[Aurornis]

You’ll have to be more specific about your price range. There are a lot of server and workstation chipsets/platforms that will have a large number of PCIe lanes, but you will pay for them.

I really suggest not seeking a lot of PCIe lanes unless you really need them right now, though. The price premium for a platform with a lot of extra PCIe is very steep once you get past consumer boards. It would be a shame to spend a huge premium on a server board and settle for slower older tech CPUs only to have all of those slots sit empty.

It’s a good idea to add up the PCIe devices you will use and the actual bandwidth they need. You lose very little by running a GPU in a PCIe x8 slot instead of a full x16 slot, for example. A 10G Ethernet card only needs 1 lane of PCIe 4.0. Even fast SSDs can get away with half of their lanes and you’ll never notice except in rare cases of sustained large file transfers.

[shadowpho]

>Even fast SSDs can get away with half of their lanes and you’ll never notice except in rare cases of sustained large file transfers

Sorta yes but kinda the other way around — you’ll mostly notice in short high burst of I/O. This is mostly the case for people who use them to run remote mounted VM.

Nowadays all nvme have a cache on board (ddr3 memory is common), which is how they manage to keep up with high speed. However once you exhaust the cache speeds drop dramatically.

But your point is valid that very few people actually notice a difference

[wtallis]

You're pretty far off the mark about SSD caching. A majority of consumer SSDs are now DRAMless, and still can exceed PCIe 4.0 x4 bandwidth for sequential transfers. Only a seriously outdated SSD would still be using DDR3; good ones should be using LPDDR4 or maybe DDR4. And when a SSD does have DRAM, it isn't there for the sake of caching your data, it's for caching the driver's internal metadata that tracks the mapping of logical block addresses to physical NAND flash pages.

[shadowpho]

Here’s a page comparison 8 modern SSD cache, notice how they all fall off once the cache is full.

https://pcpartpicker.com/forums/topic/423337-animated-graphs...

[wtallis]

That has nothing to do with DRAM; that would be completely obvious if you stopped to think about the cache sizes implied by writing at 5-6GB/s for tens of seconds before speeds drop. Nobody's putting 100+ GB of DRAM on a single SSD. You get at most 1GB of DRAM per 1TB of NAND.

What those graphs illustrate is SLC caching: writing faster by storing one bit per NAND flash memory cell (imprecisely), then eventually re-packing that data to store three or four memory bits per cell (as is necessary to achieve the drive's nominal capacity). Note that this only directly affects write operations; reading data at several GB/s is possible even for data that's stored in TLC/QLC cells, and can be sustained for the entire capacity of the drive.

[shadowpho]

Interesting, that makes sense, but my point still stands: all 8 of those devices fall down 4-8x in speed at some point, meaning for sequential transfers speed falls off and can handle 4-8x less lanes

[wtallis]

The performance drop due to SLC caching only applies to writes. Sequential reads (and often, even random reads at sufficiently high queue depth) will still more or less saturate the PCIe link. Most workloads and use cases read a lot more data than they write.

[crote]

For a lot of enthousiasts the problem genuinely is in the lane count.

For example, putting a cheap 2nd-hand dual 25G NIC in a DIY server is quite attractive. But those are using PCIe Gen3 - so unless you're giving it 8 lanes it is being bottlenecked.

Same on the storage side: that PCIe Gen4 x4 slot might technically have enough bandwidth for four SSDs in most storage applications, but the board doesn't support bifurcating it.

The total platform bandwidth is plenty, it just isn't available in the way I want to use it.

[rkagerer]

Sky's the limit. (Short of hiring a team of engineers to design and fab a one-off board, anyway).

I appreciate your advice. I use the machine for a variety of different tasks, and am looking to accommodate at least two high-end GPU's (1 for passthrough to virtual machines for running things like Solidworks), a number of SSD's, and as many PCIe expansion cards as possible. Many of the cards are older-gen, so could be consolidated to just a few modern lanes if I could find an external expander with sufficiently generous capacity. Here's a quick inventory of what's in the existing box:

- Mellanox Infiniband. For high-speed, low-latency networking... these days, probably replaceable with integrated NIC's, particularly if they come with RDMA.

