Nice! One suggestion - please add AM4 socket boards. With current memory prices, AM5 with DDR5 is becoming unattainable for some. DDR4 prices are rising as well. But not nearly as bad as DDR5.
Very cool. Seeing how almost everything from WiFi, to NVME SSDs, (to apparently USB ports sometimes?) are connected to it, is PCIe the only high-speed interconnect we have for peripherals to communicate with modern CPUs?
The high speed signals that come out of mainstream CPU chips are generally DDR, SMP, and PCIe. Outside of a very few exotic things that use QPI or HT to connect, or exotic storage might use DDR, yes high speed off-chip peripherals use PCIe.
NVLink is another one you might have heard of, although it might also fall in the exotic category. I think some systems take AXI off-chip too. So there's various other weird and wonderful things. But none you're likely to have in your PC I think.
On-chip is another story, you can connect USB or NVMe or GPU "peripherals" using an on-chip interconnect type. But I guess you are asking about off-chip.
Probably a good thing SLI fell out of fashion. No consumer boards with multiple 16x, but a few with 2 8x (gated behind a "mode" switch). A few years ago it was looking like we were on our way to full 4 16x slots. For cuda/llm/whatever does it really matter if the cards are in 1x slots?
... shouldn't the logic be opposite? "Bad that SLI went out of fashion, there's no way for two GPUs to communicate fast over pcie, and SLI would allow such fast bridge"
Whether or not SLI remained viable for gaming, Broadcom was going to jack up the prices on PCIe switches to the enterprise-only range. That's the real reason why consumer motherboards don't have more GPU slots. Mainstream consumer CPU sockets never had a wealth of PCIe lanes, there was just a brief span of years where PCIe switches were cheap so high-end consumer boards could offer several x8 or x16 slots (sharing bandwidth in ways that make diagrams like these important).
In previous decades, non-mainstream CPU sockets were also more accessible to consumer budgets; first-gen Threadripper started at only 8 cores, so it was possible to pay extra for more memory channels and IO lanes without also buying an excess of CPU cores. But that had little to do with the popularity or viability of multi-GPU consumer systems.
Can anyone recommend a specific, well-made, high-performance motherboard with loads of PCIe lanes and expansion slots, and sensible lane topology?
All the motherboards these days make me feel claustrophobic. My current workstation is pretty old, but feels like it had more expansion capability (relative to its time) than what's on the market today.
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.
>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
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.
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 ;-).
Given that you've said 'workstation', if you've got a spare $5000, a Threadripper Pro comes with 128 PCIe 5.0 lanes.
This means you can get a motherboard like the "Asus Pro WS WRX90E-SAGE SE" which dedicates 104 lanes to seven PCIe slots and 16 lanes to four M.2 slots.
For more like $3000 you can get a non-Pro Threadripper; the "Asus Pro WS TRX50-SAGE" has a more restrained 48 PCIe 5.0 and 32 PCIe 4.0 lanes, meaning the board's five PCIe slots and three M.2 slots have a mixture of speeds and lanes.
The rest of the market seems to think you just want to plug in one huge four-slot GPU and perhaps one other card.
(ps. I don't suppose they make a "supersized" version of that board with a gap beside the first one or two GPU slots? So you can install a couple double-width cards without losing the underlying slots? Or a good source for a single-width, high-end GPU like the Inno3D RTX 5090 iChill Frostbite Pro?)
Yeah my ASRock have nice map of the every lane and interface and where they are connected on the board. Especially important as some devices go thru second io expander
Wow, this is great! I don't know how they generate this but it's really impressive. One of the things that I've been surprised with is some older dual socket workstations have tons of PCI-E lanes, but none are hooked to the second CPU it seems
Very nice! Just a note (as the site says on bottom left side), this can vary depending on the CPU you use, would be nice to be able to select all different variations of supported CPUs as a future feature.
That is so incredibly useful, hardware vendors do such a bad job of properly advertising how many GPUs will actually work and with what combination of m.2 slots in use.
Nice! One suggestion - please add AM4 socket boards. With current memory prices, AM5 with DDR5 is becoming unattainable for some. DDR4 prices are rising as well. But not nearly as bad as DDR5.
Very cool. Seeing how almost everything from WiFi, to NVME SSDs, (to apparently USB ports sometimes?) are connected to it, is PCIe the only high-speed interconnect we have for peripherals to communicate with modern CPUs?
The high speed signals that come out of mainstream CPU chips are generally DDR, SMP, and PCIe. Outside of a very few exotic things that use QPI or HT to connect, or exotic storage might use DDR, yes high speed off-chip peripherals use PCIe.
NVLink is another one you might have heard of, although it might also fall in the exotic category. I think some systems take AXI off-chip too. So there's various other weird and wonderful things. But none you're likely to have in your PC I think.
