Best Nvidia CUDA Card for HQPlayer

[b0bb]

On 6/24/2023 at 11:38 AM, Apollo said:

I am looking around for a second hand gpu with min 16Gb memory for HQplayer.
Need some advice, on following 2 offers:

 

1) PNY Nvidia RTX A4000 16GB : 600€

2) Gigabyte Aurora Nvidia RTX 3090 24GB : 450€

 

which one would you recommend?

 

The 3090 card is good if you want to use the heavier filters and modulators, especially if you want 44k to 48k (or vice versa) conversion to high rate DSD (DSD512 or DSD1024).

 

It is up to 3X faster than A4000 but you will have to use all 3 8pin PCIe connectors, the card can draw up to 450-500W, Nvidia Ampere GPU architecture pretty much scales linearly with the amount of power consumed, A4000 TDP is about 140W.

 

Gigabyte 3090 is 3 slots wide, A4000 is 1 slot, with the Gigabyte card on the motherboard, the adjacent M2 SSD slots are not really usable as the cooler vents hot exhaust air onto to the areas where most auxillary M2 slots are commonly located.

 

There have been reports of the area around the PCIe latch area developing hairline cracks, the card has to be supported on the other end to reduce stress. 

Checking for hairline cracks should be done before buying any used Gigabyte cards.

https://www.tomshardware.com/news/gigabyte-rtx-30-rtx-40-series-gpu-pcbs-are-reportedly-cracking

 

The PCIe connector near the latch is a little longer and thicker compared to cards from Asus or MSI, my 3080ti has to be pushed all the way in to the slot until the retaining latch pops up to lock the card into place

[b0bb]

15 minutes ago, paradis said:

What would be a merit if we can find some Titan V cards that has 7teraflops (way more than 4090) but limited by only 12GB of memory?

 

Would Titan V type of GPU a good candidate for today considering it can be found used lesser than a 3090

It depends on how much power is used by your workload. Titan V is 250W vs 3090's 350W (Founders Edition).

7TFlops is a peak number, once the GPU exceeds its power limit, this number comes crashing down.

 

3090 does more in practice as it has an additional 100W to use, Titan V will throttle when the card's power consumption gets to 250W

I have the Titan V and 3080Ti (365W limit), the 3080Ti is more performant.

[b0bb]

On 10/24/2023 at 2:01 PM, Triplefun said:

alternatively see the cheaper K40 on ebay. The K40 requires 6 pin power, has a passive option, has 2880 cuda cores (series 3.5) BUT supports 1,4300 double precision (ie. FP64).  The RTX4070 has 5888 cuda cores but supports only 455.4 FP64.

 

So would the tesla k40 be better for HQplayer than the k40.

 

Note there is also a passive version (no fan) of the k40 !!!

I had the P100 12GB datacenter card, no fan either, however cooling is not optional, I used a fan + 3d printed  shroud for this.

Pinouts for the power connector are also not standard, needed adapter to connect the card to the pcie power connectors on my PSU.

If you are interested in an older datacenter card , the table below shows the compute capabilities of the various model.

 

P100 is the first model with HBM memory built into the GPU, holds its own quite well in HQP5, used prices are dropping on this one, about half of what I paid 4 months ago.

[b0bb]

On 10/30/2023 at 2:27 AM, paradis said:

@b0bb how would you compares the P100 to the RTX series, how performant is the P100? I understand that you've had RTX3080 for instance.

I use the RTX3080Ti, this is the 3090 with half the memory.

P100 is 50%  of the performance.

Do not have data for other RTX devices.

 

For the filters I have tried (polysinc-gauss and SincMX) to DSD512 output rate, the P100 does the job without dropouts.

 

GPU is load is close to 85% in the scenario above based on numbers from nvidia-smi and nvtop tools.

RTX3080Ti in this situation is 35-45% load.

 

P100 struggles with non integer conversions while simultaneously converting PCM to DSD eg 48k/96k/192k to DSD512