFire that we have been able to determine to be true. The resolution and refresh rates are being limited by this solution. For users hoping to buy a CrossFire setup and drive high end monitors and LCD's, they are going to be in for a disappointment. The above stated limits are there, and there is currently no work around that we can see. On a standard video card, like the Radeon X800 series, the chip itself has two internal RAMDACs running at 400 MHz (for higher end cards), as well as a TMDS that runs at 165 MHz which can power LCD's up to 1920x1200 @ 60 Hz using reduced blanking techniques. Higher end cards can utilize dual link DVI, but monitors that can accept this are not common. Where the design of CrossFire differs from a standard card is that of the compositing chip being used. The chip in question is actually a FPGA (Field Programmable Gate Array) from Xilinx. FPGA's are typically used as a prototyping tool for custom made ASICs. In this case, it appears as though ATI is just going to use this chip as is, program the functionality into it, and use it for the final products. This is not confirmed, but the spy shots from finished hardware makes this appear to be the case. This FPGA (Xilinx Spartan 3) is a solid part, but it does not have a significant amount of memory built onto it. When dealing with graphics, memory is a big deal. Any kind of frame buffer or advanced effect is going to need some hefty cache or memory, and this FPGA does not have it. Since there is no frame buffer on this chip, or memory that it can directly access, it poses a whole slew of problems. The first is that it cannot store a single frame of output from either the slave or master card. All graphics functionality has to be done on the fly, or it just won't work. The FPGA is then connected to a 240 MHz external RAMDAC (which should allow for 1600x1200 @ 85 Hz or 2048x1536 @ 60 Hz on a CRT monitor) as well as another Silicon Image TMDS (1920x1200 @ 60 Hz with reduced blanking). So then why are these review units being limited to 1600x1200 @ 60 Hz on a CRT? In the end I believe that the compositing chip is holding it back. Because it doesn't have a built in frame buffer to accept output from either of the cards, there were some compromises made on getting it to work in the first place. The way the frames have to be synched as well as passed through the compositing chip is causing the issues. ATI had to make some very heavy design decisions to get it to work in the first place, and this limits users to the undesirable resolutions and refresh rates that we are currently seeing on CrossFire setups. We will have to wait and see what the CrossFire reviews have to say when they are finally released later this month, but the news so far is pretty disappointing. ATI felt they needed to have a strong and competitive multi-GPU consumer level product that will stand up to NVIDIA's SLI, but it seems that while the idea and certain aspects of the implementation are good, it might not be the robust product many were hoping for. It seems to me that ATI has a part already developed that could be changed and used for the compositing chip. A chip that has 10 MB of E-DRAM and runs at a very fast rate, and has a hefty amount of computing power to it. Yes, I am talking about the E-DRAM portion of the Xenos chip, the heart of the X-Box 360. Unfortunately, this product is really the property of Microsoft, but it seems that with the knowledge of the engineering team involved in the project that ATI could quickly and effectively produce a compositing engine of their own that would feature 10 MB of E-DRAM and make CrossFire a truly effective solution without any of the current limitations we are starting to see. CrossFire will be around for a long time though, and we will see many iterations of it in the future. Perhaps the future does involve a compositing chip as I have described above, or perhaps the functionality will be built into the GPU's themselves. We shall see, but what looks to be the current implementation may be a big disappointment to users with higher end CRTs and LCDs. Again, I don't have CrossFire products in hand, so I can't verify this myself (plus I would probably be under NDA if I did have CrossFire cards on the premises). It does appear as though there is a mountain here, but not where many expected it to be. September 16, 2005 ATI's Crossfire Limitation- Mountains Out of Molehills - Josh Rage3D really broke the news on the possible limitations of the Silicon Image Sil 1161 receiver that ATI's CrossFire utilized. The receiver can handle a max of 1600x1200 at 60 Hz (or essentially receive 60 1600x1200 frames per second), and runs at a maximum of 165 MHz. At first this limitation looks like a huge disaster waiting to happen to ATI with the release of CrossFire, and many are screaming that the engineers fell down on the job and should all be sacked. When I was first given this information, I originally thought that ATI had made a big blunder... but upon further investigation I believe that it really isn't that big of a limitation at all. The DVI spec is actually pretty flexible, and by doing tricks such as blanking, higher resolutions can be achieved that should be normally out of the range of the chips involved. Since Silicon Image actually developed the spec, they definitely know it inside and out. The first thing we need to do is separate the fact from fiction, and really show what is going on as well as the limitations that really are there. What we should remember, first and foremost, is that we don't know the exact engineering processes and tricks that ATI is using for their CrossFire implementation. ATI has some pretty clever engineers, and I don't think that something this obvious got past them. First off the Sil 1161 receiver simply receives the pixel data from the slave board. That data is sent to the compositing engine, which then takes the final frame/s and sends it to either its built in TMDS transmitter or the RAMDAC, both of which can handle 2048x1536 @ 85 Hz. The Sil 1161 has NOTHING to do with the refresh rate of the screen that is connected to the master card. So, at