More mGPU details
One of the most frustrating parts of the event in Las Vegas was the company’s reluctance to talk about the mGPU inside nForce 780a SLI, nForce 750a SLI, nForce 730a and GeForce 8200. During one of the discussions I had with two of Nvidia’s representatives, I gathered a few more details about the graphics chip in these products.
The company again spent a lot of time talking about the GeForce 8200’s PureVideo HD capabilities, which we outlined in yesterday’s article, but this begged the question: are the GeForce 8200’s PureVideo HD features also available on the nForce 700a series mGPUs? Everything I had learned from talking to Nvidia’s partners seemed to suggest otherwise, but given that there has been a lot of confusion, it was only right to attempt to get confirmation from the horse’s mouth.
David Ragones, Director of Nvidia’s MCP business, explained that while there are some differences in silicon between the four different products, none of them are in the GPU portion of the chipset. In other words, the portion of silicon that is the GPU on nForce 780a SLI is physically the same as the one on the GeForce 8200, nForce 730a and nForce 750a SLI chipsets.
The PureVideo HD features in the GeForce 8200 are also in the nForce 700a series mGPUs
In an attempt to understand the magnitude of the task ahead, I also asked about the driver development schedule for Hybrid SLI. I was told that the task is a massive one and that’s why things aren’t quite as we would expect them to be at this early stage. Nvidia jogged my memory on the state of SLI when it first came to market late in 2004 again. At that point there was support for just three games, which was obviously less than optimal. The software development task ahead for Hybrid SLI is of the same magnitude – if not even bigger, although, it’s for different reasons.
When I asked about automatic switching between the two different HybridPower modes, David said that “as we move into future [driver] releases, we expect it to be automatic – we will use application profiling to achieve this.” He then quickly added that the user will still be able to manually control the settings too in the same way that they will be able to at launch.
As a result, it includes support for 100 percent offload for decoding all HD movies, regardless of the codec used to compress it. This is great news, because the way that Nvidia had presented the technology suggested that only the GeForce 8200 mGPU actually supported the new PureVideo HD features.
HybridPower mode enables the mGPU to surrender most of its frame buffer when you're in performance mode
Part of my chat with Nick Stam was about how memory is allocated during Hybrid SLI. There are two parts to this – the first was related to how the mGPU’s frame buffer is allocated when running HybridPower’s ‘performance’ mode. Ideally, you will want the maximum amount of system memory available when you’re playing games that use a lot of RAM. Nick explained that, “in Vista, there is dedicated memory and shared memory combined. Memory carved out at boot that is dedicated to the mGPU for refresh operations is still needed for display refresh. The shared memory is allocated only when needed and would be allocated to the dGPU, which is the ideal case for games.”
The second part of my question was in relation to how the frame is rendered to the screen in performance mode – I wanted to know whether it would be possible to bypass system memory by taking a direct path from the discrete GPU to the northbridge and then straight out to the display without the need to write the frame to the mGPU’s front buffer located in system memory.
“The mGPUs require the display surface to be stored in memory for refresh,” said Nick. “Remember that we have to refresh the display at 60Hz or more. If the display surface passed directly from dGPU to mGPU to display, then the dGPU would have to serve up 60+ fps. If the discrete GPU could only render 40 frames per second, it would look really bad.
“It is also more efficient and guarantees no tearing if the display refresh is handled by the mGPU, independent of the rate that rendered frames are served by the dGPU,” he added.