Thinking about purchasing a monster machine. Need some advice before I do so.

[t33to]

The universe is basically telling me that I need to take PC building a bit more seriously, perhaps even as a career. I've been that guy for as long as I can remember that's always helped people in my community, friends and family get the machine they need for the price they can afford. Lately I've been charging for these services and everyone seems happy with my services. Ok, end preamble.

 

What I'd like to do is build a brand new machine for myself that basically contains the latest and greatest 9900k/2990, 1080ti/2080ti and put it all on a motherboard that has a zillion options so that I can downclock it to represent the performance of lesser machines.

 

I have a lot of clients that come my way that say I need a machine that can do X and my budget is $Y. I'd like to be able to quickly spec the best machine I can for them for $Y and then downclock my machine to represent with a fair level of accuracy how that machine will perform in X scenarios.

 

What parts would you recommend? Specifically, what kind of motherboard would make this stupid easy? Is there any motherboards out there that would permit me to save 20-30 different downclocked profiles so I could easily turn my system into something as weak as a Pentium?

 

I presume all of the GPU downclocking will happen at the OS level?

 

Thanks in advanced! Cheers.

[Fasauceome]

The downclocking of your system is ok in theory, but there's a lot it misses. How certain cooling systems would work, power delivery, acoustics, etc. As a system builder, it's never helped me to try to emulate a system with hardware I had on hand. Sometimes you'll be upgrading a Skylake CPU and a 1070, and there'll be no real way to simulate that. It's best to simply learn the hardware through experience, but showing a client a "simulated" system as some form of approximation does serve as a good selling point.

[t33to]

1 minute ago, fasauceome said:

Sometimes you'll be upgrading a Skylake CPU and a 1070, and there'll be no real way to simulate that.

Would you mind elaborating on what you mean here? If I have a more boss machine isn't it simply a matter of restricting the system to emulate the lesser?

[Firewrath9]

15 minutes ago, t33to said:

Would you mind elaborating on what you mean here? If I have a more boss machine isn't it simply a matter of restricting the system to emulate the lesser?

but you can't simulate a 1070 with a 2080 ti. the cuda arcitechture is different, it's like having a lambo, and setting the max speed to 60 mph to simulate a truck.

[t33to]

1 minute ago, Firewrath9 said:

but you can't simulate a 1070 with a 2080 ti. the cuda arcitechture is different, it's like having a lambo, and setting the max speed to 60 mph to simulate a truck.

Well if I look up benchmarks for the 1070 in various titles, is it not possible to restrict the 2080ti to roughly emulate how the 1070 will perform?

 

Limiting a lambo to 60mph is pretty much exactly what I'm trying to do. I just want the client to be able to sit down and use the downclocked machine and then when they get their chosen system it's within 10-15% of the performance they tested.

[Firewrath9]

2 minutes ago, t33to said:

Limiting a lambo to 60mph is pretty much exactly what I'm trying to do. 

but the feel of a lambo is different, the seats, the acceleration, etc. it djust doesnt work.

[t33to]

Just now, Firewrath9 said:

but the feel of a lambo is different, the seats, the acceleration, etc. it djust doesnt work.

Well in this case I'm mostly just referring to the time someone spends waiting for something to load/finish, the responsiveness of the system and of course frame rates. Do you truly think it's not possible to build a rig capable of being downclocked to represent a lesser machine with 10-15% accuracy? Honest question.

[Fasauceome]

4 minutes ago, t33to said:

Well in this case I'm mostly just referring to the time someone spends waiting for something to load/finish, the responsiveness of the system and of course frame rates. Do you truly think it's not possible to build a rig capable of being downclocked to represent a lesser machine with 10-15% accuracy? Honest question.

In order to do that you'd need to change the clock speed and power for every single benchmark individually. You can't simply run a 2080 ti at 75% power and expect it to be the same in every stress test. Comparison would be torque of a truck vs a lambo. Uphill changes the whole game. There's suspension too, and weight. You have to fine tune it all to get the same "emulated" result.

[Fasauceome]

8 minutes ago, t33to said:

10-15% accuracy

This a very inaccurate amount, you'd want to emulate something like 90% for it to be practical.

