|
I wouldn't buy a zen CPU until the bug is fixed but it doesn't make me worried either. Seems like they are zeroing in on the bug and will be able to patch it. This isn't ideal but isn't shocking for a new architecture and doesn't mean much long-term, unless it doesn't get patched.
|
#
¿
Jun 9, 2017 01:57
|
|
|
|
| # ¿ Apr 24, 2024 03:34 |
|
|
Combat Pretzel posted:You expect the Threadripper dies to be revisions? I think he just wants AMD to figure out their segfault issue and work around it. Presumably fixing the issue for Ryzen should allow them to fix it for TR too.
|
|
#
¿
Jun 17, 2017 22:47
|
|
|
GCC fuckery hasn't been sorted out. Individuals report mixed results from various workarounds. Turning on load line calibration with a low setting has fixed the issue for a few people. On one of the community threads an AMD employee implied the issue wasn't widespread and they were doing service for affected individuals. Not sure I'm comfortable with that but I can believe it doesn't affect all chips. IOMMU groups do seem resolved, though that also comes down to the specifics of your motherboard. Do your research. The biggest one stopping me is NPT performance with GPU passthrough, which apparently has been a problem for AMD as far back as bulldozer. AMD hasn't properly acknowledged this issue, from what I've seen. This would be a critical failure on threadripper, which otherwise seems perfect for a gaming VM build. Desuwa fucked around with this message at 05:35 on Jun 23, 2017 |
|
#
¿
Jun 23, 2017 02:46
|
|
|
Scarecow posted:Isnt the rumoured b2 step ment to fix these issues? AMD hasn't acknowledged the NPT bug. We also don't know what the B2 stepping is supposed to fix, if anything. Searching for NPT and ryzen in the VFIO subreddit will turn up some discussion over people attempting to manage it. The short of it is that leaving NPT enabled will ruin GPU performance but disabling it will ruin CPU performance. https://community.amd.com/thread/215931 https://www.redhat.com/archives/vfio-users/2017-April/msg00019.html https://lists.linuxfoundation.org/pipermail/iommu/2017-May/021592.html e: Ah, here's the much older bug I saw: https://sourceforge.net/p/kvm/bugs/230/ Apparently it's been broken forever on AMD and nobody cared because nobody used AMD processors. There's a reasonable chance of this getting attention, with Ryzen being usable, but if there isn't at least a fix in the pipeline by the time I do my build I'm going to have to go with SKLX even if it burns my apartment down. Desuwa fucked around with this message at 05:57 on Jun 23, 2017 |
|
#
¿
Jun 23, 2017 05:47
|
|
|
Combat Pretzel posted:GPU performance is currently irrelevant for cloud providers. I didn't mention cloud providers, but at the same time I wouldn't say it's irrelevant. However, GPU compute is a completely different beast from trying to play games on a consumer GPU with passthrough and it's unknown if the bug even gets triggered with how cloud providers set up their machines. I'm hoping this will get attention just because enthusiasts are actually buying and using AMD processors, and Threadripper especially is a perfect fit for the gaming VM use case, but I'm not holding my breath for a silver bullet. If AMD does look at it and it can't be fixed until Zen+, or later, then I'll just have to go with SKLX once Volta is released. Hopefully Intel will have announced a soldered revision before that.
|
|
#
¿
Jun 23, 2017 09:25
|
|
|
eames posted:Hmm, that makes Ryzen very tempting if passing through the entire SATA controller really solves all performance issues. His idea of using a physical drive to dual boot as physical and virtual machine is pretty neat too. That is extremely promising. Also in that thread was information that Xen is working on enabling geforce cards by working around nvidia's checks like KVM does, though I'd rather just hold out for a KVM patch.
|
|
#
¿
Jun 26, 2017 00:33
|
|
|
I can see them avoiding unbalanced configurations because that might give unpredictable performance. With 12 and 16 they've got two dies with two CCXs with 3/4 cores each, which keeps everything equal. They could go down to 8 or 4 but there's little point, the only advantage over Ryzen would be support for more memory, and those customers would probably go for lower end Epycs instead for even more memory.
|
|
#
¿
Jul 14, 2017 07:25
|
|
|
If you get enough RGB LEDs you don't even need a monitor.