- High-performance RAID. I've found dedicated cards offer better features, performance, capacity, resilience and reliability than any of the mobo-integrated garbage I've tried over the years. Things like BBU's/SuperCaps, seamless migration and capacity upgrades, out-of-band monitoring, etc. e.g. I've taken my existing mass storage array created on a modest ARC-1231ML 15+ years ago, through several newer generations to an ARC-1883, with many disk and capacity upgrades along the way, but it's still the same array without ever having had to reformat and restore from scratch. Incidentally I've been particularly happy with Areca's hardware, and they've even implemented some features I requested over the years (like the ability to hot-clone a replacement disk for one expected to fail soon then swap in the new one, without having to degrade the array and wait for a lengthy rebuild process that reduces your fault tolerance while hammering all member disks; as well as some other tweaks for better compatibility with tools like Hard Disk Sentinel). I notice they're finally starting to come out with controllers oriented to SSD's, like a PCI 5.0 product (https://www.areca.com.tw/products/nvme-1689-8N.html) for up to 8 x4 M.2 SSD's that boasts up to 60 GB/s, which is interesting (though the high-queue-depth random performance still doesn't match directly-plugged drives). I know software-RAID for the solid state stuff is also an option (as is just living without redundancy), but it's been convenient outsourcing the complexity.

- Slim, low-performance accessory GPU for more displays

- A few others this crowd would just laugh at me for (e.g. a PCI I/O card that includes a true parallel port, because nothing is more fun™ for hobbyist stuff and USB-based alternatives were found to have too much abstraction or latency; a SCSI adapter for an archaic piece of vintage hardware I'd love to keep installed permanently but there ain't space, and occasional one-off use stuff like a high-bandwidth digitizer).

The motherboard had 6 PCIe slots, and I've got two more provided by an external PCIe expander (after accounting for the one lost for it's own connection). If I could find some kind of expander that took a single PCIe 5.0 slot and turned it into half a dozen PCIe 3.0 slots (some full-width) I'd be set.

I know I'm at the crazy end of how-much-crap-can-you-jam-in-one-PC, but it still seems bizarre to me that newer boards have so many fewer slots yet feel lane-constrained, when between leading-edge SSD's and high-bandwidth GPU's the demand for more lanes is skyrocketing. When I built the previous PC it felt tight but doable... these days it feels like I can barely accommodate the level of graphics and storage I'd like, and by the time I do, there's nothing left for anything else. Granted it's been a few years since I got my hands dirty with this stuff, so maybe I'm just doing it wrong?

And yes, I've heard of USB... and have a bazillion devices plugged in (including some of exotic ones like an LCD display, logic analyzer, and a legit floppy drive that does get used once in a blue moon like when I need to make a memtest86 boot disk for a vintage PC). I've actually found some motherboards have issues where the USB stack gets flakey once you have too many devices connected (even using powered hubs to mitigate power constraints).

Ok... go ahead and have at me; tell me I'm old and dusty and I should take my one GPU and one SSD and be happy with them ;-).

[zamadatix]

I just built a 4th gen Epyc build with 22x NVMe drives and a dual port 40G NIC. It was a FAR superior experience to trying to use prosumer parts, PCIe splitters, etc and didn't end up costing as much as I thought it would (though the DDR5 RAM has went up 2x in the last month since I bought haha).

Side note on RAID: Motherboard integrated garbage is still meh, but hardware options are also pretty meh. I just use software options like ZFS on Linux (or mdraid if I just need a fat RAID 0 with no protection) and get fantastic speeds, portability, no artificial drive topology restrictions, and no additional latency to accessing the drives over PCIe. On Windows the equivalents would be ReFS Storage Pools or whatever the software RAID 0 in the disk manager was called.

If you're looking for cheaper and "not the latest gen" the search tugm4770 on Ebay and filter to the CPU you want. They have MB bundles for standard EATX chassis or bundles with used supermicro servers+PSUs as well. I went the EATX route and put it in a standard PC case, it's crazy how quiet it runs (it's in my living room even)! I used to buy from this seller for the lab at my previous job, but decided to go all in at home this time. Never had trouble with them and it is by far the cheapest way to get these kinds of setups (just avoid the really old Epyc generations).

If the sky is truly the limit and you want the absolute best of the best, get a current gen high frequency optimized Epyc new (don't know the easiest place to do that). You can also do the Threadripper PRO, just depends.

Both of these Epyc options have a "how sky high?" option of getting a dual socket MB, which can give you up to 192 PCIe lanes instead of up to 128. You have to populate both sockets, and the physical footprint starts to become monstrous (especially if you use all the lanes).