priznat posted:They might have some junk cluttering up the motherboard where a x16 edge would bang into it I guess. Workstation/server boards are how I know about open-ended dautherboard slots.
|
|
# ? Apr 21, 2024 23:36 |
|
|
# ? May 4, 2024 08:35 |
|
Asrock tends to leave the x1 slots open-ended on about half of their mid-high end consumer boards. I've only ever seen ASUS do it for their workstation-marketed and enterprise things.
|
# ? Apr 22, 2024 05:46 |
|
In terms of PCIe slots the best board I found was the Asus proart b650-creator - notably better than the x670 proart if you need a third slot with more than two lanes. The b650 will do x8/x8/x4 PCIe 4.0, which might limit upgrades in the future but the current gen nvidia cards only do 4.0 and my NIC is pcie 3.0 so I needed the lanes. In practice I'm only able to do about 20gbps with iperf on a 40Gbe NIC, instead of the 32ish I was hoping for, but that might be limited by the CPU at the other end and a lack of proper offload somewhere. The x670 proart will do x8/x8 PCIe 5.0 but then the third slot is stuck at two lanes PCIe 4.0. edit: Oh yeah there are MSI boards that will also do x8/x8/x4 but I had excluded them due to historically bad IOMMU support from MSI. If you're not concerned about that, the MSI boards will probably be even better. Desuwa fucked around with this message at 16:08 on Apr 22, 2024 |
# ? Apr 22, 2024 07:11 |
|
The MSI MPG X670E Carbon Wifi could be another option, it will do x8/x8/x4 with the first two x8s at PCIe 5.0. I've been using this board populated with a 3080 in the first slot, P420 SAS controller in the second, and a X520 2x 10Gb NIC in the third without any issues so far.
|
# ? Apr 22, 2024 09:51 |
Bjork Bjowlob posted:The MSI MPG X670E Carbon Wifi could be another option, it will do x8/x8/x4 with the first two x8s at PCIe 5.0. I've been using this board populated with a 3080 in the first slot, P420 SAS controller in the second, and a X520 2x 10Gb NIC in the third without any issues so far.
|
|
# ? Apr 22, 2024 09:58 |
|
You'd also need to consider the power delivery as wider slots have more power to them, you'd need to make sure that these cute small open ended slots can handle up to 75 watts cards.
|
# ? Apr 22, 2024 12:14 |
|
Kivi posted:You'd also need to consider the power delivery as wider slots have more power to them, you'd need to make sure that these cute small open ended slots can handle up to 75 watts cards. PCIe power is all on that front stubby bit, which is the same on every size of slot. Every slot needs 75 watts by spec. My guess would be that open-end slots are much easier to break or damage. Server stuff gets put together and then shoved into racks and nobody touches it until it fails or is obsolete. DIYers are always monkeying with their PCs. And if someone puts a heavy x16 GPU into an open-end x4 slot and then is moving or shipping the PC, it probably ends in tears. Thus the general absence in consumer boards and prevalence in server and pro-grade stuff.
|
# ? Apr 22, 2024 14:01 |
|
Klyith posted:PCIe power is all on that front stubby bit, which is the same on every size of slot. Klyith posted:Every slot needs 75 watts by spec.
|
# ? Apr 22, 2024 15:23 |
|
25W ought to be enough. Even the most inefficient 400GBit ConnectX-7 barely goes past this with 26W, with other models less power hungry.
|
# ? Apr 22, 2024 16:20 |
|
Mixing different types of memory of similar specs is still a no-go? I want to add more RAM to my NAS, turns out they meanwhile switched from Micron E-die to Hynix C-die on the (mostly) specific model of module.
|
# ? Apr 22, 2024 17:09 |
|
Combat Pretzel posted:Mixing different types of memory of similar specs is still a no-go? Do you mean like one stick of each, or adding another pair of sticks? Mixed sticks in a single pair, your problem is the XMP/AMP values might be slightly different for the two. When loading XMP it just looks at one stick, it doesn't do any comparison or smarts. It should work fine in JDEC, or with speed backed down one notch from rated value. But to fully OC you may need manual settings of timings & voltage set because you have to target the worst value for both. Adding a second pair, it doesn't matter because even 4 perfectly matched sticks will get trained to slower timings, if they can run at rated speed at all.
|
# ? Apr 22, 2024 17:32 |
|
Going from two to four sticks, at JEDEC speeds. They're DDR4-3200 CL22 out of the box. Well, four of them are gonna run at most at 2666 from what the mainboard manual says.
|
# ? Apr 22, 2024 17:48 |
|
Combat Pretzel posted:Mixing different types of memory of similar specs is still a no-go? It’s a NAS, run JEDEC speeds. Combat Pretzel posted:Going from two to four sticks, at JEDEC speeds. They're DDR4-3200 CL22 out of the box. Well, four of them are gonna run at most at 2666 from what the mainboard manual says. …well, if the mobo’s original release predates zen2 I bet you could get 3200 cl22 which is the fastest JEDEC speed.
|
# ? Apr 22, 2024 21:28 |
|
Tuna-Fish posted:True.
|
# ? Apr 24, 2024 19:39 |
|
crazypenguin posted:e: and it looks like Apple's A17 Pro is 35 TOPS, and A18 will probably come out at the same time, so maybe qualcomm isn't that far ahead of everyone here For what it's worth, Apple traditionally used 16-bit TOPS as their marketing number (*), and their NPUs always double that number when doing 8-bit computations. Some think that for whatever reason, they chose to market A17 Pro using 8-bit TOPs, while sticking with 16-bit numbers for M3. The reasoning is simple: in the past Apple's reused the same NPU block in both A-series and M-series chips, and M3/A17 Pro are both N3, launched at about the same time, and share lots of other cores (same CPUs, for example). There should be no reason why the A17 Pro is about 2x as fast. Frustratingly, in the months since launch, nobody seems to have benchmarked this to confirm or deny the hypothesis. Or if they have, I can't find it. By the way, yes, this is a huge problem for NPU TOPs comparisons in general. Be sure you're comparing the same thing... edited to add this footnote: * - iirc they seldom or never explicitly said they were using 16-bit, people had to test M1 to determine that the marketing number was 16-bit TOPs BobHoward fucked around with this message at 03:24 on Apr 26, 2024 |
# ? Apr 26, 2024 03:21 |
|
|
# ? May 4, 2024 08:35 |
|
Oh interesting. I definitely didn't take the time to think carefully about whether these sources were using different units. I was confused about the difference between M3 and A17, but shrugged it off. I wonder if we'll see standardization here or if bitwidths and formats will continue to change.
|
# ? May 3, 2024 16:27 |