|
Eletriarnation posted:Yeah, most mATX cases are like "standard big-OEM desktop" size. Noticeably smaller than a full tower, but not really "small form factor". It is a nice sweet spot if you're confident that you won't need all of the ATX expansion slots but you still want a relatively normal build experience, especially since mATX boards tend to be a little cheaper than full ATX boards with the same feature set and they're definitely cheaper than ITX. Even the average mid tower case feels pointlessly bloated for a purely gaming system when the NR200 exists tbh
|
# ? Apr 13, 2024 10:57 |
|
|
# ? May 4, 2024 14:55 |
|
Arzachel posted:Even the average mid tower case feels pointlessly bloated for a purely gaming system when the NR200 exists tbh I built a machine for work in a NR200 (air-cooled) but having room to work on poo poo easily inside the case is nice. My use case mounts the tower under my standing desk, and I know my Meshify fits so the desire for 'smaller' is just some other part of my midlife crisis, probably. And I need room for all my sweet sweet radiators.
|
# ? Apr 13, 2024 11:24 |
|
movax posted:Oops… 3D Cache derail! The points made all make sense to me though. I think what would drive me insane, knowing myself, is verifying games are doing the right thing on the right cores and just taking away the problem in the first place by having one CCD will win over that part of my brain. Well, uh, you've chosen two games that get basically zero benefit from all those cores. Battletech is single-thread (and janky in general). Civ 6 is basically 2 threads. And you already should be driving yourself insane over CCDs with the threadripper. A general problem for AMD's architecture is that if a thread gets moved from a core on CCD A to one on CCD B, you have performance loss from needing to fetch cache across the IO die. This problem got mostly solved in the scheduler over the years, but I dunno if the threadrippers are still affected by it. I still see people saying they use process lasso for their threadrippers and it makes a difference. movax posted:My use case mounts the tower under my standing desk, and I know my Meshify fits so the desire for 'smaller' is just some other part of my midlife crisis, probably. And I need room for all my sweet sweet radiators. Yeah if you're gonna stick with the watercooling I wouldn't downsize to mATX. The Meshify 2C you have is already pretty small for an ATX case.
|
# ? Apr 13, 2024 14:42 |
|
The Ryzen 7000 threadrippers are atrocious for gaming because they use the giant 12 channel IO die from Genoa, which when fully enabled has four NUMA domains with three memory channels and three CCDs each. The configuration used for TR however is four total CCDs and four total memory channels, one per NUMA node. So each CCD only has a single local memory channel. One of the CCDs is the gaming targeted one and has a Fmax that is 500MHz higher than all the others, and if you turn on gaming mode in the UEFI it will disable all the other CCDs to maximise gaming performance, but you're still crippled for gaming by the awful memory topology.
|
# ? Apr 13, 2024 15:01 |
|
I would have figured that they've had seen the error in their ways with the earlier Threadrippers and go with an UMA memory controller this time around.
|
# ? Apr 13, 2024 15:29 |
|
Combat Pretzel posted:I would have figured that they've had seen the error in their ways with the earlier Threadrippers and go with an UMA memory controller this time around.
|
# ? Apr 13, 2024 15:35 |
|
Combat Pretzel posted:I would have figured that they've had seen the error in their ways with the earlier Threadrippers and go with an UMA memory controller this time around. Threadripper non-Pro is just an afterthought they are limping along so they can technically claim they still have HEDT chips, they'll never spin up a new die for it so the best we are getting is the full size IO die cut down enough for it to not cut into their high margin lines. You have to be in a weird spot of wanting lots of cores but not particularly caring about memory or IO to want to buy TR non-pro now, it's a platform that isn't really a good buy for any real usecases. There's a reason MSI literally didn't bother to make a TRX50 motherboard while the other big vendors only made one SKU each, it's going to sell close to zero chips and they know it.
|
# ? Apr 13, 2024 15:37 |
|
BurritoJustice posted:You have to be in a weird spot of wanting lots of cores but not particularly caring about memory or IO to want to buy TR non-pro now, it's a platform that isn't really a good buy for any real usecases. I actually have a use-case at work for this sort of setup right now (run many tiny web assembly functions, as parallel as we can get, over small data) but I doubt I can get GCP to deploy them for us!
