|
ATI Crossfire Xpress 3200 (RD580) Technology Preview
Q&A Session
Q. Is the PCI Express performance of the Nvidia nForce4 SLI X16 architecture impeded by having the secondary PCI Express X16 slot served by the MCP? Nvidia claims in an email blast that there are no performance issues caused by having the PCIE X16 slots connected to different chips, and that SLI performance is not appreciably affected by this extra latency. A. This is another attempt by Nvidia to rewrite the basic principles of mathematics. If this claim were true, then the press would be able to measure 8GB/s total between Nvidia's NB and SB (a 16-bit HT 1GHz link between the NB & SB equates to a 4GB/s link (in either direction) or 8GB/s total). When you actually measure it, you'll find about 1.5GB/s total bandwidth (or 750MB/s either way with true bi-directional traffic), which is about 25% of the required bandwidth for a true X16 PCI-E graphics slot. This is a pretty obvious bandwidth limitation. This is why Nvidia's dual X16 motherboards sometimes score less than their dual X8 motherboards. In addition, consider that this bandwidth needs to be shared with Gigabit LAN, SATA and all other I/O devices connected to the SB. It is clear why this is an inferior solution. Below is a chart that illustrates this bottleneck. ATI uses two X1900XTs to generate approximately 2.5GB/s of total peer to peer traffic (cards are "talking" to each other). We are measuring both one-way as well as bidirectional traffic, against the competitor at even 5X HT setting (1000MHz). The graph shows how the performance scales on the Nvidia board by changing the speed of the NB-SB link (always using the maximum 16 bits up and down stream).   Q. Why does CrossFire Express offer exceptional PCI-E scaling moving to 16 lanes, while your competitor in some cases delivers less performance with 16 lanes than with 8? A. ATI's CrossFire Xpress 3200 uses a highly refined dual X16 architecture that doesn't bleed away PCI-E bandwidth and performance to the southbridge, like competing solutions. In addition, our competitor has a long and documented history of having Hyper Transport Interface issues dating back to the AMD K7 and because of this many large motherboard manufacturers use a 600 MHz NB-SB link by default to preserve stability. Q. Nvidia claims that comparisons between dual X8 and dual X16 platforms, should use higher performance GPUs as mainstream GPUs will not generate the traffic necessary to saturate a dual x8 connection, and that for a valid test, you need to use high-end cards that can run at very high resolutions with very high 16x antialiasing and anisotropic filtering settings. Only this, they say will generate the traffic necessary to make a valid comparison between dual X8 and dual X16 platforms. A. If a mainstream GPU doesn't generate enough traffic to saturate the dual x8 connection, then why do mainstream SLi setups like 6600 SLi and 6600 GT SLi need a separate connector to provide enough bandwidth for the cards to communicate? With mainstream cards like X1600 and X1600 XT, ATI's solution offers enough bandwidth to use two cards without a connector. Nvidia's motherboard solution never provides enough bandwidth to do that. This also means that future applications, such as using one graphics card as a physics processor, are also impractical on Nvidia's platform. And as for the high end, ATI's Crossfire Xpress 3200 + X1900XTX Crossfire platform is clearly superior to any SLI solution, both at default speeds and in an overclocked scenario. Q. Nvidia claims that there is no architectural advantage to combining both X16 PCI Express slots on a single chip like the ATI RD580 architecture rather than splitting them between the SPP and MCP as NVIDIA does with the Nvidia nForce4 SLI X16 architecture. How do you respond? A. It's clear that if you were designing a dual x16 platform from the ground up, you would put both links close together on one chip and make the path as short as possible for lowest possible latency and maximum performance. This is even backed up by Nvidia's roadmap, which shows a single-chip solution in their future. Q. Is it true that NVIDIA SLI GPUs show negative performance benefits when comparing dual X8 and dual X16 platforms? Nvidia claims this isn't true? A. As stated above all tests on the Nvidia dual X16 motherboards show major performance bottlenecks 16 bit up and down stream at 1GHz as recommended by Nvidia in their benchmarking guide. However, it should be noted again that there is no shipping Nvidia dual x16 configuration in which the NB-SB link actually runs at 1GHz. The motherboard vendors prefer to set it at 600MHz, perhaps for stability reasons. Q. Nvidia claims they already have a competitive response to ATI's RD580 platform. Can you comment? A. If Nvidia already has a competitive platform, why does it's roadmap show a dual x16 PCI-E graphics part with both graphics ports on one chip?
Related Articles Editorials August 27, 2009
Editorials August 18, 2009
Editorials March 3, 2009
Editorials February 12, 2009
Add Comment
2 User Comments 1 - Posted by
Kurtis
on March 8, 2006 - 6:39 am I just noticed this article over at PCPerspective. Ryan investigates ATI's claims about bandwidth bottlenecks on NVIDIA's SLIx16 boards and comes to the conclusion that these claims were not based in fact. Have a look: http://www.pcper.com/article.php?aid=212&type=expe... 2 - Posted by
Frosty
on March 8, 2006 - 5:24 pm i knew it those bastards lied!
Add Comment To add a comment without being a member, you may omit the password field, but you must enter your name (or nickname) along with your comment. * Denotes required fields. |
Recent Reviews Video Cards September 23, 2009
Gaming Devices July 17, 2009
Portable Media Players July 16, 2009
Thumb Drives July 14, 2009
Recent News VICE Nov. 20, 2009 - 7:17 pm
Wired Nov. 20, 2009 - 7:07 pm
Engadget Nov. 20, 2009 - 6:58 pm
Telegraph.co.uk Nov. 20, 2009 - 6:48 pm
BBC Nov. 20, 2009 - 6:38 pm
Destructoid Nov. 18, 2009 - 11:55 pm
Wired Nov. 16, 2009 - 11:56 pm
Ultimate Typing Championship Nov. 16, 2009 - 11:42 pm
|
|||||||||||||||