On-chip is another story, you can connect USB or NVMe or GPU "peripherals" using an on-chip interconnect type. But I guess you are asking about off-chip.
> PCIe the only high-speed interconnect we have for peripherals to communicate with modern CPUs?
In a pedantic/technical sense, no. Practically speaking though, yes.
I’ve been struggling to find an AM5 board that can run three MI50s at 4x. This is perfect thank you.
Him are you sure about some of the PCI slots? I think some marked as 4x get downgraded to 1x on these boards…
Further edit - this maybe accurate - how are you getting this / confirming it?
Probably a good thing SLI fell out of fashion. No consumer boards with multiple 16x, but a few with 2 8x (gated behind a "mode" switch). A few years ago it was looking like we were on our way to full 4 16x slots. For cuda/llm/whatever does it really matter if the cards are in 1x slots?
GPUs in 16x slots is still important for LLM stuff, especially multi-GPU, where lots of data needs to move between cards during computation.
... shouldn't the logic be opposite? "Bad that SLI went out of fashion, there's no way for two GPUs to communicate fast over pcie, and SLI would allow such fast bridge"
Whether or not SLI remained viable for gaming, Broadcom was going to jack up the prices on PCIe switches to the enterprise-only range. That's the real reason why consumer motherboards don't have more GPU slots. Mainstream consumer CPU sockets never had a wealth of PCIe lanes, there was just a brief span of years where PCIe switches were cheap so high-end consumer boards could offer several x8 or x16 slots (sharing bandwidth in ways that make diagrams like these important).
In previous decades, non-mainstream CPU sockets were also more accessible to consumer budgets; first-gen Threadripper started at only 8 cores, so it was possible to pay extra for more memory channels and IO lanes without also buying an excess of CPU cores. But that had little to do with the popularity or viability of multi-GPU consumer systems.
Can anyone recommend a specific, well-made, high-performance motherboard with loads of PCIe lanes and expansion slots, and sensible lane topology?
All the motherboards these days make me feel claustrophobic. My current workstation is pretty old, but feels like it had more expansion capability (relative to its time) than what's on the market today.
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.
>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
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.
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 ;-).
Given that you've said 'workstation', if you've got a spare $5000, a Threadripper Pro comes with 128 PCIe 5.0 lanes.
This means you can get a motherboard like the "Asus Pro WS WRX90E-SAGE SE" which dedicates 104 lanes to seven PCIe slots and 16 lanes to four M.2 slots.
For more like $3000 you can get a non-Pro Threadripper; the "Asus Pro WS TRX50-SAGE" has a more restrained 48 PCIe 5.0 and 32 PCIe 4.0 lanes, meaning the board's five PCIe slots and three M.2 slots have a mixture of speeds and lanes.
The rest of the market seems to think you just want to plug in one huge four-slot GPU and perhaps one other card.
Now we're talking - thanks!
(ps. I don't suppose they make a "supersized" version of that board with a gap beside the first one or two GPU slots? So you can install a couple double-width cards without losing the underlying slots? Or a good source for a single-width, high-end GPU like the Inno3D RTX 5090 iChill Frostbite Pro?)
Some builds I kept tabs on:
Let's Encrypt documented their early 2021 whitebox that used 128 PCIe 4.0 lanes, mainly for storage: https://letsencrypt.org/2021/01/21/next-gen-database-servers...
Troy Hunt (HaveIBeenPwned) recently solicited upgrade advice from the internet and settled on an Asus Pro WS TRX50-SAGE WIFI (which doesn't appear to be in the MoboMaps database yet): https://gist.github.com/troyhunt/a6e565981e4769976e9cffb705f...
How can I contribute the data for the boards I own which are not on the site?
I wish all manufacturers clearly gave info like this up front. AM4 boards would be nice.
Yeah my ASRock have nice map of the every lane and interface and where they are connected on the board. Especially important as some devices go thru second io expander
For disclosure, this was created by "Ronin Wilde" - https://www.youtube.com/watch?v=cgdXj75VSMo
I found it useful and thought others might also like it.
Whoa. This is so cool and helpful. Too bad my board is Intel. Is there a way to contribute to this?
I dropped a message to the creator :fingers_crossed: they open the motherboard database so we can make contributions
Wow, this is great! I don't know how they generate this but it's really impressive. One of the things that I've been surprised with is some older dual socket workstations have tons of PCI-E lanes, but none are hooked to the second CPU it seems
Very nice! Just a note (as the site says on bottom left side), this can vary depending on the CPU you use, would be nice to be able to select all different variations of supported CPUs as a future feature.
That is so incredibly useful, hardware vendors do such a bad job of properly advertising how many GPUs will actually work and with what combination of m.2 slots in use.
Warning: addicting site :)
Legendary!
[flagged]