NVIDIA 780a: Integrated Graphics and SLI in One

Integrated Graphics

Beginning now, all new NVIDIA chipsets will ship with integrated graphics (which NVIDIA is now calling the mGPU), regardless of what market segment they are targeted at. It's a particularly bold move by NVIDIA but much appreciated given that the mGPU in all of its chipsets will receive PureVideo HD and thus can fully accelerate H.264/MPEG-2/VC1 decode streams.

While it's unlikely that many would purchase a high-end motherboard based on the NVIDIA nForce 780a SLI chipset and simply use its integrated graphics, the mGPU in the 780a is the same GPU used in the 750a, 730a, 720a and the GeForce 8200 based motherboards, so the discussion here is far reaching.

AMD 780G vs. NVIDIA 780a Graphics Architecture

AMD has built a superior Integrated graphics part this time around, both from a technical standpoint and in terms of realized performance. It isn't that AMD really went much further than NVIDIA in terms of engineering something great: they just selected a higher performance core to integrate into their chipset than NVIDIA did.

Neither AMD nor NVIDIA told us exactly how the built their interface to the system bus and system memory, but the lack of a local framebuffer does mean that fast and as low latency as possible communication with system memory are required. In both cases, the discrete GPU from which the integrated part is derived uses a 64-bit width connection to local memory. In both cases, since system memory offers a 128-bit wide but these parts make use of a wider bus to help compensate for increased latency to system memory. Increasing local (on die) cache would also help here, but since IGP solutions are as low cost as possible it doesn't seem likely that we've got loads more cache to play with.

We used 3dmark's single texture test to try to get an idea of memory bandwidth. The test largely removes computation overhead and ends up just pulling in as much data as possible as fast as possible and throwing it up on the screen. The result in MTexels/sec shows that NVIDIA has a bit of an advantage here, but the gap isn't huge. This means that performance differences will likely come down to compute power rather than bandwidth

Past here, NVIDIA and AMD integrated hardware diverge. AMD's solution is based on the RV610 graphics core. In fact, it is an RV610 core shrunk to 55nm and integrated into the Northbridge. This means we get 8 5-wide blocks of shader processors (SPs -- 40 total). In the very very worst case, we get 8 shader ops per clock (which isn't likely to happen in any real situation). Compare this to NVIDIA's G86 based 8 SP offering with a maximum of 8 shader ops per clock and we see quite a difference emerge. AMD's IGP can handle 8 vector instructions per clock and then some, while the similar code could run at 2 instructions per clock on NVIDIA hardware.

Of course, this difference isn't as decimating to NVIDIA as one might think at first blush. We must remember that NVIDIA cranks up it's shader clock to ridiculous speeds while AMD's shaders all run at core clock speed. With AMD and NVIDIA core clocks both coming in at 500MHz, NVIDIA's shader core runs at 1200MHz. In spite of the fact that AMD's part can do more operations per clock (probably averaging out to somewhere between 3x and 4x; it heavily depends on the application), NVIDIA is able to do 2.4x as many clocks per second which closes the gap a bit.

The only discrete part with 8 SPs is the GeForce 8300 which is OEM only. As of this writing, NVIDIA has not confirmed details other than core and shader speeds and the number of SPs in the part with us. They have stated that their integrated hardware is simlar to the 8400/8500 in order to optimize the benefit of Hybrid SLI, so it's possible the number of texutre and ROP units are 8 each. Of course, if half the number of SPs is "similar" to the 8400 and 8500 parts, we can't really be sure until NVIDIA confirms the details. We do know that AMD's hardware has 4 texture and 4 render outs since it is RV610. With so few SPs, and the competition sticking with 4/4 texture/render units, we suspect that this is what NVIDIA has done as well.

What is clear is that either way, AMD's hardware is more robust than NVIDIA's offering. Our performance tests reflect this, as we will soon show.