|
edit: poo poo, pagesnipe. Here, dog tax. Not mine, it's fuckin' hot out here, though. edit edit: I'm also a moron, disregard. 
SwissArmyDruid fucked around with this message at 21:17 on Jun 18, 2021 |
#
?
Jun 18, 2021 21:15
|
|
|
|
| # ? May 8, 2024 23:00 |
|
|
long board is loooooong
|
|
#
?
Jun 19, 2021 04:04
|
|
|
love the implication that the board is for bitcoin mining by GPU which hasn't been a thing for like 4 years at this point
|
|
#
?
Jun 19, 2021 04:18
|
|
|
Also DDR5 is out for your future Alder Lake platform (and then Zen 4 sometime like a year from now)! DDR5-4800 CL-40-40-40-77 (JEDEC standard at launch) 16x2 $400 Buy now, be futureproofed!!! (dont do this) https://videocardz.com/press-release/teamgroup-launches-its-elite-u-dimm-ddr5-4800-memory
|
|
#
?
Jun 19, 2021 04:22
|
|
|
Cygni posted:long board is loooooong I love this new Milkcart ATX form factor.
|
|
#
?
Jun 19, 2021 19:13
|
|
|
Intel announced that they're adding HBM to their Sapphire Rapids Xeon processors and I wonder if that has anything to do with trying to keep up with AMD's 3D stacking cache tech
|
|
#
?
Jun 29, 2021 05:18
|
|
|
Most of the things this thread thinks Intel does to keep up with AMD are really about trying to stay competitive with NVIDIA. HBM for cpus is the coming thing in the scientific computing space; see the fujitsu arm64 super. Many science codes are memory bound and putting more generic dimm channels can’t keep up with rising core counts.
|
|
#
?
Jun 29, 2021 05:25
|
|
|
Intel release a CXL based Optane product you cowards
|
|
#
?
Jun 29, 2021 06:16
|
|
|
PCjr sidecar posted:Most of the things this thread thinks Intel does to keep up with AMD are really about trying to stay competitive with NVIDIA. I started typing out, "I'm not entirely convinced that this is even a smart thing for Intel to do, from what I remember with AMD GPUs, HBM is like, 10W per stack, doesn't that begin to rob the cores they're attached to of overhead", but then I figure, Intel has probably already done the math and figured out that they can eliminate the memory bottleneck this way without resulting in a CPU bottleneck instead.
|
|
#
?
Jun 29, 2021 09:58
|
|
|
WhyteRyce posted:Intel release a CXL based Optane product you cowards Wasn't Micron big on putting Optane in CXL before the split? Intel To Disable TSX By Default On More CPUs With New Microcode Intel is dropping TSX from Skylake through Coffeelake, which effectively means there are no major processors on the market that support it. Another win for i7-5775C gang, the one true Intel processor. From looking into this there is a market for a series of Comet Lake engineering samples that still had TSX enabled, they dropped it before QS so those original chips fetch a pretty penny for people really into RPSC3. 10900K ES QTB2 will set you back an 11900k or two. Several interesting differences about the 10th generation processors: ES version has more features There are some weird chips out there. 10600K's with the native six core coffeelake dies and silicon TIM and 10900K's with experimental heatspreaders.
|
|
#
?
Jun 29, 2021 14:30
|
|
|
I am legitimately horny for HBM on die for a client pchem workload, and this has been teased for the last loving five years come ON DO IT
|
|
#
?
Jun 29, 2021 14:58
|
|
|
SwissArmyDruid posted:I started typing out, "I'm not entirely convinced that this is even a smart thing for Intel to do, from what I remember with AMD GPUs, HBM is like, 10W per stack, doesn't that begin to rob the cores they're attached to of overhead", but then I figure, Intel has probably already done the math and figured out that they can eliminate the memory bottleneck this way without resulting in a CPU bottleneck instead. Intel also doesn't seem terribly concerned about just slapping bigger wattage numbers on their processors. So assuming that the cooling they're planning for it can take the extra load, I don't think more power draw is really going to stop them.
|
|
#
?
