|
With non-ECC DDR5, the only proof you have of the on-die ECC working is by running memory benchmarks and start looking for evidence of the memory slowing down.
|
#
?
Oct 13, 2023 20:10
|
|
|
|
| # ? May 30, 2024 12:00 |
|
|
Kingston sells DDR5 ECC UDIMMs under their Server Premier brand. The 4800MT ones are Hynix M-dies and clock fine to 6000MT with the Buildzoid timings (works fine here), the newer 5200MT ones are Hynix A-die and support even tighter timings ostensibly.
|
|
#
?
Oct 13, 2023 23:58
|
|
|
Once you start overclocking ECC RAM, you risk losing the accuracy of the parity bits and could start getting silent corruption.
|
|
#
?
Oct 14, 2023 00:08
|
|
|
Combat Pretzel posted:Kingston sells DDR5 ECC UDIMMs under their Server Premier brand. The 4800MT ones are Hynix M-dies and clock fine to 6000MT with the Buildzoid timings (works fine here), the newer 5200MT ones are Hynix A-die and support even tighter timings ostensibly.
|
|
#
?
Oct 14, 2023 00:09
|
|
|
Kibner posted:Once you start overclocking ECC RAM, you risk losing the accuracy of the parity bits and could start getting silent corruption. You don't increase the odds of silent to noisy correction when you do that. Which means that so long as you don't get any correctables, you are not getting silent ones either.
|
|
#
?
Oct 14, 2023 00:18
|
|
|
Would ECC make it easier or harder to do 4x32@6000?
|
|
#
?
Oct 14, 2023 01:18
|
|
|
Subjunctive posted:Would ECC make it easier or harder to do 4x32@6000? any 4x@6000 on AM5 is basically impossible right now, better off with 2x64.
|
|
#
?
Oct 14, 2023 02:33
|
|
|
wargames posted:any 4x@6000 on AM5 is basically impossible right now, better off with 2x64. Isn’t that the same as 4x32?
|
|
#
?
Oct 14, 2023 02:48
|
|
|
hobbesmaster posted:Isn’t that the same as 4x32? I assume they meant only use two dimm slots. Using all four is what is causing a lot of the issues with reduced ram speed.
|
|
#
?
Oct 14, 2023 02:50
|
|
|
Kibner posted:I assume they meant only use two dimm slots. Using all four is what is causing a lot of the issues with reduced ram speed. All 4 ranks is usually the issue, not slots per se. is 64GB not always dual rank?
|
|
#
?
Oct 14, 2023 02:52
|
|
|
hobbesmaster posted:All 4 ranks is usually the issue, not slots per se. is 64GB not always dual rank? No, fully occupied slots are a problem for signal quality in the normal daisy chain topology. 2x dual rank are easier to run at high speed than 4x single rank. edit: that was the case with DDR4... looks like DDR5 is vice-versa and the dual rank sticks generally have lower peak speed? But this is on the ram side, not the memory controller having trouble running it. And you can get high speed dual rank sticks if you spend 
Klyith fucked around with this message at 03:32 on Oct 14, 2023 |
|
#
?
Oct 14, 2023 03:00
|
|
|
I think most 32GB sticks of DDR5 are dual-rank too? The way DDR5 is organized is different from DDR4 though so I don't know how much the distinction matters. Either way, 2x64GB is the way to go if you want to overclock your memory.
|
|
#
?
Oct 14, 2023 03:07
|
|
|
hobbesmaster posted:Isn’t that the same as 4x32? you can run 4x64. or 4x24 but any of the 4 dimm configs is broken, stick with 2 dimms of the highest capacity you can get and you can get to 6000+
|
|
#
?
Oct 14, 2023 04:36
|
|
|
Multi-channel width vs. speed has always been a tradeoff, and AM5 is no exception. Would I like my 7700X box to be able to run 4x16s at 6000 MHz? Sure. Is 2x16 enough? Yeah, for now. When it's not, I'm sure AM5 will be mature enough that I can pop an 8800X3D and 4x16 at 6000 MHz in and it'll probably Just Work. But in the meantime, I can run Streets of Tarkov at high graphics at 1440p at 90-110 FPS and that's more than good enough for me.
|
|
#
?
Oct 14, 2023 09:02
|
|
|
Kazinsal posted:Multi-channel width vs. speed has always been a tradeoff, and AM5 is no exception. You can do 64GB at 6000 right now, you just need to go 2x32GB. You can even do 6000 at up to 96GB if you use 2x48GB. I'm running 2x32GB 6400c28, using inexpensive greenstick Hynix memory I bought from a server supplier.
|
|
#
?