|
# ? Apr 13, 2024 16:20 |
|
Subjunctive posted:I actually have a use-case at work for this sort of setup right now (run many tiny web assembly functions, as parallel as we can get, over small data) but I doubt I can get GCP to deploy them for us! Isn't that usage case of "tons of cores without significant I/O or memory bandwidth" about to be completely devoured by both ARM64, whether Ampere or in-house, or E-core Intel? I haven't looked at the GCP pricing comparison, but I do feel like we're heading for a future where for those kinds of workloads it's not cost effective to run them on mainstream Intel or AMD. I guess we'll see how Zen 5c SKUs and pricing shake out, or if the big cloud guys buy from Ampere or just roll their own ARM.
|
# ? Apr 13, 2024 16:39 |
|
BurritoJustice posted:
Partly because a lot of anticheat systems will not let you set affinity.
|
# ? Apr 13, 2024 16:49 |
|
Twerk from Home posted:Isn't that usage case of "tons of cores without significant I/O or memory bandwidth" about to be completely devoured by both ARM64, whether Ampere or in-house, or E-core Intel? I haven't looked at the GCP pricing comparison, but I do feel like we're heading for a future where for those kinds of workloads it's not cost effective to run them on mainstream Intel or AMD. Yeah, we’ll probably move that service to ARM, realistically.
|
# ? Apr 13, 2024 16:53 |
|
Twerk from Home posted:Isn't that usage case of "tons of cores without significant I/O or memory bandwidth" The memory bandwidth is fine, it's latency that gets killed by having to hop across the IO die to a different controller. That's real bad for games, but plenty of Real Work type things are much less bothered.
|
# ? Apr 13, 2024 18:10 |
|
beware of a lot of the cheap new B650 AM5 boards: https://www.youtube.com/watch?v=naX-DnKekCM Best cheap board is probably the Asrock B650M-HDV/M.2 at $119 or the Asrock B650M Pro RS at $134. e: 5 of the boards can't even run a single cold pass of Cinebench without throttling jesus christ Cygni fucked around with this message at 19:58 on Apr 13, 2024 |
# ? Apr 13, 2024 19:52 |
|
So, I have an "ASUS B650-A GAMING WIFI" motherboard because apparently allcaps makes it better? Anyway, I'm running a slightly older BIOS on it (v1811) and I see they are up to 2413, a couple "enhance system stability", a couple AGESA versions, and the logofail fix. My 7800X3D system is running well, stable, stock clocks/voltage, memory is running at 6000/CL30 off its EXPO profile just fine, is there any reason I should update to the newer BIOS? Or if it isn't broke, don't fix it?
|
# ? Apr 13, 2024 20:25 |
|
Klyith posted:The memory bandwidth is fine, it's latency that gets killed by having to hop across the IO die to a different controller. That's real bad for games, but plenty of Real Work type things are much less bothered. Bandwidth is also not incredible, you've got four times the cores of a 7950X and only twice the memory bandwidth.
|
# ? Apr 14, 2024 00:35 |
|
BurritoJustice posted:Bandwidth is also not incredible, you've got four times the cores of a 7950X and only twice the memory bandwidth. 88 PCIE lanes is the real reason to ever look at it
|
# ? Apr 14, 2024 00:57 |
|
Somebody hacked P2P communication between RTX 4090s for deep learning, time to load up all the x16 slots for a "cheap" AI machine.
|
# ? Apr 14, 2024 01:33 |
|
no idea where else to put this, but the e2-6110 in this QC6000M is a world of pain all in itself. only 27% less clocks than my old athlon 5350s (that are ostensibly a generation older than this) but still manages to be more than 27% slower. I don't even remember the athlon 5350 choking this hard on doing ANYTHING on windows 10. it runs nuclear throne at a brisk 15fps, just over half of the framerate it gets on the athlon 5350. no benchmark will run. this is prime i'm floating installing a stripped down windows 7 to give it a modicum of a fighting chance, but i'm enjoying the experience enough as-is.
