Best Nvidia CUDA Card for HQPlayer

[Rovo]

one remark:

your CPU has P- and E-cores. It can be that in the BIOS there are some settings required or adviced when using such a CPU. Others can jump in to advice. 

[Apollo]

7 minutes ago, Escarbille said:

 

Hello,

 

@Apollon, why add a CG to your 13900, I think if you limit yourself to DSD512, the cpu must know how to make most filters work...

Yes indeed but with some filters (poly-sinc-gauss-xla) and 48kHz/24bit I get some dropouts.

Although the system should be cooled well enough, I think there could be some thermal throtthling.

If the A4000 is not bringing the expected results, I can send it back to vendor.

[Miska]

12 hours ago, Rovo said:

- in the Guided storage configuration UNCHECK setup disk as LVM group (this could otherwise interfere when you are using a HQPlayer USB key/dongle)

 

Not as simple, but I do manual partitioning and create 512 MB UEFI partition and rest as ext4 rootfs, all as regular GPT. And no swap although the installer will complain/notify about this. Swap is bad idea on this kind of setup in any case. It is not necessary and if it would end up being used it would just spoil performance.

 

If you already have swap enabled in some installation, it is easy to disable by commenting out the corresponding line in /etc/fstab

 

[paradis]

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

[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.

[paradis]

Just now, b0bb said:

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 consumption gets to 250W

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

That's great to know, as the multi-teraflops number on Titan V intrigued me, and it's so important to hear from someone that owned it to say 3080ti is a more performant option in real world

 

Many thanks

[Miska]

54 minutes ago, b0bb said:

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.

 

And of course for the same W figure, newer generations can do more, because the newer semiconductor processes consume less power. Ada generation (40-series) is manufactured in TSMC 4 nm process. Ampere (30 series) professional cards are manufactured in TSMC 7 nm process and consumer cards in Samsung 8 nm process.

 

Titan V (Volta series) was manufactured in TSMC 12 nm process.

 

Newer generations also have more advanced features, so the generational performance difference is even bigger.

 

Transistor counts also went up, from Volta's 21.2 billion to 76.3 billion in Ada. As the process shrinks, it is possible to fit more transistors in same silicon area. While the silicon is still smaller, Volta's 815 sqmm to Ada's 608 sqmm.

 

[paradis]

nice! as i've that wish that i found a gem just based on the FP64 specification :-)

 

Looks like 4080ti or 4090 still the current best all rounder for me, as these both performed better than A4000/A5000 workstation range from the specs, and price do not differs much

[AudioDoctor]

If you want to spend the money why not get an Ada generation A6000. It's essentially a 4090 with double the RAM, 48GB, and limited to 300 watts power usage.  Or even better, two of them. Although I don't know if HQP would take advantage of them both and all that RAM.

[paradis]

That's too much for me to spent on without knowing the immediate benefit, I am only on a RTX3070 today but just wish to test some filters that my system can't do today, like sinc-L that's known to be a memory hog

 

4090 is absolute the ceiling top I am willing to spend

[AudioDoctor]

2 hours ago, paradis said:

4090 is absolute the ceiling top I am willing to spend

 

Then you have answered your own question. The Zotac 4090 is ~500 bucks less than the ASUS 4090 at my Microcenter and likely just as performant.

[paradis]

indeed, earlier i was just curious at the Titan V claimed multi-teraflops capability which is even higher than 4080/4090 range hence was the original ask :-)

 

Thanks for all the feedbacks

[Miska]

8 hours ago, paradis said:

indeed, earlier i was just curious at the Titan V claimed multi-teraflops capability which is even higher than 4080/4090 range hence was the original ask :-)

 

You could try... But prepare to be disappointed... Just like if you compare number specs of AMD GPU vs Nvidia GPU on similar things. From numbers it would seem AMD beats Nvidia. But in practice, as I have tested, it doesn't perform any better, maybe a little worse. Probably because they have not invested to ROCm as much as Nvidia has invested to CUDA.

 

[Miska]

Some good news, if someone is looking for an inexpensive GPU to complement less powerful CPUs with still fair amount of RAM:

https://slashdot.org/story/23/09/01/1937248/nvidia-quietly-cuts-price-of-poorly-reviewed-16gb-4060-ti-ahead-of-amd-launch

 

[juwa]

My new computer for HQPlayer is ready. It has an Intel CPU i9 13900k and 64GB RAM. The i9 13900k is very powerful and can perfectly execute a number of filters and modulators up to DSD512.

 

I plan on upgrading for maximum possible performance with a GPU. So far, the Geforce RTX 3090 has been the optimum because of its over 10,000 CUDA cores and 24 GB of memory. Does anyone have experience with the RTX 4070TI or do I have to go for the largest model in the new series, the RTX 4090?

 

The price of the RTX40xx models is still very high. A used RTX3090 is significantly cheaper. Would have liked to know if the extra charge was worth it.

 

Thanks for your feedback.

[SPAZ]

Unfortunately, I only have experience with the 4090 but that and the 13900k I can do DSD1024 with no stutters by checking off Cuda offload and multiprocessor cores.

[pis99]

On 9/17/2023 at 11:21 PM, juwa said:

My new computer for HQPlayer is ready. It has an Intel CPU i9 13900k and 64GB RAM. The i9 13900k is very powerful and can perfectly execute a number of filters and modulators up to DSD512.

 

I plan on upgrading for maximum possible performance with a GPU. So far, the Geforce RTX 3090 has been the optimum because of its over 10,000 CUDA cores and 24 GB of memory. Does anyone have experience with the RTX 4070TI or do I have to go for the largest model in the new series, the RTX 4090?