Jun 29, 2021 15:34
|
|
|
BurritoJustice posted:Wasn't Micron big on putting Optane in CXL before the split? They still say they are in on CXL but have dropped Optane so I don't know what they plan on putting on that
|
|
#
?
Jun 29, 2021 15:47
|
|
|
Sapphire Rapids is their mondo multi tile server CPU. It’s leaked to have a TDP upwards of 350W. So yeah, I don’t think the HBM power use is really a concern (and no, still isnt coming to desktop… probably!)
|
|
#
?
Jun 29, 2021 16:43
|
|
|
DrDork posted:Intel also doesn't seem terribly concerned about just slapping bigger wattage numbers on their processors. So assuming that the cooling they're planning for it can take the extra load, I don't think more power draw is really going to stop them. You can only cool so much in a given space and environment. Since these factors won't be changing much, the power is going to have to come out of the package budget.
|
|
#
?
Jun 29, 2021 18:41
|
|
|
Cygni posted:Sapphire Rapids is their mondo multi tile server CPU. It’s leaked to have a TDP upwards of 350W. So yeah, I don’t think the HBM power use is really a concern (and no, still isnt coming to desktop… probably!) any server is a desktop if you're willing to glue legs onto it
|
|
#
?
Jun 29, 2021 19:26
|
|
|
It doesn't really matter, both Intel and AMD sell server CPUs with 250+W TDPs, they can probably just drop 50MHz off the clocks to carve out 10W for a HBM stack. Anyway, here's a relevant article.
|
|
#
?
Jun 29, 2021 19:26
|
|
|
The reason we don't see more HBM is mainly that it is really loving expensive as balls. The cost is not always worth it, as it does not magically make everything faster. Mostly it benefits bandwidth bound workloads that are not too sensitive to latency and show some data reuse. Don't forget that last bit, because getting the data into the package is often the bottleneck and HBM isn't addressing that problem. So yes, it make sense to slap it on a AI or HPC platform, but probably not for your regular datacenter or CSP platform.
|
|
#
?
Jun 29, 2021 20:21
|
|
|
Also saw that little note that Sapphire Rapids is delayed to Q2 2022 from Q3/Q4 of this year. So if we follow the same schedule as Ice Lake SP, it will come out approximately in 2025.
|
|
#
?
Jun 29, 2021 20:25
|
|
|
Potato Salad posted:any server is a desktop if you're willing to glue legs onto it Makes zoom calls more dramatic with all the airflow.
|
|
#
?
Jun 29, 2021 21:36
|
|
|
priznat posted:Makes zoom calls more dramatic with all the airflow. that's why you build a hot aisle enclosure in your home office
|
|
#
?
Jun 29, 2021 21:49
|
|
|
canyoneer posted:that's why you build a hot aisle enclosure in your home office Any aisle I’m in is a hot aisle because im morbidly obese
|
|
#
?
Jun 29, 2021 22:19
|
|
|
I was in a discussion recently about an enthusiast server that needed a really high single core frequency. This seems like a job for a 10700K, right? Are there any oddball Xeons you can get off AliExpress for cheap that push 5GHz?
|
|
#
?
