Best Nvidia CUDA Card for HQPlayer

[NellyWhads]

If anyone in this thread could help me with a quick test I would appreciate it:

I am thinking about getting a GPU to handle up-sampling using HQP, but my goal is to have a completely passively cooled system. The case (HDPlex H5) can handle 95W TDP per heatsink, and has two heatsinks. I can use one for the CPU (to be decided) and one for the GPU. I will be getting the Holo May soon, so the higher the sample rate the better (PCM 1.5M and DSD 512 would be golden. DSD 1024).

 

I do not think I can count on a CPU-only build due to the limited heat dissipation of the cases I want to use (95W TDP). Can push it to 125W TDP, but I would really rather not.

 

Could someone try TDP limiting their A4000 and CPU to 90W and see what is the upper limit of up-sampling possible using one of the intense filters (sinc-L, ext3, xla)? @Cornell77 and @Rovo if you see this and could maybe help with an experiment in your systems, I would truly appreciate it.

[NellyWhads]

9 minutes ago, AudioDoctor said:

 

This is what I built in. The case is not easy to get, you will have to wait, and it's a pain in the ass to build in. But once that's done you're good to go. I have an AMD 5900X and a 3080Ti in mine, undervolted and 100% passively cooled.

 

https://www.monsterlabo.com/the-beast

Woah. It's good to know something like this exists. It is definitely overkill. Like building a full gaming PC.

From my time with Audiolinux and HQPlayer so far, I can sinc-L up-sample to 768k with no issues. Just on CPU using an i7 7700k. How much I want to spend to get up to 1.5M and DSD 512.... is another question.

 

Thank you for sharing this option, I will keep an eye on the used market. Maybe I'll get lucky.

[NellyWhads]

5 minutes ago, AudioDoctor said:

 

The company does have a Discord. I haven't looked to see if anyone is selling any used cases, but the company will build the thing for you, and ship it to you from the UK where they are based as well.

 

I would not call it overkill either. If you want to push HQP to the upper end of its capabilities you're going to need some power to do so. Additionally, I am pretty sure the new generation of CPUs and GPUs will be too hot to use passively at all without going to massively overkill extremes that make this look tiny.

From my research so far your final comment here definitely checks out. There are a lot of advances in manufacturing processes, etc., but in order to build something which will last as my last build did, it will need to be relatively top spec.... welp.

 

The A4000 is such a good proposition. If they would spec a build with something like the A6000 in it for me, then I think it would be a good way to go. I much prefer to build my own system by hand though. Let's see what they say, I'll reach out.

 

In the mean time, if anyone can help with that A4000 TDP limit test, I would definitely appreciate it. Just to see what's reasonably possible in the H5 case.

[NellyWhads]

27 minutes ago, AudioDoctor said:

I found it, Quadro 4000 GPU and a PCI card.

https://www.monsterlabo.com/post/01-sep-assembly-of-the-beast-w--rtx-4000-rednet-pci-card-audio

This is cool, but the P4000 only has 8GB of memory and is listed on their site as a compatible card. A4000 has 16GB and would be ideal to work with for DSD upsampling.

If I were to scale back the requirements to just 1.5M PCM up-sampling and DSD passthrough, could I get away with just a 95W TDP CPU? What's the fastest/highest spec'd CPU I can run with that thermal limit (I'm Googling this as we speak as well). No Apple hardware for me. Need linux base so I can spend hours and days tuning kernel parameters to perfection in a LPSU + no-moving-parts build.

 

Edit:

 

28 minutes ago, AudioDoctor said:

I don't know how the May DAC handles DSD being an R2R DAC. Some R2R DACs convert DSD to PCM before the DAC stage.

 

The May has 2 separate pairs (stereo) of R2R implementations. One L/R pair for PCM, and a second standalone implementation for DSD. My "realistic" goal is to feed each implementation the highest possible sample rate (PCM 1.5M or DSD 1024) given the appropriate source content. 