Oct 14, 2023 10:54
|
|
|
ITX boards stay winning
|
|
#
?
Oct 14, 2023 11:17
|
|
|
I’m running 4x16 at 6000 right now on AM5, though for one or two BIOS versions it was touch and go. Just using the EXPO profile, Asrock riptide B650e.
|
|
#
?
Oct 14, 2023 12:44
|
|
|
Hmm, maybe I’ll try it again.
|
|
#
?
Oct 14, 2023 23:39
|
|
|
Gyrotica posted:I’m running 4x16 at 6000 right now on AM5, though for one or two BIOS versions it was touch and go. Just using the EXPO profile, Asrock riptide B650e. my Gigabyte board won't post if I do expo with 4x16, so I am at stock ram speeds.
|
|
#
?
Oct 15, 2023 08:28
|
|
|
wargames posted:my Gigabyte board won't post if I do expo with 4x16, so I am at stock ram speeds. oh geez, you can probably do better than that unless your ram is really bad. Load expo, and then before rebooting go into wherever gigabyte keeps the main memory clockspeed setting and lower it a notch or two. That gives you timings appropriate for high speed operation, but at a slightly lower clock which means everything has a bit more wiggle room. It's the easiest way to get recalcitrant ram to work -- this was what I told people to do with first-gen ryzen back when that was real picky about what ram would run XMP.
|
|
#
?
Oct 15, 2023 13:19
|
|
|
I didn't see this mentioned anywhere and a CPU thread is as good a place as any to do it: Canon announced the development of commercially relevant nanoimprint lithography machines. They're currently capable of printing 5nm-equivalent features and they mention expectations of 2nm. It doesn't look like that will happen before EUV gets there, but the machines are cheaper to both buy and run. Considering the spiralling costs of EUV nodes and the fairly anemic improvements they bring (Intel 20A is supposedly only going to have a 5% single thread improvement over 4nm because of lower clock speeds), giving people more cheaper silicon may be the easier path to higher performance CPUs/GPUs. e: there's also been some recent talk about free electron lasers as a source of EUV. Asianometry put out a video on it recently too. Interesting times on the lithography front. ConanTheLibrarian fucked around with this message at 09:21 on Oct 17, 2023 |
|
#
?
Oct 17, 2023 09:04
|
|
|
ConanTheLibrarian posted:I didn't see this mentioned anywhere and a CPU thread is as good a place as any to do it: Canon announced the development of commercially relevant nanoimprint lithography machines. They're currently capable of printing 5nm-equivalent features and they mention expectations of 2nm. It doesn't look like that will happen before EUV gets there, but the machines are cheaper to both buy and run. Considering the spiralling costs of EUV nodes and the fairly anemic improvements they bring (Intel 20A is supposedly only going to have a 5% single thread improvement over 4nm because of lower clock speeds), giving people more cheaper silicon may be the easier path to higher performance CPUs/GPUs.
kinda nuts that that can work on a nanoscopic level
|
|
#
?
Oct 17, 2023 10:04
|
|
|
speaking of trying to do a synthetic 3d cache test, it appears the 3d cache benefit in factorio only lasts as long as you can fit your map in it (duh), which apparently gets harder over time, with a lot of benchmarkers using 10K maps which some don't think reflects a lot of use cases. i haven't played factorio myself, but this is what some nerds have to say:reddit posted:As someone with a 7950X3D, I really hate the Factorio hype for X3D CPUs because it is based around a fundamental misunderstanding on how to benchmark Factorio. https://old.reddit.com/r/Amd/comments/178bh52/ryzen_7_7800x3d_wipes_the_floor_with_core/ so rather than limit 3d cache on your end, you could increase the size of the map that normally gets cached i do wonder what optimization goes into fitting all this stuff in the 3d cache, especially given how it varies
|
|
#
?
Oct 17, 2023 14:00
|
|
|
Rinkles posted:
Happy to be proven wrong though.
|
|
#
?
