I’m running a Proxmox home lab server with my trusty 1920X and ASRock X399M Taichi. Can’t really complain, but one does visit eBay to take a look at things from time to time, and it seems the 1950X is available for around $350. I wouldn’t mind spending that amount for a few more cores.
Looked at the 2950X and the cheapest is $660+ (plus shipping and import tax). Every review I saw the 2950X barely moved the needle forward, so it doesn’t look like a winner. I was curious because it has a higher boost, and I’m mostly running short bursty loads, but double price is surely a no-go.
Out of curiosity I looked at the 2970WX which is pushing $870+. Who doesn’t like more cores, but the reduced clock speed tells me it’s more for a marathon, and not a 100m dash.
I ditched my TR 2950x for a dual Xeon 2683v4 and couldn’t be happier! 32 core 64 thread, compared to 16/32, and big ram is a lot easier. The Xeon works with registered buffered ECC RDIMMs and doesn’t need exotic UDIMMs like the Ryzen. Also the board has 24 dimm slots, so I was easily able to jump from 128GB non ecc to 512GB ecc.
At first glance, the 2683 doesn’t look like much compared to the 4ghz 2950, but I tell you, the xeon multitasks SO MUCH BETTER… I can have several virtual machines running stress tests maxing out their assigned CPU and RAM, and other VMs don’t notice, and the host doesn’t notice. The threadripper would start getting all jittery if one VM was super busy, it would effect all the other VMs.
I don’t know why exactly, other than the Threadripper is a workstation CPU designed focused mainly on intense lightly threaded workloads, and the Xeon is a server CPU focused on heavy multithreaded workloads.
Bought the Xeon setup off ebay for less than I’m selling my TR setup for. It took quite a bit of patience to hit the right deals for the motherboard, and CPUs, so it’s not something you can rush out at do with a few clicks. The motherboard I got for $279 is now selling for over $600 on ebay, but I’m sure it’ll come down again.
Biggest gain, moving from Gen1 to Gen2 TR, would be better memory support
The price discrepency [1950 vs 2950] does throw a wrench, if that cost be justified
Going for them WX chips, could lead to possible trouble
Those added cores, do not have a hard connection, with the IO sub-die / Memory
The infinity fabric can get put to a struggle, for all that cross chatter
Do note that added power draw will be noticeable, compared to the __50x
I’d go for the higher base core clock, over that blanket of cores
Also the __50x bump can [effectively] maintain, entirety of your current hardware
Currently have 4x16GB Kingston 2666 CL19 ECC memory which I have no problems with, and the board can only take up to 64GB. So I guess whatever “better memory” the Gen2 could provide, I could barely take advantage of. These Kingston Micron E-die sticks can go really fast, I’m sure I could up the speed if that was ever my limiting factor.
So the 2950X not being significantly faster, or cooler, or more energy efficient than 1950X, around 2x price is not warranted.
Maybe something about the CCX layout and the Infinity Fabric? That CPU is two CCDs, each with two quad-core CCXes, right? If your workload is saturating the Infinity Fabric by a VM being spread over more than four cores, presumably it could introduce latencies in another VM that is also trying to access RAM or another CCX as well even if the two are on paper on entirely different cores? Just a guess.
It’d be interesting to see how the 5950X would compare, with the 8 core CCXes and the substantially larger cache.
If only AMD had been more generous with the PCIe lanes.
My use cases require a lot of IO bandwidth, and the AM4 platform is severely lacking. The 5900x and 5950x provide a lot of compute for the price, but very little bandwidth. Again, chips optimized for single user workloads, the only IO is one flash memory device and one or two GPUs.
God yes. When I was on Sandy Bridge I had two GPUs in SLI both in x16 and an m.2 (on a PCIe card I still have in a drawer somewhere xD) and still had PCIe lanes to spare. Then I upgraded and had to go to x8 for the GPUs and still had to juggle between the CPU and the chipset. I know the x16 matters less for GPUs on PCIe 4 than it did on PCIe 2, but it’s never good to see features being downgraded on a £700 product 10 years apart.