 

Given your feedback and a better understanding of hardware requirements I'm updating my ask to the following:

A system that can up-sample PCM to 1.5M and DSD to 1024 given the matching input format. The more complex the filter it can handle, the better, so that as Jussi comes out with new ideas in the future, I don't have to worry about hardware components and can focus on other unnecessary diminishing returns (such as motherboard clock upgrades and LPSU).

[NellyWhads]

13 minutes ago, pis99 said:

I currently have 12900KS AIO 360 with RTX A6000. The A6000 gets very hot when running Sinc-L filter at 1024DSD. I have to find out another cooling solution for A6000 so I feel comfortable for long playing time. A4000 maybe different though.

Thank you for your input - given your feedback and @AudioDoctor's, I think it's more appropriate to target within the same family of source content. I can't afford an A6000 anyways lol. And if I get an A40, I might as well just use it for Deep Learning instead of up-sampling using filters lol

[NellyWhads]

Just now, AudioDoctor said:

 

If that were the case, a 5600x would be ideal. Perhaps a 7600X if you want a new one. It has a TDP of 105 watts, but I don't think you would be running it near max performance for upsampling to 1.5mhz and DSD pass through.

How do I find spec requirements for DSD 64 -> DSD 1024 up-sampling?

 

If the 7600X can handle it, I can just stick with my current Streacom FC8 case (87W TDP) and upgrade CPU + MOBO + RAM and focus on the next steps (Clocks, LPSU, etc.) instead of rebuilding in a new case.

[NellyWhads]

4 minutes ago, AudioDoctor said:

I don't think there are specs available as the filter and modulator combinations are too numerous. I just know that the higher you go, the more power is needed to crunch through the math.

Hm, I was thinking if you're starting with DSD64, interpolating to a higher 1-bit sample rate would be significantly less computationally expensive compared to starting all the way at RedBook. Perhaps @Miska could provide some insight based on implementation of the most taxing filters.

 

Edit: I wonder if there are any statistical/machine learning methods cooking in his kitchen... an inference engine for up-sampling like Sony's DSEE could be pretty spiffy cool... maybe I should look into it

[NellyWhads]

 

3 minutes ago, AudioDoctor said:

 

I have no idea, as I generally don't upsample the DSD, I just leave it as is and only upsample the PCM.

Now hold on just one second... if you're only doing PCM (I assume with something that's heavy hitting like a sinc- or -ext3 filter), and a 5600x could probably handle it.... why the passively cooled beast? lol

[NellyWhads]

Just found this (looks like DSD64 -> DSD1024 will still need a GPU):
 

 

[NellyWhads]

4 minutes ago, AudioDoctor said:

 

I upsample PCM to DSD. I do not upsample DSD to higher rate DSD.

Got it. If you have some time to experiment, would you mind trying out CPU-only DSD 64 to DSD ___, whatever your system can handle?

If your DAC is capped at some limit, that's understandable. Also, I completely understand if it's too much of a hassle to reconfigure and invest time to test. I've asked the other author to share his thoughts as well in the linked thread.

[NellyWhads]

3 minutes ago, AudioDoctor said:

 

It's a click of the "CUDA" button. I can do that but not tonight.

Many thanks,

 

I'll hold off on the local HDPlex H5 used deal for a few days and see what Jussi and others say based on their experience. The less components to worry about the better, but being able to get "the best possible sound" is something we continue to hunt.

[NellyWhads]

1 minute ago, pis99 said:

I also have a 3090 card that works equally well. Can do Sinc-L at 1024DSD! 3090 used one even for mining could be way cheap than A4000 and it has 24G mems for Sinc-L. 

Would you mind sharing the load your card experiences when up-sampling from Red Book to DSD 1024? I would take a shot with the 3090 (given the memory availability) if I could cool it passively in the H5 (meaning it needs to operate at 95-100W TDP). This is less than half of the spec of the card.

[NellyWhads]

1 minute ago, AudioDoctor said:

That will NEVER happen.

Well there we have it. Only really possible in The Beast. 

 

Seems the right way to approach this is to build a 7600X into the Streacom FC8 and spend some time with PCM 1.5M + DSD pass through. After spending some time with the May (and maybe testing some offline up-sampled files on DSD 1024) I can see what to do next.

