What is PureVideo?

The Old PureVideo Architecture

Every GeForce 6 and 7 with PureVideo (excluding the AGP versions of the 6800 Ultra/GT) and 6100 has featured a dedicated programmable video processing unit (VPU) on the physical chip. This programmable VPU is separate from the other parts of the GeForce that deal with video (like the scaler) and is also separate from the GPU shader part of the chip. It’s like having a dual-core CPU – one core is the regular GPU, and another core is a completely different VPU or something similar to the Vector Units found in the PlaySation 2’s Emotion Engine.

Although NVIDIA has never provided specific details about the underlying architecture of PureVideo, our best analysis suggests that the GeForce’s video processing pipeline involves several elements :

1. Fully-programmable VPU (H.264 decode, VC-1 decode, some deinterlacing)
   
2. GPU Shaders (some deinterlacing, noise reduction, sharpening)
   
3. Fixed-function video processing elements (video scaler, HDCP, MPEG-2 IDCT)
   
4. Software (Bitstream processing, H.264 IDCT)
   

The New PureVideo HD

With the GeForce 8500 and 8600, NVIDIA has introduced a new second-generation PureVideo solution. It’s still called “PureVideo HD” but it’s a substantially different product.

The GeForce 8500/8600 now features a second generation VPU, codenamed VP2. Although architectural details for the VP2 are limited, we do know that NVIDIA claims that secondary video streams can also be decoded on-chip (i.e. Picture-in-Picture – useful for certain HD-DVD and Blu-Ray). Assuming that there’s no increase in efficiency from the software side, the ability to decode secondary streams suggests that the VP2 is twice as powerful as the existing VPU inside the older GeForce CPUs. NVIDIA has had a history of running some deinterlacing and quality-enhancing features on the VPU as well as the shader. With time, we may see new features added to VP2.



The real innovation is the new H.264 Bitstream Processing Engine. This dedicated unit provides full H.264 decoding, including context-adaptive variable-length coding and context-adaptive binary arithmetic coding (CAVLC/CABLC) reverse entropy. Reverse entropy is the most computationally intensive step of the process and accounts for more than half of the decode time. Remember those movies that took a Core 2 Duo to its knees? With the GeForce 8600GT, you’re looking at just 20% CPU utilization on that same system. Want another perspective? That same clip that took the Core 2 Duo to its knees? A GeForce 8600 and a Celeron 347 (3.06GHz Cedar Mill – like a 65nm Prescott) will be fine…

The GeForce 8600’s shaders are also put to good use with support for nearly all of the quality features found in a GeForce 8800GTX-class machine. You have things like content-based HD inverse telecine (the ability to recover the full 1080p24 source hidden in a 1080i60 broadcast), NVIDIA’s more advanced spatial-temporal deinterlacing, capable of passing the “guitar string” test on the HQV Benchmark, and even support for unusual SD cadences. The only thing that’s currently missing is HD noise reduction and sharpening. (Although with NVIDIA’s history of the 6600GT, we may see new features with new drivers over the lifetime of the product).

Finally, the AES128 engine provides a fixed function pipeline for accelerating the content protection found in AACS and potentially BD+. The 8600GT line now mandates full HDCP-support with the full encryption key support. Importantly, the GeForce 8600 now allows 30” Dell owners to enjoy full HDCP protected video.

When it comes to software, NVIDIA no longer requires a NVIDIA-branded PureVideo software decoder. (They haven’t since last year). With Windows XP, PureVideo quality features will be available to any DXVA application but H.264 decode will only be enabled with partner software such as Intervideo, Cyberlink, and Nero via a proprietary interface. With Windows Vista, NVIDIA is writing to the Vista-only DXVA2. This allows any developer to tap into NVIDIA’s H.264 decode abilities as well as NVIDIA’s video processing features. With the standardized DXVA2 interface, even enthusiasts can write to this API, and I wouldn’t be surprised to see software such as ffdshow or DScaler being written to take advantage of the new technology. Officially, PureVideo quality features are enabled by the built-in Windows Vista MPEG-2 decoder, but I still seem to have had better results with Intervideo’s decoder.