 

The price of the RTX40xx models is still very high. A used RTX3090 is significantly cheaper. Would have liked to know if the extra charge was worth it.

 

Thanks for your feedback.

3090 will be working just fine for Hqplayer. I have 3090 and A6000 in different computers. Not much difference can be found with 13900K cpu. You can reach 1024dsd/sinc-L for 44.1 with proper cooling.

[AudioDoctor]

On 9/17/2023 at 10:21 AM, juwa said:

My new computer for HQPlayer is ready. It has an Intel CPU i9 13900k and 64GB RAM. The i9 13900k is very powerful and can perfectly execute a number of filters and modulators up to DSD512.

 

I plan on upgrading for maximum possible performance with a GPU. So far, the Geforce RTX 3090 has been the optimum because of its over 10,000 CUDA cores and 24 GB of memory. Does anyone have experience with the RTX 4070TI or do I have to go for the largest model in the new series, the RTX 4090?

 

The price of the RTX40xx models is still very high. A used RTX3090 is significantly cheaper. Would have liked to know if the extra charge was worth it.

 

Thanks for your feedback.

 

Given a choice between the two, I would pick the 3090 over a 4070. If you're dying to buy a new Ada generation card, I would get at least a 4080.

[paradis]

14 hours ago, pis99 said:

3090 will be working just fine for Hqplayer. I have 3090 and A6000 in different computers. Not much difference can be found with 13900K cpu. You can reach 1024dsd/sinc-L for 44.1 with proper cooling.

Any difference in term of loading time between 3090 and A6000 @ sincL 1024? What modulator used if I may ask?

[pis99]

15 minutes ago, paradis said:

Any difference in term of loading time between 3090 and A6000 @ sincL 1024? What modulator used if I may ask?

No difference I can tell between 3090 and A6000 on loading time for 1024 sinc-l. Sinc-L loading time at 1024 is always longer around 20 -30 seconds. The difference between A6000 and 3090 is temperature and power consumption. A6000 is cooler and less demanding on power.

I like asdm7ec-super or asdm7ec-light 512fs.

[paradis]

7 minutes ago, pis99 said:

No difference I can tell between 3090 and A6000 on loading time for 1024 sinc-l. Sinc-L loading time at 1024 is always longer around 20 -30 seconds. The difference between A6000 and 3090 is temperature and power consumption. A6000 is cooler and less demanding on power.

I like asdm7ec-super or asdm7ec-light 512fs.

Thanks, that's really important information as it means 3090 has all the necessary grunt to perform at one of the heavy filters, if there's no apparent difference with the exception of power & heat, 3090 can be a sweet spot 

[juwa]

How I understand it, in order to use all filters and modulators, a GPU with a performance equivalent to at least an RTX 3090 is required. Then only the GPUs of type 3090 or 4090 come into question.

 

The 3090 series has a power consumption of 350 watts. The 3090TI and 4090 series require 450 watts. Could be important because of the performance of the power supply and the expected cooling performance.

 

Can anyone who uses one of these GPUs say something about the temperatures on the GPU. Maybe they don't get so hot during operation. 

 

[Quadman]

2 hours ago, juwa said:

Can anyone who uses one of these GPUs say something about the temperatures on the GPU. Maybe they don't get so hot during operation. 

I run a 12900KS cpu with a 280 artic freezer AIO cooler and a EVGA 3090 GPU with 3 stock fans.  I don't overclock except I select XMP 1.  My ram is 4500 Mhz DDR4.  with Jussi's latest 22.04 images, I can do DSD1024 with many of easier EC5 modulators (DSD5ec, ASDM5EC, ASDM5ECv2, ASDM5ECv3, ASDM5EC-light)  No EC 7 modulator works that I have tried without stuttering.  My Favorite filters that always work with the above are Sinc-MGa, poly-sinc gauss-hires-mp and ip, poly-sinc-ext3.  GPU temps run pretty much in the 59-61 degree C range, CPU temps, although I have not installed sensors since a new fresh Ubuntu install, generally run under 64 degrees C.  I suspect even lower than that now as generally only 2 cores of my CPU have significant loads (60-80%)  the rest just cruise.  GPU usage is usually in the upper 20% range and often in the 30% range and occasionally I briefly see usage in the low 40% range.  GPU ram usage is not significant with the above filters, meaning most GPU's with 12GB ram could easily run what I run above.  The longest filters use the GPU ram the most. (Sinc-L, Sinc-LI, Sinc medium and sinc-Long) with those a 24gb card would be useful.

 

I haven't run windows in a while but back when I did I could never do the heavier filters that were successfully running with ubuntu, so I stick with ubuntu. 

[cpcat]

On 9/1/2023 at 5:29 PM, Miska said:

Some good news, if someone is looking for an inexpensive GPU to complement less powerful CPUs with still fair amount of RAM:

https://slashdot.org/story/23/09/01/1937248/nvidia-quietly-cuts-price-of-poorly-reviewed-16gb-4060-ti-ahead-of-amd-launch

 

Will this help me vs what I have now for Sinc-L at 512 DSD?
 

Too slow for my CPU?  I don’t overclock, just use XMP-1 and AVX enabled.

[Miska]

1 hour ago, cpcat said:

Will this help me vs what I have now for Sinc-L at 512 DSD?
 

Too slow for my CPU?  I don’t overclock, just use XMP-1 and AVX enabled.

 

It is possible that it could help. But we don't know until someone tests it.