[SOLVED] DirectStorage

  • Thread starter Deleted member 2720853
  • Start date
D

Deleted member 2720853

Guest
Can I use DirectStorage if my NVMe is connected to the southbridge and not directly to the CPU?
 
Solution
Hi AlexaKitty.

I think that this will not work on NVMe not connected with the CPU lanes.

That's what I'm getting from the articles I can read about it.
D

Deleted member 2720853

Guest
I guess I'll wait and see. Hopefully it works, since I can very easily install a 2 TB NVMe on my second slot, rather than have to take out both my GPU and CPU cooler to swap drives. I would also like to keep my PCIe 4 NVMe boot drive directly attached to CPU.
 
Well if it's like that it's good but there is no real information about this from Nvidia. This is why some says it will work and others no. Who is right? No idea.
NVIDIA has it's own, but similar technology that seems to skip copying to system RAM.

However I'm still not convinced it requires the NVMe drive to be directly connected to the CPU's PCIe lanes. In PCIe topology there's the so-called root-complex which is the main controller for the entire bus. The chipset is acting like a switch. So the system should be able to directly initiate a transaction from one end point to another regardless of where it's actually connected:
640px-Example_PCI_Express_Topology.svg.png


It'd be like saying you can't connect to the internet unless you're connected to the modem directly.
 
  • Like
Reactions: Maxxify and Nemesia
Using CPU lanes will provide a bit lower latency but otherwise it's similar. Linux of course has P2PDMA which shares similarities with this technology, and notably it depends on chipset support. I see a lot of people posting other places (TPU) "mansplaining" how the PCIe root complex works who apparently don't know about P2PDMA, an exception being the SSD reviewer from AnandTech Billy Tallis, with two follow-up posts on the subject here and here. As hotaru says above, the implication is that you do not need a direct CPU connection with proper support as the PCH is a PCIe switch. In fact, the AMD Zen chipsets like X570 (which support P2P DMA explicitly on Linux) use a PCH that is literally the same as the CPU's I/O die.
 
Last edited:
D

Deleted member 2720853

Guest
Using CPU lanes will provide a bit lower latency but otherwise it's similar. Linux of course has P2PDMA which shares similarities with this technology, and notably it depends on chipset support. I see a lot of people posting other places (TPU) "mansplaining" how the PCIe root complex works who apparently don't know about P2PDMA, an exception being the SSD reviewer from AnandTech Billy Tallis, with two follow-up posts on the subject here and here. As hotaru says above, the implication is that you do not need a direct CPU connection with proper support as the PCH is a PCIe switch. In fact, the AMD Zen chipsets like X570 (which support P2P DMA explicitly on Linux) use a PCH that is literally the same as the CPU's I/O die.
X570? Would that apply to my B550 board too? So that theoretically means I can go ahead and get a 2 TB NVMe on my chipset, and it'll work? Not going to do it until I have 100% confirmation obviously
 
We'll have to see what Microsoft decides to do with it as they'll be having compatibility categories and such no doubt, but there's no actual roadblock to its support in my opinion. Ideally you would be using a primary M.2 socket with CPU lanes (even though all historic consumer Intel boards have NO CPU LANES FOR M.2) for best performance.
 
Last edited:
D

Deleted member 2720853

Guest
We'll have to see what Microsoft decides to do with it as they'll be having compatibility categories and such no doubt, but there's no actual roadblock to its support in my opinion. Ideally you would be using a primary M.2 socket with CPU lanes (even though all historic Intel boards have NO CPU LANES FOR M.2) for best performance.
Impossible unless I find a way to take out my GPU.
 
The B550 has a primary M.2 socket with CPU lanes, so no real issue there, unless you mean a second NVMe drive over the chipset. In that case, again, I think it would be dumb for that not to be supported, but I hesitate to confirm since this is Microsoft. I will ask around on the subject though. For me personally I have a Hyper bifurcating lanes from my GPU so it's not a big deal.