|
|
#
¿
Jul 26, 2017 07:27
|
|
|
Klyith posted:The main problem with both that one and the Define S is that dust filters behind a solid, not-easily-removable front panel suck. On my main PC I want to clean the dust filters every 2-3 months, good luck with that when you have to pry off the front panel. The Define S front panel is easy to remove though? Just pull it off. The connections aren't that strong, more than enough that the front panel is securely attached but not enough to be difficult to remove. The downsides to the define S are the lack of dust filters if you open the top vents and the bottom dust filter being inaccessible from the front or side.
|
|
#
¿
Feb 27, 2018 18:33
|
|
|
I'm basically in the same boat, doing my first custom loop except with two gpus (not SLI, the other is my 1080ti) for a Windows gaming VM. Not sure where you are but the 2950x is still available on Amazon for shipping some time next week. I stayed up a bit late to ensure I could order it as soon as it was available. Wish I could have just pre-ordered it.  E: misread, looks like you did manage to order one and you were just commenting on it being sold out. Desuwa fucked around with this message at 05:41 on Sep 3, 2018 |
|
#
¿
Aug 31, 2018 20:45
|
|
|
Combat Pretzel posted:I find that temperature offset stuff on the Threadripper (and Ryzen?) pretty irritating. I've found most windows software already handles it, the only one that hasn't yet is msi kombuster's on screen display. Same for my motherboard which is how I'm controlling the fans. Linux software doesn't yet but they do tend to lag behind.
|
|
#
¿
Sep 11, 2018 19:59
|
|
|
I was playing with PBO and I think I might just leave my 2950x at stock or with only a very modest boost, like 20W just to feel like I gained something. The boosts don't seem to be worth pushing it since I'm not constantly loading all cores, doubly so because if I load my UPS too much its fan will turn on and that thing is loud.Combat Pretzel posted:The whole temp stuff seems weird. Package temperatures on my TR seem to be more correct, without the 27°C offset. But when doing random low effort crap, like watching a video and using a browser, it randomly tends to jump 8-10°C and then comes back down. But when playing a game, which loads the CPU, it's steady and goes slowly up (when the water cooling is trying to get to steady state). The temperature reported by the package isn't from a single sensor but instead takes the maximum of a bunch of different sensors that aren't individually exposed. So it's not your entire chip getting 10C hotter, it's just one small part, which is entirely normal with modern CPU boost algorithms. What's happening is there's tons of power and thermal headroom at idle so when a task does come in a single core will run flat out until it's done with as much power as a single core can handle. You hit another steady state, where the energy leaving the cores equalizes with the energy entering the water, very quickly, and below that temperatures can swing wildly.
|
|
#
¿
Sep 12, 2018 05:08
|
|
|
https://rog.asus.com/forum/showthread.php?105105-Code-0E-Load-VGA-Bios Not really impressed for what I paid for this board, especially given that it's not a new board and ASUS makes Turing cards themselves. One of the reasons I picked it was because it had already had a year out to work out all the kinks. I'm doing GPU passthrough so I need a second physical GPU. Good thing I thought to test this poo poo before draining my loop. This should have been caught, and they should have had a day one Turing compatibility BIOS. I'm seriously considering going through the effort of ripping this out and replacing it with a taichi, which would certainly ship faster than ASUS will ship a new BIOS.
|
|
#
¿
Oct 6, 2018 10:10
|
|
|
I'm not so sure about how good Window's NUMA scheduler is, at least for Threadripper style NUMA where the latencies between the nodes are very low. Turning on NUMA in my BIOS tanked CPU performance on Windows (using geekbench anyway), but I never did reinstall and I know Windows can be weird about CPUs "changing" out from underneath it. They only mention Windows in that article so this might just be a lower level bandaid on a poor NUMA balancing implementation in Windows.
|
|
#
¿
Oct 7, 2018 21:24
|
|
|
Klyith posted:It's absolutely a bandaid, but I wouldn't call it a low-level one at all. It's like manually setting CPU affinity in task manager, but automatic by an app. That works well enough for people using threadrippers for gaming or video editing or other single-user stuff, but would totally not fly in more complicated environments. Huh, I assumed it was lower level than that.
