PCIe question

[madhatter300871]

I have a motherboard with PCIe1.1 slot. I currently have a PCIe 2.0 nvidia 8400GS installed into that slot, so the slot does support PCIe 2.0 cards.

I am about to upgrade my graphics card, am I OK to choose a PCIe 2.0 and a PCIe 3.0 card, are they backwards compatible with a PCIe 1.1 slot, or have a mis-understood ?

Many thanks.

[LoRd_MuldeR]

Yes, PCIe is backward compatible, at least in theory. If you combine devices with different PCIe versions, then the "slower" one will prevail.

FWIW: I have a GeForce 600 series card (GK104) with PCIe 3.0 support myself. And it's running fine on my Intel P35 Express board with only PCIe 1.1 support. Your mileage may vary ;)

(From what I've heard, there sometimes are problems with PCIe 2.1+ cards and old PCIe 1.0/1.0a boards, while the situation with PCIe 1.1 boards seems much better)

[madhatter300871]

Thanks for clarifying. I'm not looking anything bleeding edge, just think it's time to upgrade the old 8400GS. It's being pushed quite hard decoding HD x264 video and using madVR as my renderer. It works, I just feel like treating myself.

Is it a fair assumption to say that any newer Nvidia card will at least match the 8400GS, a newer card couldn't perform worse could it ? (I know that's a subjective question, I just mean generally).

In your experience, what are they points I should look at for, such as cuda cores, processor clock, graphics clock, bandwidth, or is it simply a matter of as many cores and as much ram as I can afford to buy. I seem to have adopted the theory that a newer card can only be better

[LoRd_MuldeR]

x264 doesn't use the GPU at all. At least that's the current situation. Apparently some work is going on the create a CUDA-based lookahead, but it's not there yet.

In other words: If x264 is your target application, the GPU doesn't matter at all. And even if you want to use a the graphics card for decoding, e.g. via DGDecNV, any cheap model will do the job.

MadVR is another topic, because it actually does use the GPU for rendering. On the other hand it has a lot of options to tweak GPU load...

(FWIW: On my old GTX 260 the "Jinc" mode, especially above 3 taps, was too much, though "Lanczos" was working fine. Now with my GTX 660 Ti even "Jinc" mode with 8 taps works flawlessly)

[madhatter300871]

I play movies encoded with x264 and LAV filters uses CUVID for decoding. Doesn't that mean that x264 decoding is being done by the GPU ?

I know the x264 application doesn't use the GPU while encoding, I was thinking purely about playback.

So, any old Nvidia will do the job of decoding 1080p HD, but a more substantial card is needed if I want madvr to do some serious processing. Have I understood correctly ?

Price is an issue, i was looking at getting a GT620, can get one for under £50. Do you have any thoughts ?

[LoRd_MuldeR]

I play movies encoded with x264 and LAV filters uses CUVID for decoding. Doesn't that mean that x264 decoding is being done by the GPU ?

Nope. There's a dedicated decoder "chip" on the graphics card for "hardware" H.264 decoding (and some other formats). NVideo calls it the "PureVideo HD" engine.

It's not like they have implemented a H.264 decoder in CUDA/OpenCL to run on the actual GPU. It's a separate hard-wired logic. And it's pretty much the same between "cheap" and "high end" models.

(Sometimes "cheap" models even have a newer incarnation of the "PureVideo" engine, as compared to the current "high end" model)

So, any old Nvidia will do the job of decoding 1080p HD, but a more substantial card is needed if I want madvr to do some serious processing. Have I understood correctly ?

Yes, even older NVidia cards will decode H.264 videos in "HD" resolution just fine - as long as you use some decoder that actually makes use of the "PureVideo" hardware and only if your specific H.264 stream conforms to what the "hardware" decoder can handle. Hardware decoders are much more picky than pure software decoders!

About MadVR: Yes, to use to more "GPU intensive" settings (without frame drops) you need a powerful GPU, because it does the processing entirely via GPU shaders.

[madhatter300871]

Well I never knew that there was a dedicated chip for hardware decoding, thanks for that explanation. :D I have had a good read of doom9 about producing bluray compatible x264 streams and, having followed it, I haven't had any hardware decoding problems so far (touch wood!). Does the GPU handle DXVA, or is that done by the dedicated chip as well (just wondering) ?

I'll have a good search through the madvr page here at doom9, see if there are any graphics card recommendations. The 8400GS allows me to select Spline 4 tap scaling in madvr for SD content, it chokes on HD and I can only select bilinear, that's why I was thinking of treating myself to a new card.

[LoRd_MuldeR]

Again: The "hardware" decoding is done by dedicated circuits (function units), called "PureVideo HD" in NVidia jargon. AMD calls theirs "Unified Video Decoder" (UVD).

In tools like GPU-Z you can even see that "GPU Load" and "Video Engine Load" are two separate things. Note that "GPU Load" doesn't go up when the graphics card is used only for video decoding.

DXVA (DirectX Video Acceleration) is a programming interface (API) that Windows applications can use to access the graphics card's hardware decoder. DXVA can work with NVidia and AMD cards.

NVidia cards can alternatively use CUVID (CUDA Video API). And on Linux/Unix there is the VDPAU (Video Decode and Presentation API for Unix) for that purpose. VDPAU currently is NVidia only.

The "hardware" decoder (video engine) will be one and the same, regardless of what API is being used to access it.

[madhatter300871]

Thanks, again, for the explanation. I use LAV filters, set to use CUVID for decoding.