Jun 30, 2021 02:03
|
|
|
~Coxy posted:I was in a discussion recently about an enthusiast server that needed a really high single core frequency. you can do a 10700K sure, but it won't have ECC, and depending on what you're doing that may be a requirement. if you do need ECC there are also similar client-platform-based Xeon processors - for example Xeon-W 1250P, 1270P, and 1290P in Comet Lake, or Xeon E-2288G/E-2286G/etc in Coffee Lake. https://en.wikipedia.org/wiki/List_of_Intel_Comet_Lake-based_Xeon_microprocessors https://en.wikipedia.org/wiki/List_of_Intel_Coffee_Lake-based_Xeon_microprocessors the gotcha with those processors is going to be the boost behavior because I think all server/workstation platforms are going to respect Tau/etc and after some period of time they will stop boosting. The base frequency is what is officially guaranteed and that is only about 4 GHz tops. it's not gonna be like on a consumer board where it runs all day at max turbo. intel's frequency-optimized server-Xeon SKUs were never mass-market parts as far as I know, and I don't know the current name of them (since they were off-roadmap parts to begin with, the public mostly knows about them 5-10 years after the fact). Generally what is available here is not going to be super competitive against Epyc considering AMD publicly offers frequency-optimized parts, their IPC is higher, and their boost behavior is better. This is gonna be a curveball recommendation but if you want to go Intel and you really only care about a few cores, you don't want SMT, and you want constant high frequencies, recent Intel i3 processors offer validated ECC support and they offer high base clock frequencies (with no turbo boost - which in this case is an advantage since it means it will never clock down) at relatively moderate prices. You can do a Supermicro whitebox server board with an 8350K or 9350K (or -KF, etc) and that will get you 4 GHz baseclock on a validated platform (same guarantees as a Xeon), with 4C4T. You don't have to overclock it but the K versions clock higher at stock (and on i3s they still offer ECC). Really though if this is for a serious job the best thing is going to be a frequency-optimized Epyc. A 72F3 is 8 cores at up to 4.1 GHz boost, and AMD's IPC is higher so that's faster than a 4 GHz Intel, and AMD's boost behavior is better than Intel's, even under load you should maintain higher than base clocks (which is 3.7 GHz). This goes all the way up to 75F3 which is 32 cores running at between 3 GHz and 4 GHz. Yes, it will be several grand, but that's actually cheap compared to what Intel's off-market frequency-optimized SKUs would cost you. https://en.wikichip.org/wiki/amd/cores/milan#Milan_Processors if you are just doing personal stuff you could also look at some of the whitebox AM4 server boards, and these will boost even higher, but I've heard some bad stuff about the quality of the validation involved (eg someone in the NAS thread owned one and said they could get it to boot with ECC UDIMM memory but it didn't seem to actually correct errors). One more weird option is getting an Asus X99 WS, X99E WS, or X99 WS/IPMI board and a Xeon E5-1660v3. $125, 8 cores, and you can overclock it to run 4.7 GHz all-core if you want, or a bit slower if you want to be really safe. It supports ECC RDIMMs and so on as well, if you want to stack in a bunch of memory, and you get up to 7 PCIe slots as well. Probably not the most thoroughly validated platform either, but it gets you a "server" style platform with tons of memory and a lot of PCIe at a much lower price, and because you can overclock it you can guarantee that it will always boost. If it's just an enthusiast toy that's a decent option for the price. Paul MaudDib fucked around with this message at 09:10 on Jun 30, 2021 |
|
#
?
Jun 30, 2021 06:53
|
|
|
What funky-rear end kind of workload needs a Xeon and is single-threaded, nowadays?
|
|
#
?
Jun 30, 2021 08:24
|
|
|
Paul MaudDib posted:snip Thanks very much for the thoughtful post. The application is a games server, of course. It's currently a homebrew 3820K in a rack, so the overclocked Xeon might be the way to go.
|
|
#
?
Jun 30, 2021 09:30
|
|
|
~Coxy posted:Thanks very much for the thoughtful post. probably don't need ECC for games for a games server yeah the 10700K, or an AMD "server"-ish board like an X470D4U (this is the aforementioned board with questionable ECC support though) with a 5600X/5900X or similar is probably the way to go. X99/1660v3 is fine too but the other options will outperform it if you don't need the RAM. I'm guessing "games" is probably minecraft or source engine or similar single-thread limited stuff and those are the best options to just punch the single-threaded at the minimum cost. Paul MaudDib fucked around with this message at 09:39 on Jun 30, 2021 |
|
#
?
Jun 30, 2021 09:36
|
|
|
You don't "need" ECC for anything other than system stability. If you value having applications or the OS not crash, or being able to know if a DIMM is having issues, you want ECC.
|
|
#
?
Jun 30, 2021 10:03
|
|
|
It's a proxmox host with ceph, so it is probably worth getting a ECC capable solution. Although I just checked the current host and it actually doesn't seem to have ECC to begin with so...
|
|
#
?
Jun 30, 2021 11:14
|
|
|
~Coxy posted:It's a proxmox host with ceph, so it is probably worth getting a ECC capable solution. In a previous post you mentioned overclocking it. If you're gonna do that, then IMO there's no reason to do ECC - you're deliberately reducing system reliability and data integrity, and not in a way which is significantly mitigated by ECC.
|
|
#
?
