The Radeon X1900 Family
As we mentioned earlier, ATI is releasing four Radeon X1900 cards today. Pictured below are the new flagship Radeon X1900 XTX and a Radeon X1900 CrossFire Master card. A lower clocked Radeon X1900 XT and an All-In-Wonder Radeon X1900 256MB card are being introduced as well, but they haven't arrived in the lab just yet. The XT looks just like the XTX, though, and the All-In-Wonder X1900 is a dead-ringer, at least physically, for the AIW X1800 XL reviewed here.
The card pictured here is a 512MB Radeon X1900 XTX (MSRP $649). At its heart is an R580 GPU comprised of approximately 380 million transistors, built using a .09 micron manufacturing process. The GPU on this card is equipped with 48-pixel shader processors, 8-vertex shader processors, and a 256-bit 8-channel GDDR3/GDDR4 memory interface. The internal Ring Bus memory controller is 512-bits, however.
The Radeon X1900 XTX's core is clocked at an impressive 650MHz and its memory is running at a robust 1.55GHz. To sustain these high clock speeds, the Radeon X1900 XTX sports the same beefy dual-slot cooler used on the Radeon X1800 XT. The PCB is much larger than X800 series Radeons, and this card is equipped with Volterra's multi-phase voltage regulator chipset (under the thin, red heatsink). The lower clocked Radeon X1900 XT (625MHz / 1.45GHz) will look similar, but will be priced initially with an MSRP of $549.
We were curious to see how large the Radeon X1900 GPU really was after hearing that the core is composed of over 380 million transistors, so we popped the heatsink off of the Radeon X1900 Master Card to take a closer look. Using a trusty old ruler, we found the Radeon X1900 GPU to be roughly 18.5mm x 18.5mm, or 342mm2. Conversely, a GeForce 7800 GTX, which is built on TSMC's .11 micron line, is a bit larger. We measured a GeForce 7800 GTX (G70) at approximately 19mm x 18.5mm, or 351.5mm2. And we found an X1800 to be roughly 288mm2. If yields at TSMC are good, it could be more cost efficient for ATI to produce X1900s than it is for NVIDIA to currently make the GTX, which could push street prices down in time. Then again, packing 512MB of 1.55GHz GDDR3 RAM on a flagship card won't be cheap for the foreseeable future. NVIDIA also has plans to shrink the G70 down using a 90nm process, which would then give NVIDIA a significant edge.
The Radeon X1900 Master Card looks very similar to the X1900 XTX on the surface. The GPU on our X1900 Master Edition card was clocked at 625MHz and its memory was clocked at 720MHz (1440MHz DDR), just shy of the 650MHz / 1550MHz of the Radeon X1900 XTX. The aproximate 100MHz memory clock disparity should have a relatively small effect on performance, so we won't dwell on it here. And CrossFire doesn't require a matched-pair of video cards to function, so the difference in memory clock speed shouldn't pose a compatibility problem either.
To bring CrossFire to the X1900, ATI used the same compositing engine introduced on the X1800. The second generation compositing engine used on the X1900, is similar but superior to the one used on the older Radeon X850 XT Master cards. If you remember, because X850 cards were equipped with single-link DVI outputs, X850 CrossFire was limited to a max resolution of 1600x1200 with a lowly refresh rate of 60Hz. Graphics cards in the X1K family of products are equipped with dual-link DVI outputs, however. Having dual-link DVI outputs means more bandwidth, which the new compositing engine capitalizes on to offer higher-resolutions than the first generation CrossFire implementation.
The compositing engine on the Radeon X1900 XT CrossFire Edition Master card consists of a handful of chips. The biggest chip in the group, in the middle of the picture, is a Xilinx Spartan XC3S400 FPGA. The XC3S400 is a more capable FPGA (Field Programmable Gate Array) than the one ATI used to enable CrossFire on the X850 XT. The XC3S400 is the chip that's programmed to do the actual compositing work. In total, this chip has ~ 400K logic gates, which is fairly low-end by today's standards for an FPGA. And at a cost below $7, the overall retail price point of the board isn't affected too adversely. We should note that it's upgradeable via firmware as well, so ATI could theoretically incorporate more features into X1900 CrossFire moving forward. To the left of the Xilinx FPGA is the flash ROM chip, that actually contains the necessary programming and configuration code.
Above and to the left of the Xilinx FPGA in the picture, are a pair of Silicon Image SiI 163B TMDS receivers. These are the chips that receive data from the slave card. The "data" is the information being transmitted from the slave card's dual-link DVI output and on through the custom dongle pictured above. The Silicon Image SiI 163B TMDS receivers are clocked at 165MHz and are capable of processing images at a resolution of up to 1600x1200 @ 60Hz. But because there are two of them that work in tandem, the maximum resolution is doubled to 2560x1600. The three smaller Silicon Image chips to the right are a pair of SiI PanelLink TDMS transmitters and an Analog Devices RAMDAC, which then power any displays connected to the output on the CrossFire dongle.
As you can see, although the Radeon X1900 has 48 pixel shader processors, it's raw pixel fillrate is only marginally higher than an X1800 XT, because it has the same number of ROPs (Raster Operation units) and texture units -- 16. The X1900 XTX's faster memory and core, should give it an advantage over the X1800 XT though, even when the 48 pixel shader processors aren't being fully utilized.