In the next coming months, roughly between 5-7 months, I plan to buy a really high perf PC with the purpose of running a lot of simultaneous VM’s, networking labs in GNS3, and very likely some introduction to AI. I would also like as a bonus, to play games in perhaps a windows VM or multiboot. Currently, I’m thinking:
7900XTX (Particular model dependent on prices, driver updates, etc.)
7950X
2TB*3 NVMe drives or some other config, whatever is best price to performance
MPG X670E
Arctic liquid freezer ii 360mm
Either 96GB (48*2) DDR5-6000 or 128GB DDR5-(idk) dependent on whether the 128GB can perform fast enough.
1000W Seasonic Platinum PSU
Lian Li Licool 3
Any thoughts on bottlenecks? I have tried looking into the memory as much as I can but I’m still unclear if the 128GB in 4 sticks config on AM5 is going to be too slow relative to the rest of the build, because I don’t know to what extent BIOS updates can help this. I don’t mind 96GB if it means higher performance.
Budget is roughly 3000ish euros, 4000 euros at a huge push if its totally worth it, ANY ADVICE OR CRITICISM GREATLY APPRECIATED
Wait a little while. PCIe Gen 5 NVMe drives are coming (no firm ETA as of yet), possibly before the end of this year? GPU market is also fairly fluid, so the 7900XTX may not be your best option then. I’d reckon 4 sticks DDR5 will be possible w/o performance issues by then too, as drivers (microcode) have matured more and dev’s have a better understanding of the DDR5 platform.
Also, new/upgraded platform launches are coming, again no firm ETA, but both Intel and AMD are working on it. That may mean a decent cost savings for the ‘old’ platform(s) like DDR5 and now happening to DDR4. Mind, it’s still possible to build a high end rig on DDR4 (Threadripper, -Pro!) that outperforms DDR5 systems, at a lower cost.
Won’t happen. Consumer boards only have two channels. Technically you only have 2 slots. 2DPC will always be slower. So get two DIMMs. 4 DIMMs is a desperate move if you really need the capacity.
Boards only have a single 5.0 M.2 Slot. I doubt this will change much with future boards (signal integrity won’t bring 5.0 to the other slots). And a single slot isn’t worth a damn if you want multiple drives and RAID them.
System looks solid as far as I’m concerned. I went with 7900xt and 7900x Ryzen and 64GB memory. Pretty much one floor lower than my system.
Memory…if you need capacity, you really need it. And clock speed is of secondary concern. Couple % drop in gaming or using swap all the time? I always go for capacity.
Relevant to the GPU market comment, do you mean that perhaps a high memory 7800xt or 7700xt or something will be available? It is something I’ve thought of, seeing as videogames aren’t my number one priority. Gen 5 NVMe looks so awesome too, so that is a great point to look out for. What really piqued my interest is your comment on a threadripper system on DDR4. Afaik threadripper costs way more than my budget allows, any elaboration? Also, isn’t that a bad choice down to single core performance? I was thinking 7950x because it doesn’t have bad single core and also has great multicore.
Data permanence isn’t a huge concern of mine, so RAID wasn’t even an option I considered (for data safety), only for performance, as in some RAID setup across multiple drives that purely goes for speed. I don’t mind losing any data on my personal PC as i will also be buying a NAS a little down the line for sensitive stuff.
There’s quite a few AM5 boards with 2 5.0 NVMe slots and a few with more (but those start eating into the GPU lanes). AM5 CPUs have 24 usable lanes, so 16+4+4 CPU-connected lanes.
With a NAS as backup, just do RAID0 across all disks. If you want local redundancy, RAID10 is really nice for 4 drives.
I got 3x drives as RAID0 and a SATA SSD boot. Throughput is insane. I just don’t have need for more than 20GB/s. Writing more than a TB at once sees some slowdowns, but that’s just how consumer drives work.
Last time I checked it was one slot and some using x4 GPP lanes for a second NVMe (and often located beneath the GPU, so you also get a cooling problem). If the board has 2x 5.0 M.2 via CPU, you don’t get Thunderbolt or a x4 PCIe slot, means you only have only a single slot on the board. And maybe x1/x4 chipset slots if manufacturer is “generous”.
I don’t think is PCIe 5.0 NVMe is particularly useful because throughput of 3-4 4.0 drives isn’t perceived as limiting by anyone. With less cost and easier cooling. Random I/O (which is why we love NAND Flash) is the same on all generations.
RDN3 may come with vfio reset bugs. RTX 4000s are better.
MSI x670e Carbon? That board is so good that I bought 2.
Don’t go with 4 dimms 128GB. You would be lucky to get 5200MT/s with 128GB.
For 48GB dimms, hynix dies are preferable, Althought they are still rare as today. They are rated way over 6000 with 2 sticks. Even with 4 dimms, they should run 6000 just fine.
Even with 4 dimms, they should run 6000 just fine.
192GB of 6000(+) DDR5?? I thought they would have the same issue with the CPU memory controller running more than 2 sticks.
RDN3 may come with vfio reset bugs. RTX 4000s are better.
What do you mean by vfio reset bugs? Also the reason I prefer AMD’s GPUs is the memory abundance for less cost and also that they more or less compete in rasterization against cards far more expensive from NVIDIA.
Regarding 4 dimms 128GB, the limitations are on the dimms.
The example is that 4x16GB can run 6200 just fine, but not with 4x32GB.
The new Hynix 24Gb chip is so good electrically that, even with dual-rank dimms, it does not behave any worse than single-rank dimms. (At least with the current Ryzen 7000)
In regards to the VFIO reset, I’m not all too bothered because worst come to worst, I just dual boot, and plus the thread mentions the 7900XT, not XTX though I suppose the problem may quite likely be the same on both.
That’s actually really interesting, I’d love to take a deeper dive into the performance differences between a 32GB stick and a 48GB stick. Maybe it’s something to do with the data distribution? I’m not particularly clued in when it comes to how newer memory works, so excuse the ignorance.
I noticed, but high end GPU’s are also useful for AI stuff. Alternatively, consider the specialised AI GPU’s from nVidia, but none of those would fit in your budget. Having said that, I don’t have a crystal ball and I’m not some well-known YT-er with a double digit million subs, so I have no idea what nVidia and AMD have in the works on the GPU front. I do know they are working on their respective next gen GPU’s, but no clue nor indication how far away their (planned nor actual) release date is.