the very least, CrossFire users can enjoy refresh rates on CRT monitors greater than 60 Hz at resolutions above 1600x1200 because that is based on the transmitters on the master card, which are essentially identical to every other standalone, high-end ATI card. You can find the refresh/pixel specs here. What we are basically talking about here is bandwidth between the slave card and the master card. Something else to take into account is that the DVI standard assumes that the cable will be at least a meter long between the transmitter and the receiver. In CrossFire the cable between the slave and master cards will most likely be around 10 cm long at most. ATI could very well have tweaked the 1161 to increase its bandwidth. But even if the chip itself is stock, we are still only dealing with this problem as bandwidth. The receiver can handle 1600x1200 @ 60 Hz, which means that it can receive 60 frames of 1600x1200 pixel content per second. When used with the output of the master card, this can give a theoretical maximum of 120 fps at 1600x1200 resolution when using alternating frames. Most high quality 21" CRTs can usually handle a maximum of 100 Hz at 1600x1200, so using AFR with v-sync enabled, 100 fps is theoretically achievable (depending on the application of course). Let's look at other situations. The maximum bandwidth at 1600x1200 @ 60 Hz is around 115 million pixels per second (Mpps). ATI is using three different methods for rendering with two cards. The first is alternate frame rendering, which in most cases is the best scenario due to geometry scaling. The second is SuperTiling, which splits the scene into even and odd "supertiles" (though there is information pointing to SuperTiling not working with 12 pipeline cards- apparently SuperTiling requires even numbers of quads to work correctly). The last is Scissor mode, which splits the screen in half. Again, we are essentially dealing with bandwidth here. The DVI spec is fairly flexible with how it is implemented, and in SuperTiling and Scissor modes, the entire frame doesn't have to be transmitted (which is essentially done pixel by pixel anyway). So, in either of these modes, 1/2 of the total frame needs to be transmitted. Let's stick with Scissor mode so it will be easy to understand. One of the biggest complaints raised by people around the net is that they won't be able to use their 23" and 24" widescreen LCD's at 1920x1200. Let us assume we are using one of these LCD's at that resolution. The Sil 1161 can handle a maximum of 115 Mpps theoretically. Using Scissor mode, the slave card has to render a 1920x600 area, if it is doing so at 60 fps, it is taking up a maximum bandwidth of 69 Mpps, which is far below the theoretical 115 Mpps maximum. Using tricky algebra, we can see that using a 1920x600 "frame" a maximum frame rate of 99.9 fps can be achieved and still reach the theoretical bandwidth limit. Due to the way Scissor scales, the entire composited scene will have a maximum framerate of 99.9 fps. Considering most of the widescreen LCD's can usually handle only 60 Hz (or 60 fps with v-synch enabled) at the higher resolutions, it is not all that big of a deal. Let's move on to a higher end CRT or LCD that can display 2048x1536. In Scissor mode each card has to render 2048x768 pixels, and at 60 Hz that comes out to be 94 Mpps. Again, this is well below the maximum 115 Mpps that the receiver can handle. The maximum fps that the receiver can handle in this mode is 73 fps. While this is short of the 75 fps that is needed to run with v-sync enabled at 75 Hz, it isn't terribly far off either! Most users will not notice much of a difference (this is again assuming the cards can run the application at that speed). Now, things do change around a bit when using alternating frames. At 1920x1200 the receiver can handle a maximum of 50 fps (1920x1200 at 50 Hz, assuming the Sil chip can be programmed as such). At 2048x1536 the maximum rate is 37 fps (2048x1536 at 37 Hz). In alternating frame mode the CrossFire solution will have a max FPS of 100 when used at 1920, or a max FPS of 75 when used with 2048. We are essentially just dealing with bandwidth, as the pixel information transmitted basically contains color information (RGB) and location (clock). As such, the ways with dealing with this information can be pretty flexible. This of course assumes that the Sil 1161 is running at 165 MHz all the time. In summary, CrossFire users should not be hamstrung into using 1600x1200 @ 60 Hz with their high end monitors and LCDs. When digging deeper, we see that it is a probably a fairly flexible solution, and the Silicon Image chip being used won't be that big of a hurdle. Now, in saying that, I think that ATI should have used the Sil 1171 receiver that can handle 2048x1536 @ 60 Hz, as it would have given them a LOT more headroom and would have made this controversy a non-issue. Something to definitely consider though is that we have no idea what the R520 CrossFire cards will be using, and they could very well be designed with the 1171. This may be a bit of a stumble for ATI as they definitely are limiting their performance in these higher resolution situations with the 1161 receiver. My overall feelings towards CrossFire remains the same though. I do feel that it is essentially a knee jerk reaction to SLI, and there are some aspects that were not well thought out or planned for. I believe that NVIDIA's SLI is more flexible, and the over the top connector allows far greater bandwidth than what CrossFire currently enables. I think what is most important though is that we have competition at the high end with scalable solutions. There will be areas where the CrossFire will shine over SLI, and vice-versa. Still, CrossFire has not been released yet, and SLI in its present state is a far cry from what it was when introduced nearly a year ago. ATI still has a big mountain to climb, especially if the leaked benchmarks for the X1800 XT are correct. Feel free to discuss this here.
If you have found this article interesting or a great help, please donate to this site.
Copyright 1999-2005 PenStar Systems, LLC. |
|||