|
# ? Apr 14, 2024 02:35 |
|
Arzachel posted:Even the average mid tower case feels pointlessly bloated for a purely gaming system when the NR200 exists tbh i tried and later gave up on "large enough to fit a full size GPU" ITX long ago they dont really save much space, and that's still pointless at the end of the day when they still need masses of cables plugged from a fixed position anyway Palladium fucked around with this message at 05:20 on Apr 14, 2024 |
# ? Apr 14, 2024 03:55 |
|
Klyith posted:Well, uh, you've chosen two games that get basically zero benefit from all those cores. Battletech is single-thread (and janky in general). Civ 6 is basically 2 threads. I am bad at computer now, clearly... My reasoning at the time was "I don't play the latest AAA games / am usually 3-4 years behind, so picking a CPU that at least is parity with the most recent console generation will let me repeat my 2600K experience" ... and I didn't deep-dive enough to pick up the following: BurritoJustice posted:The Ryzen 7000 threadrippers are atrocious for gaming because they use the giant 12 channel IO die from Genoa, which when fully enabled has four NUMA domains with three memory channels and three CCDs each. The configuration used for TR however is four total CCDs and four total memory channels, one per NUMA node. So each CCD only has a single local memory channel. I guess it's a sign of getting older / maturity to actually pick based on some requirements vs. 'haha biggest number' but now I'll just plan for a nice Q4 build on Zen 5... hopefully X3D parts don't arrive in Q1 2025 like this past cycle. hobbesmaster posted:88 PCIE lanes is the real reason to ever look at it I used the PCIe lanes as an excuse to shove in as much used NVMe as I could. AFAIK though there is not a 65 W TDP part in any EPYC line up with the big-rear end I/O die for maximum lanes -- I want to do an all-flash / NVMe build at some point and have some EPYC parts on the shelf, but changed my mind on running a high-TDP part. Beginning to think a PCIe switch is the solution and just use an under-volted Ryzen 5 or something.
|
# ? Apr 14, 2024 18:43 |
|
I want an 9950X3D with V-Cache on both dies.
|
# ? Apr 14, 2024 18:57 |
Combat Pretzel posted:I want an 9950X3D with V-Cache on both dies.
|
|
# ? Apr 14, 2024 19:09 |
|
Combat Pretzel posted:I want an 9950X3D with V-Cache on both dies. Yes, the ERA brick meme but with fat SRAM stacks on everything.
|
# ? Apr 14, 2024 20:15 |
|
movax posted:I'll just plan for a nice Q4 build on Zen 5... hopefully X3D parts don't arrive in Q1 2025 like this past cycle.
|
# ? Apr 15, 2024 18:34 |
|
amd revealed some info on their 8000 prosumer line and their inclusion of npu'squote:AMD announced its Ryzen Pro portfolio today as it extended the 'Hawk Point' 8040-series to commercial laptop and workstation users while simultaneously offering its Ryzen 8000 'Phoenix' APU models for commercial desktop PCs. As we've seen in the past, the Pro series is based on AMD's existing consumer-oriented processor models but comes with additional features that tailor them for the commercial market. With the consumer versions of these processors, AMD was the first x86 company to bring its AI-processing neural processing unit (NPU) to the mobile and desktop PC market. Those same AI acceleration features are now headed to commercial users, allowing AMD to lay claim to being the first with professional CPUs armed with NPUs for laptops and workstations. https://www.tomshardware.com/pc-com...mmercial-market ... they still don't meet microsoft's tops requirement, though: quote:AMD's in-built XNDA engine powers the NPU, and the company touts that its mobile processors have an advantage over Intel's competing Core Ultra processors with 16 TOPS of NPU performance, outperforming the Meteor Lake NPU's 11 TOPS. AMD also maintains a slimmer lead in overall system TOPS, which includes both the CPU and GPU AI processing power, taking the lead with 39 TOPS over Intel's 34 TOPS. AMD's desktop Ryzen 8000 APUs also have an in-built NPU engine that delivers 16 TOPS, whereas Intel has yet to release a processor for desktop PCs with an integrated AI engine. no idea how microsoft arrived at 45 tops as the sweet spot, maybe someone has some qualified guesses
|
# ? Apr 16, 2024 16:27 |
|
