Dual Chips on a Single PCB: Panacea or Poison?
Dual-processor graphics cards have never become very popular and widely spread, because they consumed a lot of power, cost more and besides, the driver quality left much to be desired, and the performance growth they provided was incomparable with the price growth all these innovations caused.
Theoretically, a dual-chip solution on Volari V8 Ultra looks not bad at all: two 128bit memory buses provide the total of 256bit with the bandwidth of 32GB/s (in case they use GDDR-II 1000MHz). And the number of rendering pipelines increases up to the impressive number of 16, thus ensuring enormous fillrate of 5.6Gpixels/s. As for the supported memory types, single-chip Ultra solutions were supposed to work with DDR memory chips working at 750MHz, and the non-Ultra models – with 650MHz memory. The more powerful dual-chip products were intended to work with GDDR-II supporting 900MHz or even 1000MHz frequency.
The approach that allowed XGI to host two chips on a single PCB was called BitFluent. According to the technical documentation available on XGI web-site, the chips are connected with one another with a special 32bit bus working at 133MHz (the efficient frequency is equal to 533MHz), and providing the bandwidth of 2.13GB/s. From the hierarchical point of view, the first chip connected to the AGP bus is the primary one, while the second one connected to the primary chip with the above described bus is the secondary one. Here I would like to point out that AGP 8x and BitFluent feature a lot of similar technical details, that is why the latter is very likely to be almost a full analogy of the AGP 8x.
According to XGI, each VPU processes its frame. Moreover, both VPUs work in parallel and neither of them has to wait for the counterpart to complete its work. Of course, each VPU features its own frame-buffer and memory bus.
BitFluent technology boasts a number of enhancements, which allow minimizing the latencies and increasing the inter-chip bus efficiency.
It is a pretty funny observation, I suppose, but even though all multi-processor graphics systems intended for both: the mainstream users and professionals, use different algorithms ensuring parallel functioning of the two chips, the questions about their efficiency are always the same.
Here is a list of potential problems XGI’s solution may face:
- They might need to use twice as much memory onboard, as ATI and NVIDIA solutions carry. Since each chip has its own frame-buffer, it means that each chip uses a limited amount of its own memory, which can be not enough for high-performance rendering. For example, in order to process over 128MB of textures without involving the system memory resources, the card should have 512MB of memory onboard: 256MB for each of the two chips.
- The potential bottleneck of this system is BitFluent bus connecting the two chips with one another. Even if the frame is “optimized” with contemporary compression techniques, it will still occupy some memory, so that a 2.1GB/s bus between the chips can turn out a grave bottleneck limiting the performance.
- If one processor needs some data from another processor for the next frame processing, the system performance will drop to or even below that of the single-chip solution.
- Complex PCB layout.