Jun 30, 2021 11:48
|
|
|
BlankSystemDaemon posted:What funky-rear end kind of workload needs a Xeon and is single-threaded, nowadays? HFT It's a market important enough to get its own SKU.
|
|
#
?
Jul 1, 2021 10:24
|
|
|
Beef posted:HFT It's one of those things where, if you have to ask the price, you can't possibly afford it.
|
|
#
?
Jul 1, 2021 12:13
|
|
|
BlankSystemDaemon posted:The BlackCore, which is the mythical system that Intel sells to high-frequency traders, isn't single-threaded; it's got 10 cores at 5.4GHz for all cores. I'd kill to see their memory OC, the 5.4 all core is heard of but 5.0 cache at the same time is amazing. Given how latency bound everything is their memory must be insanely tuned. The amount of binning and testing to have those overclocks be 24/7/365 stable is crazy. E: I looked into the spreadsheet and they're using a Z490 Aorus Master, so a consumer 4 slot board so the ram numbers can't be too insane. Props to gigabyte for their penultimate consumer board making it into this high end BurritoJustice fucked around with this message at 13:34 on Jul 1, 2021 |
|
#
?
Jul 1, 2021 13:28
|
|
|
BlankSystemDaemon posted:The BlackCore, which is the mythical system that Intel sells to high-frequency traders, isn't single-threaded; it's got 10 cores at 5.4GHz for all cores. Five figures or six? I *want* to afford it.
|
|
#
?
Jul 1, 2021 16:11
|
|
|
BlankSystemDaemon posted:The BlackCore, which is the mythical system that Intel sells to high-frequency traders, isn't single-threaded; it's got 10 cores at 5.4GHz for all cores. What are they doing there for cooling? At 5.4 GHz that seems like borderline phase change cooling territory but I guess maybe it’s just barely feasible with a hyper-binned chip since even SiliconLottery has a few chips in the 5.3 range (and that’s all-core while they may be advertising single core numbers here…). Plus SiliconLottery is actually a bit conservative on their voltages, and these guys may be willing to use more voltage than most people would consider safe - they probably don’t care if it dies in a year. Note that’s still using a 10900K which is still a normal part, probably just a platinum tier bin. previous HFT offerings used off-roadmap chips that typically were server chips with all cache enabled and a limited number of cores (to maximize cache per core) running at super high frequencies (otherwise unheard-of for server chips). Of course today the server chips are on Skylake-X which takes an IPC hit compared to consumer Skylake due to the mesh… and Ice Lake-SP won’t clock all that great.
|
|
#
?
Jul 1, 2021 17:25
|
|
|
The BlackCore and ICC ones are vendor made I think. Stuff like the auction-only 9990XE is direct from Intel for HFT: https://www.anandtech.com/show/14980/the-intel-core-i9-9990xe-review
|
|
#
?
Jul 1, 2021 17:30
|
|
|
Cygni posted:Also saw that little note that Sapphire Rapids is delayed to Q2 2022 from Q3/Q4 of this year. So if we follow the same schedule as Ice Lake SP, it will come out approximately in 2025. Even Charlie at Semiaccurate seems to think that the Q3 2022 is likely going to be the real availability and he would be shouting from the rooftops if availability was going to be any further away than this.
|
|
#
?
Jul 6, 2021 20:00
|
|
|
Bofast posted:Even Charlie at Semiaccurate seems to think that the Q3 2022 is likely going to be the real availability and he would be shouting from the rooftops if availability was going to be any further away than this. i was goofin at intel's expense, a joke if you will 
|
|
#
?
Jul 6, 2021 21:01
|
|
|
Cygni posted:i was goofin at intel's expense, a joke if you will My bad, friend
|
|
#
?
Jul 8, 2021 08:33
|
|
|
|
| # ? May 8, 2024 23:00 |
|
|
 So it looks like the mobile mask is 6C+8c with a big iGPU, but the desktop mask is gonna be 8C+8c with a small iGPU? Also no HT mentioned? Also are we gonna go with this big-C little-c notation?
|
|
#
?
Jul 9, 2021 09:07
|
|