quote:AMD's NPU 16 TOPS I just wanted to edit that quote down to parse the performance numbers a little easier. I don't know if all this NPU stuff if going to matter, but at least back when the iphone XS came out, it's bump in NPU (to 5 TOPS) was really quite nice because it cut down on the latency for face id unlock. We'll see if literally anything as useful as face id comes of any of this, I guess. So far looking grim. Also interesting to see Qualcomm promising so much more than everyone else. Wonder if it's BS somehow. kliras posted:no idea how microsoft arrived at 45 tops as the sweet spot, maybe someone has some qualified guesses I think performance estimation is actually pretty easy with these things, so they probably just had a particular model in mind and a latency target. 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 crazypenguin fucked around with this message at 17:36 on Apr 16, 2024 |
# ? Apr 16, 2024 17:32 |
|
Looking forward to having to paying more for the UNPU cpus that don't waste silicon on stuff I'm not interested in yet. "Overclocking? Funny you ask, we sell these ~*unlocked*~ cpus just for ~*enthusiasts*~ like you!" please imagine a choir of angels wherever you see ~* *~
|
# ? Apr 16, 2024 19:04 |
|
On the topic of PCIE lanes, is there any AM5 motherboard that offers slots for both x16 and x8? I've been looking but the best I can seem to find is an x16 slot and an x4. I didn't realize this would be a problem when I got a 7800x3d recently, but the motherboard I purchased only does x16 and then 3 PCIE 3.0 x1 slots. I know my use case is stupid, but I have an x8 Mellanox 25gbps card that wouldn't work at x1 for obvious reasons.
|
# ? Apr 21, 2024 15:41 |
|
There are a relative handful of X670/E mobos that have a PCIe4 x4 lane (but full length physical x16) slot but that is it as far as I know. This time around everyone seems to have crammed m.2 slots everywhere they could instead of full PCIe slots, even at PCIe 1.0 or 2.0, like they used to.
|
# ? Apr 21, 2024 15:56 |
|
Thanks! I'll take another look at the X670 boards. I get that most people will want more m2 slots these days. A PCIe4 x4 slot should work bandwidth wise, I'll just have to decide if my stupid use case is worth the gamble and spending more money on.
|
# ? Apr 21, 2024 16:10 |
|
JSON Bourne posted:On the topic of PCIE lanes, is there any AM5 motherboard that offers slots for both x16 and x8? I've been looking but the best I can seem to find is an x16 slot and an x4. Physical or electrical? There are 28 PCIE 5.0 lanes on AM5 CPUs. 4 lanes go to the chipset leaving 24. AMD’s slides call the other 24 lanes “16 graphics, 8 NVMe”. It’s not clear if that is a suggestion or requirement but as you saw motherboards all seem to be doing that. The 4 PCIE lanes are downgraded to 4.0 lanes for the chipset. The chipsets can be thought of as simple PCIE switches. B650 motherboards have 1, and X670 have two in series. 4x PCIE 4.0 in principle is equivalent to 8x 3.0 so what you want might actually be possible but with the death of SLI it’s questionable if even asrock would bother… edit: i apologize for ever doubting asrock, their expensive x670e boards have you covered https://pg.asrock.com/mb/AMD/X670E%20PG%20Lightning/index.asp edit2 no they don’t, goddamnit why can’t you search by motherboard block diagram or something Edit3: x670e asrock taiichi https://www.asrock.com/mb/AMD/X670E%20Taichi/index.asp#Specification hobbesmaster fucked around with this message at 16:21 on Apr 21, 2024 |
# ? Apr 21, 2024 16:10 |
|
JSON Bourne posted:On the topic of PCIE lanes, is there any AM5 motherboard that offers slots for both x16 and x8? I've been looking but the best I can seem to find is an x16 slot and an x4. A true x8 lane slot? Not really. The only thing you're going to get is the high-end X670E mobos with 2 x16 PCIe 5.0 slots, which will split the x16 from the CPU into a pair of x8s. All of the mobos with "x16" slots that are really only x4 or x1 are that way because they hang off the chipset and the chipset link is x4 gen4. Will the card not work with reduced lanes though? PCIe is supposed to auto-negotiate lanes and stuff. Even at x4 gen3 you're getting 4 GByte/s, which should cover the 25gbit/s of your card. Unless the card is gen2 because it's old, at which point you'd just have to live with reduced speed or get a more modern card.
