So the specs for the new intel mainstream platform are out, and we’re finally getting more lanes than any other previous (and existing) mainstream platform.
Sees slide
Oh
Oh…
Damn it intel, can’t we have more pci-e lanes off the cpu? I just want to be able to run 2 graphics cards at full bandwidth, why does the chipset need 24 lanes? I just want to do vfio things in peace…
I can’t really see how many lanes there will be. The numbers are vague and meaningless.
Adding usable + PCH + DMI lanes together is stupid and only useful if you want to present a large number. Oh right, that’s a marketing slide.
Without a proper block diagram, it’s pretty much guess work. But my guess is 20x 5.0 lanes (x16 slot + 1x M.2) and some PCIe 4 x4 slot or M.2 and rest is PCH.
Pretty much just another tuesday with little changes but new and fancy slides. Is more lanes on the chipsetreally better if the DMI stays at PCIe 4.0 x8 and not increase as well? Do we get an x8 PCH slot with this “new league of data lanes”?
More questions than answers. I remain skeptic
check the numbers on USB. You can run a datacenter just by using 800 series and USB.
Well I got the information I know off this video (https://youtu.be/zhIXt1svQZg?feature=shared&t=1009), I just didn’t want to link it from the get-go since other creators are in these threads too. The video mentions 20 lanes coming from the cpu itself and another 24 from the chipset, which is why I’m disappointed.
ASRock’s got their manuals up. Arrow Lake’s 5x20 + 4x4 + DMI 4x8 + 2 USB 4/Thunderbolt 4 (3x8) and then there’s usually 4x16 user assignable off the Z890 (one 4x17).
| mobo | CPU | Z890 PCIe 4.0 | Z890 PCIe 3.0 |
|---|---|---|---|
| Z890 Taichi | 5x16, 5x4 M.2, 4x4 M.2 | 3xM.2, M.2/x4 | 2x gen 1, 5 GbE, 2.5 GbE |
| Z890 Taichi Lite | 5x16, 5x4 M.2, 4x4 M.2 | 3xM.2, M.2/x4 | 2x gen 1, 5 GbE, 2.5 GbE |
| Z890 Steel Legend | 5x16, 5x4 M.2, 4x4 M.2 | 2xM.2, 2x4(16) | 2x gen 1, WiFi, 2.5 GbE |
| Z890 Pro RS | 5x16, 5x4 M.2, 4x4 M.2 | 2xM.2, 2x4, x1, WiFi | 2.5 GbE |
| Z890 Nova | 5x16, 5x4 M.2, 4x4 M.2 | 3xM.2, M.2/x4(16) | 2x gen 1, WiFi, 2.5 GbE |
| Z890 Live Mixer | 5x16, 5x4 M.2, 4x4 M.2 | 2xM.2, 2x4(16), 2x gen 1, WiFi, 2.5 GbE | |
| Z890 Riptide | 5x16, 5x4 M.2, 4x4 M.2 | 3xM.2, M.2/x4(16), WiFi, 2.5 GbE | 2x gen 1 |
| Z890 Lightning | 5x16, 5x4 M.2, 4x4 M.2 | 2xM.2, 2x4(16) | 2x gen 1, WiFi, 2.5 GbE |
Also, the couple MSIs I look at by default.
| mobo | CPU | Z890 PCIe 4.0 | Z890 PCIe 3.0 |
|---|---|---|---|
| Z890 Carbon | 5x16 or 5x8/x4/M.2, 5x4 M.2 | 2xM.2, x4(16) | (no block diagram) |
| Z890 Tomahawk | 5x16, 5x4 M.2, 4x4 M.2 | 2xM2, 2x4(16) | (no block diagram) |
All of these boards have four SATA ports.
If the motherboard implements vfio correctly, chipset pcie can do vfio, too. The thing is board manufacturers won’t validate vfio for the desktop market. We get it working just by chance.
For example, some of m.2 under the chipset on my x670e carbon board get individual iommu group, but the chipset pcie slot doesn’t have its dedicated group, mixed with usb stuffs.
That doesn’t solve my issue though. I want 2 graphics cards running at full x16 bandwidth. Intel could do that with their new core ultra platform but they chose to give more bandwidth to the chipset than the cpu.
Get a Xeon board if you want Intel. Ain’t gonna happen below workstation level
I feel you on PCI-E lanes but no one is going to give them to us anytime soon. Just get use to it or pay more for HEDT if available as there is no hope in sight for this.
Sigh yeah, seems like the only solution is to put used car money into threadripper or sapphire rapids refresh, and god forbid if playing games is the main thing you want to do with your machine.
Probably dual x8 gen 5 is good enough.
For AM5, you can also get 2 x4 Gen 5 for m.2, thus total 24 Gen 5 lanes are about 96GB/s bandwidth. For your reference, dual channel ddr5 6000 can only gets 96GB/s in theory. You will run out memory bandwidth very quickly.
There is some performance loss between x16 and x8 though, it’d get in my head too quickly and I’d be rather annoyed.
I think this is the deeper reason for things being the way they are. There’s not much point having PCIe bandwith if the memory can’t keep up. The only niche use cases where memory wouldn’t be limiting would be GPUs talking to each other or some infiniband NIC. There’s directstorage too, but that would be limited by the x4 ssd link any way. On the server platforms as well we see memory channels scale with pcie lanes.
That said, at least Intel provides x8 chipset lanes so the motherboard vendors can work with that. I just wish chipsets would get PCIe 5 but I guess it would make the chipsets way too expensive.
Depends on the application. What do you actually want to do?
PCIe 5 GPUs coming soon. I highly doubt 5x8 would significantly bottleneck anything you would want to do barring a compute workload spread over 2 GPUs. Even then it’s probably more latency bound than bandwith…
Otherwise, with Turin out and perhaps soon Threadripper 9000, look out for a sale on zen 4 HEDT/server?
Gaming at high fps (planning on coupling my next main pc with a 1440p 144/165 Hz monitor) and video editing. I plan on having 2 graphics cards in the system, one for the linux host and one for a windows virtual machine (vfio things, wendell has covered such a use case quite a lot). I’ve thought about used server gear but can the cpus from said platforms keep up with high fps gaming or would mainstream be my only option (sapphire rapids refresh looks pretty decent on paper but the platform’s still expensive).
Do you need two discrete GPUs or would the iGPU be fine on the host? I use an iGPU on the host and it’s fine. If the VM is not running I can still use the dGPU for compute or gaming with prime offload.
Unless you need a lot of cores and memory HEDT is not worth the price tag for the minimal difference of x8 vs. x16.
Having a discrete gpu on the host would be a much more comfortable experience for me. I’m getting nitpicky here but if I want to do an activity in a game that doesn’t require my full attention but still be able to fire up davinci resolve, or even another game, and leave the other one running in the background. I’ve thought about the possibility of doing a pseudo-optimus setup like on my notebook (basically what you’re describing, having the on-board gpu take care of the host and if the discrete one is available, it can run whatever program needs a beefier gpu), but I’m still not sure. One thing I agree though, is hedt not worth the price. I’m only considering it because of the amount of pci-e lanes, mainstream platforms nowadays have plenty of cores for what I’d like to do. Hell, my main rig’s a 6-core pc.
This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.
