a NAS for backups, snapshots of VMs, and VM disks.
mGPU Computational Power for some of the applications I need (at least 4 GPUs)
VM Hypervisor to Emulate multi-computer lab Scenarios off a single machine.
I’m confused about choosing a platform right now.
Build a TR-PRO 3000/5000 from scratch up
OR
Wait for Ryzen 7000 and push my current 5950X along with 2 GPUs to this use case and add 2 GPUs on my new build as the application that requires multiple GPUs can run in a distributive manner by dividing work to GPUs over the network.
And upgrade networking to 10 Gbps
@GigaBusterEXE could you go into more details on what this curve optimizer is? I know I haven’t done it because I have left all the BIOS setting at factory settings except the settings that have to do with virtulation.
Would you mind quantifying the effect a bit? For example, in Cinebench R23 multi-thread test, what % performance gain you get after curve optimizing vs default curve setting.
I’m probably gonna make a thread on it some point next week but to keep things short and on topic
With Ryzen 5000 they introduced the curve optimizer into PBO2
PBO allows you to boost higher than stock boost by unlocking the tdp power limits
As chips boost they need more voltage and amperage the higher the frequency, and even though you are unlocking the power limits, there are still reasonable limits you have to stay inside
Another behavior of frequency is that the voltage increase is not linear, with each bump in frequency the bump in voltage needs to be higher than the last, that’s why it’s called a voltage frequency curve
But chips aren’t finely tuned from factory they just get a generic bin, so generally they’ll get more voltage than they need
So the curve optimizer let’s you shift the curve on the voltage side to give it less voltage per clock
Volts x amps = watts
So if you give it less volts then you’ll get get less watts at same amps
Which means your frequency ceiling is much higher now especially all core avx clocks which demand a ton of amps
TLDR curve optimizer let’s you run your CPU leaner so it can run faster
Might have been like 56K points and like 48-49k without
5975wx stock is 280w on a noctua single tower cooler
But on a custom loop, just pbo it was 500w, with curve, it was 410w
Time of writing sage does support pbo and ASRock creator does
There is no Ryzen 7000 option for that workload. You want a platform with lots of PCI lanes and slots for all the GPUs, NIC, etc. Ryzen can’t deliver.
So your options are EPYC, Threadripper or Intel (Xeons). EPYC has much cheaper CPUs if you don’t rely on extreme clock speeds. Most server stuff usually wants cores rather than high clock speeds because it runs several things in parallel.
An EPYC 7443P might be plenty for your use case at much lower cost for the CPU while still having 128 lanes and 8-channel memory.
First off, there are no B550 or X570 motherboards that will support a 4 GPU setup, period - The motherboards assume two or three 3-slot GPUs at best. I assume the same will be true for the 7000 series. This does not have to be a deal breaker though, it is theoretically possible to set up a four x4 compute card by clever use of mITX and single slot water cooling, however this is a very niche case application and does require a specialist board, possibly with some EE skills required. x4 is good enough for all but the most demanding workloads, double so for PCIe 4.0 and quadruple so for PCIe 5.0. Literally.
In the long run though, I think it’s better to just go with a cheap-ish EPYC board and 16-32 cores depending on workload. It doesn’t have to be the latest 64 core available. Clock speeds over 2 GHz are irrelevant in your use case, PCIe lanes and cores however, those matters a lot for your use case.
Thank you for all your suggestions. I’ll keep a note of all this for now. Due to budget constraints, I might have to push the build dates further. Will keep you all posted on whatever I end up building.
Thanks for the suggestion!