|
# ? Apr 21, 2024 16:13 |
|
Yea, seems like best you'll be able to find is a PCIe 3.0 4x hanging off a chipset. Any additional CPU lanes are reserved for the five hundred M.2 slots each higher end mainboard comes with. Also, I'm running a Mellanox ConnectX-3 in such a slot just fine. --edit: X670E TUF from Asus here. Combat Pretzel fucked around with this message at 16:23 on Apr 21, 2024 |
# ? Apr 21, 2024 16:17 |
|
At least AM5 is more straight forward than some AM4 boards. I’m investigating throwing a plex or something server together with parts i have lying around and the spare motherboard is a b450 tomahawk…
|
# ? Apr 21, 2024 16:25 |
|
Wow, thank you all! Appreciate the extra insight here. hobbesmaster posted:Physical or electrical? You're experiencing what I have been going through so many different spec sheets this week. It's been hard to keep it all straight. Thank you for the extra technical detail here. Klyith posted:A true x8 lane slot? Not really. The only thing you're going to get is the high-end X670E mobos with 2 x16 PCIe 5.0 slots, which will split the x16 from the CPU into a pair of x8s. I should have been more specific, a true slot isn't really a requirement. What clued me into the PCIe lanes potentially being the problem was a log line indicating "NIC supported width: 8, NIC current width 1". It's a not that old of a card, so it is good to know that negotiation should work. Really good to understand about the chipset being a gen4 x4 link, I think I was missing that distinction before. Combat Pretzel posted:Yea, seems like best you'll be able to find is a PCIe 3.0 4x hanging off a chipset. Any additional CPU lanes are reserved for the five hundred M.2 slots each higher end mainboard comes with. Knowing I'm not the only one trying this makes me feel less crazy. I have a ConnectX-6DX so I would hope it would work like yours did.
|
# ? Apr 21, 2024 16:48 |
|
hobbesmaster posted:edit: i apologize for ever doubting asrock, their expensive x670e boards have you covered https://pg.asrock.com/mb/AMD/X670E%20PG%20Lightning/index.asp Huh? According to their manual, PG lightning routes the PCIE3 connector (mechanically PCIe x16, electrically PCIe 4.0x4) directly from the CPU. JSON Bourne posted:On the topic of PCIE lanes, is there any AM5 motherboard that offers slots for both x16 and x8? I've been looking but the best I can seem to find is an x16 slot and an x4. 3.0 x4 is just barely enough for a single link of 25Gbps, with overhead. I recommend going for one of the motherboards where the PCIe slot hangs directly off the CPU, otherwise the network card has to share bandwidth with everything else on your system.
|
# ? Apr 21, 2024 17:10 |
|
Tuna-Fish posted:Huh? They’re looking for PCIE 3.0 8x. Tuna-Fish posted:Huh? It’s x4, not x8. The mellanox card needs 8 electrical lanes.
|
# ? Apr 21, 2024 21:39 |
|
hobbesmaster posted:They’re looking for PCIE 3.0 8x. It doesn't need 8, it gracefully degrades down to 4. If you have a single 25Gbps link, 3.0x4 is sufficient. (32Gbps bandwidth) Just need physical x16 and electrical x4.
|
# ? Apr 21, 2024 23:19 |
I don't understand why PCIe slots that're sub-x16 aren't open-ended - it's not some fancy new technology, it's just a tiny bit of less plastic that you change the mold a bit to be able to form. Theoretically, you could even mechanically drill it out as a modification - though I'm not sure I'd recommend it. I have the ASrock X670E PG Lightning, and it's a pretty good card for anyone with lots of daughterboards.
|
|
# ? Apr 21, 2024 23:28 |
|
|
# ? May 4, 2024 14:55 |
|
BlankSystemDaemon posted:I don't understand why PCIe slots that're sub-x16 aren't open-ended - it's not some fancy new technology, it's just a tiny bit of less plastic that you change the mold a bit to be able to form. They might have some junk cluttering up the motherboard where a x16 edge would bang into it I guess. I’ve done it for work stuff on slots that didn’t have parts in the way and just snipping out the plastic worked best with some fine cutters. A lot of enterprise boards are more likely to have the open ended slots too.
|
# ? Apr 21, 2024 23:31 |