X-bit labs

There Will Be Speed: The Clash of Modern Multi-GPU Technologies (page 6)

Category: Video

by Alexey Stepin , Yaroslav Lyssenko, Anton Shilov

[ 03/24/2008 | 05:01 PM ]

PCB Design

The new card is as large as the Nvidia GeForce 8800 GTX and ATI Radeon HD 3870 X2 but occupies two slots, like the Nvidia GeForce 7950 GX2.

You may remember that the GeForce 7950 GX2 used to be criticized for certain design flaws, particularly for the poor cooling of the GPU installed on the bottom PCB. The developer has corrected those flaws in the new GeForce 9800 GX2. As opposed to the predecessor card, the PCBs of the new one face each other, and the chips give their heat away to the single heatsink in between them. Nvidia thinks this design most optimal as it simplifies the wiring of the PCBs and reduces the thermal load on the graphics card to achieve higher GPU clock rates.

We disagree with this statement, though. Yes, with this design, each GPU heats up each PCB only, but they also heat up each other through the common heatsink, which negates the possible positive effect. As indirect proof of our point, the new card’s clock rates are reduced considerably relative to the GeForce 8800 GTS 512MB, let alone GeForce 9800 GTX. The achievement of a higher memory frequency through optimal PCB wiring doesn’t sound like a hefty argument, either. Yes, the memory frequency of the reference Radeon HD 3870 X2 is clocked at a lower frequency, but PowerColor’s version is this card features GDDR4 clocked at 1125 (2250) MHz, and the single-PCB design was not a problem for it.

It took us quite a lot of time to take our GeForce 9800 GX2 apart in order to see its internals.

The PCB being so large, we guess both the graphics chips could be accommodated easily on one PCB, especially as the PCI Express switch employed by Nvidia is far smaller than the PLX PEX8547 installed on the ATI Radeon HD 3870 X2. Each PCB has a large shaped hole for the cooler (located inside the sandwich of two PCBs) to get fresh air from. So, the theoretically available area is not utilized very efficiently notwithstanding Nvidia’s claims about optimized wiring. The PCBs are connected with two flexible cables and fastened together with metallic poles.

Each PCB uses a dedicated high-frequency three-phase power circuit based on a Volterra VT1165MF controller. Such controllers are also installed on the ATI Radeon HD 3870 X2. Tiny Intersil ISL6269CRZ chips are responsible for the memory. The use of a high-frequency digital power circuit helped do without large electrolytic capacitors. It is important for the GeForce 9800 GX2 considering its high component density. Each PCB is equipped with a power connector: the bottom one carries a 6-pin plug and the top one, an 8-pin PCI Express 2.0 plug. As opposed to the Radeon HD 3870 X2, you cannot do with two 6-pin cables – the card won’t start up. Near the 6-pin power plug there is a 2-pin connector for the audio-over-HDMI feature which has already become standard.

The plastic casing makes it difficult to plug the power cables into the GeForce 9800 GX2 because the slit for the connector is very narrow. A connector with a large lock just won’t fit into the plug. This refers to our Enermax Galaxy DXX EGX1000EWL power supply, for example. We just had to cut the locks off the PSU’s connectors. It is also impossible to unplug a connected power cable without a thin-tipped screwdriver for the same reason. This problem is mentioned on the Nvidia's official website in the list of PSUs certified for the use with the GeForce 9800 GX2.

The PCI Express switch installed on the bottom PCB is marked as BR04-300-A2.

We couldn’t find its specs, but we know it supports PCI Express 2.0 and can switch 48 lanes. The GPUs are marked as G92-450-A2 whereas the GeForce 8800 GTS 512MB carries chips marked as G92-400-A2. This might mean a new version of the G92 with a higher frequency potential if it were not for the same revision number. Both chips are dated the 52nd week of 2007, i.e. late December. As we mentioned above, the GPUs have a main domain frequency of 600MHz and a shader domain frequency of 1500MHz. The GPU configuration is standard with 128 ALUs, 32 (64) TMUs, and 16 ROPs.

The PCBs of our GeForce 9800 GX2 carry GDDR3 memory manufactured by Samsung. These K4J52324QE-BJ1A chips have a capacity of 512Mb (16Mbx32), a voltage of 1.9V and a rated frequency of 1000 (2000) MHz. The memory is indeed clocked at this rated frequency, proving a bandwidth of 64GB/s to each GPU.

The total amount of graphics memory is 1024MB, but 3D applications can only use 512MB due to the specifics of homogeneous multi-chip architecture. As a result, the new card may prove slower than Nvidia’s GeForce 8800 GTX/Ultra and GeForce 8800 GTS 1024MB in some games that need more than 512MB of memory.

The reference GeForce 9800 GX2 is equipped with two DVI-I and one HDMI port, but the latter can be replaced with a DisplayPort. The top PCB has a seat for a translator chip of that interface. The card doesn’t support analog video output, which is normal for today. However, there is an empty seat for some connector on the bottom PCB, near the DVI port. Perhaps it can be used to implement the support of S-Video/RCA interfaces.

Besides the connectors, there are two LEDs on the card’s mounting bracket: a dual-color and a blue one. The former indicates power-related problems. It is green when both power cables are attached and red when one or both cables are missing or when both cables have 6-pin connectors. The second indicates a significant drawback of the GeForce 9800 GX2 in comparison with the ATI Radeon HD 3870 X2: it doesn’t support more than one monitor in SLI mode. This LED highlights the Master DVI port you should connect your monitor to in SLI and Quad SLI modes. Nvidia is expected to do away with this limitation in the next versions of ForceWare.