 

Only question that remains is what CPU can do on its own given DSD input stream and DSD max up-sample out. I'll let you get back to me on that when you have some time to experiment - no rush.

 

Many thanks

[NellyWhads]

1 minute ago, AudioDoctor said:

 

NO doubt. I like sinc-L, but I like poly-sinc-gauss-long and poly-sinc-gauss-hires-lp, more as I find them a bit more musical with minimal loss of the attack and speed of sinc-L. 

 

However, this is all personal so there are no wrong answers.

I think my exploration so far with filters (even in this discussion) is to have the option to run the system in a given configuration, and know that it will be able to handle some reasonable extra stress as Jussi gets more and more crazy. Even if those filters don't end up being what I use all the time.

[NellyWhads]

29 minutes ago, Miska said:

Every time you double the sampling rate, both filter and modulator loads roughly double. So you can see that the load increases rather quickly as function of sampling rate.

I guess it makes sense for source rate; more samples to process = higher memory requirement. I the CPU can keep up then it's just a memory concern. 32GB system memory and you're happy for DSD 512. Everything else is just CPU processing speed per W of power.

 

Can HQP benefit from using a fast enough GPU with DMA? Or is hardware GPU memory a requirement due to the streaming implementation? Maybe in the future we can get away with lower clock speed/TDP GPUs which have DMA available if that is the case.

[NellyWhads]

3 hours ago, Rovo said:

@NellyWhads

I am running HQPLayer with sinc-L and ASDM7EC512+fs. DSD512 is possible for multiples of 44.1kHz, but for multiples of 48kHz I get dropouts every approximately 45 seconds, so I have limited this one to DSD256. Power draw from the wall is approximately 180W, this is for the motherboard, CPU, RAM, one SSD and the GPU (fans powered separately). According to the NVidia tool the GPU is using approximately 70W, so I expect the CPU to draw approximately 95 to 100W. I am using a BeQuiet Dark Rock Pro 4 and the CPU temperature is approximately 60 degrees Celsius. I hope this information is helpful. 

 

At the moment I am designing my own case to passively cool the HQPlayer upsampling PC. I will be using larger heatsinks than in the HDPlex H5 case. Perhaps because I can, but I do not know if it is necessary. 

Thank you very much Rovo, this helps a lot.

 

Would it be possible to test how your load looks/what stable upsampling you can achieve if you start with DSD64 source content? The filters and modulator you are using are just right for the magnitude of system load test, so no change is necessary from those preferences.

[NellyWhads]

1 hour ago, pis99 said:

Here you go! 12900KS/3090

Cool! Roughly 150W GPU dissipation needed it seems in total on the 3090 (using some good judgement). Less than I expected to be honest.

 

Any chance you can try CPU only? After speaking to Jussi I'm currently targetting the i13900T when it is repeased for a CPU-only build. Can give it extra memory if there is a chance it can handle DSD -> DSD on its own.

[NellyWhads]

On 11/25/2022 at 5:53 AM, lsantista said:

Did you have any chance of trying this?

 

I have made no moves to build a server yet.

 

I am of two minds:

1. Take a lot of folks' advice and separate the Server PC from the Up-sampler PC. The added cost will come particularly from the network cards (JCAT NET XE x2) and matching power supplies.

2. Wait for the i9 13900T and hope I'm happy enough with it doing double duty for Roon and HQP.

 

Either way, I'm waiting for a 'sign' in the form of local used part availability, especially an HDPlex H5 somewhere in Canada.

 

All in good time I imagine. I'm a patient person

[NellyWhads]

9 hours ago, AudioDoctor said:

 

The server side doesn't need to be high power at all. I am running Minimserver on a 2011 Mac mini i5 running Ubuntu server that never reaches 2% CPU usage just serving up files to the HQPlayer machine.

 

I already have everything set up for a fully passive server PC, so that part cost I'm not counting (even though they're nice to haves. 4x 8tb SSDs are not necessary.. but it's already there. It's just the added network cards and power that will be the cost incurred for one approach over the other.