1700 and 1700x are $279 US.
Thanks for the heads-up. 280 USD is ~360 CAD. Meanwhile it’s on sale at amazon.ca for 340 CAD, and a few days ago at 310 CAD. Tricky bastards.
It’s odd to say this after experiencing fglrx, but I would like to go with AMDGPU open source drivers. The problem is ML mostly uses CUDA and it’s unclear for now how AMD’s ROCm platform will turn out.
Even if I want to tough it, R9 Fury/X/Nano are practically impossible to find, and used RX580 are listed at or above retail prices. It’s just insane, and many even mention mining/modifications.
For the RAM, how much of a real world difference does the speed actually make? I’d like a hassle-free upgrade path to 64GB eventually which is why 16GB DIMMs are appealing. I notice in the QVL that 16GB (double sided) modules run at slower speeds than even their ratings. Is that an intrinsic limitation of Ryzen or will future modules change this?
Not odd at all. AMDGPU is excellent software. I think that you’ll need AMDGPU-pro (the proprietary software) to get OpenCL support, just a heads up.
Depends heavily on the workload. I’d say 15% at the very most, for memory sensitive workloads, assuming we’re comparing 2400MHz to 4000MHz.
I don’t have any experience with the 16GB modules. What I can say is that Ryzen memory compatibility started off AWFUL and has gotten much better over the year it’s been out. AGESA 1.0.0.6 was “the RAM update” that pushed memory compatibility the most.
When it comes to memory and the QVL, you can sometimes get more speed out of your memory than it’s QVL’d for, you can’t always get what it’s rated for and in some rare cases, you can get faster speeds than it’s rated for. It’s really just a shitshow. I’d look around and get some midrange ram that other people have validated to work with your system if you’re concerned.
And that’s another headache if OpenCL is needed. But it seems OpenCL is being superceded by ROCm, which works with AMDGPU, and not -pro. Comparing support for OpenCL vs CUDA is just tragic.
In most except highly saturated cases, it seems like the RAM speed performance difference is in the single digits. Maybe aiming for at least 2400-2666 MHz is a good enough compromise.
For both RAM and motherboard. I have a lot of reading to do.
Good to know. I’m not really following Machine Learning too closely, so these ABI are falling by the wayside for me.
I’d say that’s a safe bet, especially considering your budget concerns. It’s a good place to save without sacrificing much performance at all.
It’s more a prioritization of capacity and upgrade ease, and also a strong aversion to paying inflated prices whether I can afford to or not.
1 Like
I get that. That still makes it a budget in my eyes though.
I forgot to mention, I unfortunately need wifi. Should I look for USB, PCIe, or integrated on the motherboard? Linux support is of course critical, and if PCIe, hopefully not disabling other useful slots.
For the motherboard, I’m looking for very good fan support/control in “BIOS” with the ability to turn fans off completely at idle. How is the Asus Prime X370 Pro for this?
If it’s integrated on the board, go for it. (the WLAN on my X370 Taichi is very good) I wouldn’t go out of your way to buy a board with it unless it’s going to save you money though.
PCIe usually shows up in x1 or x4 form and will only take up one slot.
If you go USB, get an Alfa Networks unit. I use them all the time.
1 Like
For wifi, maybe you have an old router somewhere? you might be able to use that as a bridge.
That is not an option on that board and I’m not sure if any board will do that.
1 Like
The Taichi doesn’t have support for that either.
Hmm, that is an option. Although it would mean more bright LEDs I’d need to tame, and using more power+plugs.
For Asus QFan I saw screenshots showing manually adjustable fan curves (Fan % vs temp), but I guess 0% doesn’t mean totally off?
I’m considering this semi-passive cooler for the CPU, and hoped something similar is possible for the case fans.
The Corsair RMx series PSUs seem well-recommended, and they are apparently also semi-passive up to 40% load.
The fan curve won’t let you go to 0%, I think 20% is slowest. But honestly, unless you buy shit fans at that point they should be inaudible. I use noctua for everything, including CPU heatsink, but it isn’t the most budget friendly option.
Which models? Would you be sleeping beside them?
So it looks like motherboards don’t turn fans off, but some, like the Arctic’s fan, don’t turn on with less than a threshold PWM %. Hopefully their turn-on is graceful. I have even more reading to do.
For the heatsinks NH-U14S and NH-U9S on Ryzen but I have other models in use too.
NF-A8, NF-A9, NF-F12 and NF-A14 in PWM or ULN for fans.
I could easily tune my tower system to run almost noiseless in idle.
I don’t know about the rack case ones but they are quiet for what they are.
Oh, you might want to take a look at Seasonic’s new 650W 80+ Titanium fanless PSU.
1 Like
That is an impressive amount of fans. Noctua must love you. I like the improvised video card support.
Thanks for the model numbers, I’ll look them up. Noctua and BeQuiet fans seem to be the recommended ones.
About the Seasonic fanless PSU, yes that very appealing, but at double the price of the Corsair semi-fanless it gives some pause.
1 Like
That’s not improvised. That’s professional grade. He graduated top of his class at Jerry Rigg University.
2 Likes
That is a Corsair Carbide 400Q without the front panel for more airflow. It’s a nice case but if you really want to stuff two GPUs into your system, maybe look for something bigger?
Nice. I was considering the Fractal Design Define C, which is similar in layout and slightly more compact. In comparison it seems the 400Q has a bit more room for longer video cards. Do you have a preference between the two, or other suggestions?
I don’t plan to use two GPUs now, but I wanted to keep the option open for later. If forgoing the option of multi-GPUs would give a significant advantage in other ways, I might consider that too.