[t33to]

12 minutes ago, fasauceome said:

In order to do that you'd need to change the clock speed and power for every single benchmark individually. You can't simply run a 2080 ti at 75% power and expect it to be the same in every stress test. Comparison would be torque of a truck vs a lambo. Uphill changes the whole game. There's suspension too, and weight. You have to fine tune it all to get the same "emulated" result.

 

Yes that's why I asked for a profile system. So I could spend the time to tune it and save the profile and switch easily between it all.

 

9 minutes ago, fasauceome said:

This a very inaccurate amount, you'd want to emulate something like 90% for it to be practical.

 

Yes I mean 10-15% off the actual lesser machine. So 85-90%

[Fasauceome]

Just now, t33to said:

 

Yes that's why I asked for a profile system. So I could spend the time to tune it and save the profile and switch easily between it all.

 

 

Yes I mean 10-15% off the actual lesser machine. So 85-90%

Ok well here's the problem:

There are still variations within graphics cards themselves. One 1070 is not the same as another. There's silicon quality variance, up to as much as 5% in many cases, and there's ambient temps of your environment, and there are all the different SKUs of the card itself. Unless you plan on making hundreds upon hundreds of profiles tuning memory clock speeds, core clocks, fan curves, and power limits all for just projected results that won't even be 100% accurate (cause they can't be) then there's no reason to really invest in that strategy. The sheer fact that different graphics cards use different technologies is enough to throw an enormous wrench in this plan, such as GDDR5 vs GDDR6, cuda cores vs SMs, etc.

[t33to]

1 minute ago, fasauceome said:

Ok well here's the problem:

There are still variations within graphics cards themselves. One 1070 is not the same as another. There's silicon quality variance, up to as much as 5% in many cases, and there's ambient temps of your environment, and there are all the different SKUs of the card itself. Unless you plan on making hundreds upon hundreds of profiles tuning memory clock speeds, core clocks, fan curves, and power limits all for just projected results that won't even be 100% accurate (cause they can't be) then there's no reason to really invest in that strategy. The sheer fact that different graphics cards use different technologies is enough to throw an enormous wrench in this plan, such as GDDR5 vs GDDR6, cuda cores vs SMs, etc.

I understand, and thank you for your perspective, but the alternative is just to look up benchmarks which as you say will have natural variance anyways but the client cannot have any kind of real experience with the machine before it's paid for and built.

 

And yes, the plan is to make as many profiles as needed. A large time investment to be sure, but any career or business usually is.

 

If I go and look up benchmarks for a given CPU/GPU combination can I not just keep downclocking my system until I achieve those results. I'm not referring to matching clocks for clock verbatim for the sake of it. I'm just talking about emulating real world performance, frames, load times, render times so the client themselves can witness it personally.

[Fasauceome]

2 minutes ago, t33to said:

I understand, and thank you for your perspective, but the alternative is just to look up benchmarks which as you say will have natural variance anyways but the client cannot have any kind of real experience with the machine before it's paid for and built.

 

And yes, the plan is to make as many profiles as needed. A large time investment to be sure, but any career or business usually is.

 

If I go and look up benchmarks for a given CPU/GPU combination can I not just keep downclocking my system until I achieve those results. I'm not referring to matching clocks for clock verbatim for the sake of it. I'm just talking about emulating real world performance, frames, load times, render times so the client themselves can witness it personally.

Well, if you feel not only that you can obtain that information but also meaningfully and accurately compile it, then be my guest. I don't think the insanely overpowered rig would be bad, just bad for this purpose, and if you discover your initial goal is out of reach, you'll still have a badass PC to show for it.

[t33to]

Just now, fasauceome said:

Well, if you feel not only that you can obtain that information but also meaningfully and accurately compile it, then be my guest. I don't think the insanely overpowered rig would be bad, just bad for this purpose, and if you discover your initial goal is out of reach, you'll still have a badass PC to show for it.

Haha, very true.

 

So can you recommend any hardware that would make this as straight forward as possible? Motherboards? Memory?

[Firewrath9]

2 hours ago, t33to said:

Haha, very true.

 

So can you recommend any hardware that would make this as straight forward as possible? Motherboards? Memory?

Budget? prob at least 32Gb of ram, and a 10c+

[t33to]

1 hour ago, Firewrath9 said:

Budget? prob at least 32Gb of ram, and a 10c+

$3500?

[Firewrath9]

17 minutes ago, t33to said:

$3500?

k i'll build you one tommorrow, got to sleep.