Oct 17, 2023 17:17
|
|
|
kliras posted:speaking of trying to do a synthetic 3d cache test, it appears the 3d cache benefit in factorio only lasts as long as you can fit your map in it (duh), which apparently gets harder over time, with a lot of benchmarkers using 10K maps which some don't think reflects a lot of use cases. i haven't played factorio myself, but this is what some nerds have to say: It doesn't even seem accurate to say that the 3D cache benefits AMD on small map sizes because the game is going to be locked to 60 no matter what for most users. Hardware Unboxed added another test in their 14th-gen review for this reason.  The 3D cache still saves AMD from being too far behind here, but it's not quite enough to actually give them the lead anymore. It's an interesting look into why the extra cache can be so beneficial, though.
|
|
#
?
Oct 18, 2023 06:05
|
|
|
Yeah, I think pretty much everyone has been guilty of underestimating the impact of the cache Intel has been adding the past few years. At least for Factorio, they've done a lot to shrink the window where an AMD CPU is markedly superior.
|
|
#
?
Oct 18, 2023 14:44
|
|
|
Watched the GN review of the 14700k last night and it's pretty telling that the 5800X3D is still up there in a lot of the gaming benchmarks.
|
|
#
?
Oct 18, 2023 14:52
|
|
|
Non-pro threadrippers are reportedly returning, with an official announcement coming within a day: https://videocardz.com/newz/amd-to-launch-threadripper-7000-pro-and-non-pro-zen4-cpus-designed-for-wrx90-and-trx50-motherboards
|
|
#
?
Oct 19, 2023 00:21
|
|
|
3D vcache and L3 cache in general is gonna double every generation right? Cache wars
|
|
#
?
Oct 19, 2023 02:57
|
|
|
Shaocaholica posted:3D vcache and L3 cache in general is gonna double every generation right? Cache wars i think the issue is the physical space it takes up
|
|
#
?
Oct 19, 2023 03:13
|
|
|
Cache rules everything around me
|
|
#
?
Oct 19, 2023 03:23
|
|
|
Wibla posted:Watched the GN review of the 14700k last night and it's pretty telling that the 5800X3D is still up there in a lot of the gaming benchmarks. I've seen new i7 review where going from 6000 to 8000MT RAM gave games a couple extra FPS. I know they compare like for like, but I doubt anyone pairs those "new" CPUs with slow RAM when you want to game. I'm also pretty smug at cheaply upgrading my 3700x to 5800x3d when AM5 debuted..
|
|
#
?
Oct 19, 2023 07:17
|
|
|
alex314 posted:I'm also pretty smug at cheaply upgrading my 3700x to 5800x3d when AM5 debuted..  Although for me it was from a 3600X and a while before AM5 launched. I'm honestly a little surprised they made a fair supply of 5800X3Ds - part of why I got mine early was it finally hit a good sale and I was concerned supply would dry up with the release of the next generation.
|
|
#
?
Oct 19, 2023 07:29
|
|
|
CaptainSarcastic posted:
It's good business for them, they could sell a ton of cpus to people with old am4 setups
|
|
#
?
Oct 19, 2023 07:58
|
|
|
alex314 posted:I'm also pretty smug at cheaply upgrading my 3700x to 5800x3d when AM5 debuted.. I paid full retail (because  ), but still so worth it.Only issue I have now is that vmware player triggers a BSOD. Haven't bothered looking into it too much though.
|
|
#
?
Oct 19, 2023 08:24
|
|
|
Wibla posted:
I remember playing with BIOS to enable or disable something before vmware started working. It wasn't causing BSOD though, just "you can't run VMs fool".
|
|
#
?
Oct 19, 2023 08:31
|
|
|
Wibla posted:
Maybe virtualisation extensions got reset to disabled during a bios update. Worth a look
|
|
#
?
Oct 19, 2023 08:50
|
|
|
threadripper announcement finally official https://twitter.com/Techmeme/status/1714993728237293698 official announcement: https://ir.amd.com/news-events/press-releases/detail/1162/amd-introduces-new-amd-ryzen-threadripper-7000-series    https://www.youtube.com/watch?v=oUqWE9HJ83I kliras fucked around with this message at 14:27 on Oct 19, 2023 |
|
#
?
Oct 19, 2023 14:18
|
|
|
The fact Threadripper vanilla is a cutdown Threadripper PRO which is a variant of Epyc should bode well for it. It must cut down on the R&D costs so it's more viable long term, while benefitting from all the quality control work that the Epyc market demands.
|
|
#
?
Oct 19, 2023 14:38
|
|
|
|
| # ? May 30, 2024 12:00 |
|
|
Is ‘480mb’ threadripper pro cache usable by all cores?
|
|
#
?
Oct 19, 2023 15:08
|
|