|
|
#
¿
Oct 9, 2018 10:10
|
|
|
Malcolm XML posted:They use lesser dies right? The io die looks to be what would be crippled or harvested. TR has always sort of been a yield improvement program The opposite, actually, TR is harvested from the top 5% of Zen dies. This leads to funny things like the 2950x having a higher max turbo than the 2700x, despite having twice as many cores. I guess they don't bin the 2700x as much as they did the 1800x, even when there's no 2800x.
|
|
#
¿
Dec 2, 2018 04:06
|
|
|
Asus seems to have dropped the ball on that feature. My zenith extreme doesn't have it, which kinda screwed me over because I wanted to pass NUMA node 1 through to my VM but I'm forced to put the GPU on node 0. There were pages in my BIOS that sounded like they were supposed to let me select the GPU but they did nothing.
|
|
#
¿
Dec 17, 2018 21:09
|
|
|
Yeah I'm done with Asus garbage after this disaster. It came down to the Taichi or the zenith and ultimately I made the wrong choice. It took them over a month after the launch of the 2080ti to get the board working properly with those cards - using them in any kind of multi-GPU setup caused the board to fail to POST. From what I understand this was also a problem on Asus' other x399 boards. I actually ordered a Taichi to replace it but the fix came out with enough time to cancel the order. If not for needing to pull apart my water loop I'd still have replaced it. I do have headers for temperature probes, at least.
|
|
#
¿
Dec 17, 2018 22:20
|
|
|
Harik posted:Looking to build a box for virtualization: host+guest GPU, NVMe, 10GbE. Probably means I'm looking at a TR2 2950x. I've got this setup running right now and after the initial hiccups with AMD breaking Linux with the same Agesa update that enabled support for TR2 it's been mostly smooth sailing. I'm assuming you're doing this for gaming in a Windows VM with a Linux host, in which case there are plenty of good resources on level 1techs or if you google for vfio. It's really not too difficult now, and the most important thing is buying the right motherboard with good IOMMU groups. From personal experience I recommend going with the ASRock Taichi board, I went with the zenith extreme and it hasn't been stellar. It's recommended that you stay away from MSI boards as they're not known for good IOMMU groups. For 10GbE the cheapest option is picking up some used mellanox connectx-2s on ebay, though it gets more expensive if you want more than just a direct connection between two machines.
|
|
#
¿
Feb 4, 2019 20:20
|
|
|
Mr Shiny Pants posted:Which firmware you running? I decided against upgrading because of some problems with Agesa. Are they fixed? I'm running the 1501 firmware for the zenith extreme, which isn't the newest one but everything seems to work. The only specific problem I'm aware of with the newest Agesa (which is still the TR2 version) is it advertising support for an EPYC-only feature, and causing then-new Linux kernels to hang. You could compile a kernel without support for that feature to work around it, but it really hosed me over when I was trying to get root on ZFS working, since I had never worked so closely with Linux's boot process before. The PCIe reset thing was fixed in one of the newer Agesa versions, either the one with support for TR2 or the one before it. It took AMD like a year to fix it though. Ryzen having a much larger install base and getting more attention with respect to bug fixes is a potential reason to go with Ryzen over Threadripper. Except for those two issues my only other significant problems have been Asus specific which is why I recommend the Taichi. The specific issues are not being able to choose which GPU is used for the BIOS, and the board wouldn't POST with two graphics cards if one was Turing for over a month after those cards were released.
|
|
#
¿
Feb 5, 2019 05:39
|
|
|
I had one of my ECC sticks go bad on me - and I'm glad it was ECC - leaving me down a quarter of my memory and I'm feeling the need to build a new system. I'd pay the early adopter tax but the initial batch of motherboards all look terrible and I'm not even sure I can make use of them. Connectivity-wise they look like budget PCIe 3 motherboards with a $500-1000 premium for PCIe 5 and 5-6 m.2 slots no one will manage to use. Even just getting two x16 slots that can do x8 + x8 is surprisingly rare when it should be the most obvious feature, especially on $600+ boards. Then I need another x8+ physical slot (x4 electrical would be fine), at least okay iommu groups - ruling out MSI with their bios update shenanigans, and ECC support if at all possible. From level1techs' first look, asrock seems to have working ECC but asus seems to have dropped it, which sucks for me since the asrock boards have bizarre and unusable pcie slot choices. If AMD hadn't killed off consumer threadripper - twice in a row - I'd try to wait for that, but even if I trusted them it'd probably be another year from now and pull 400 watts.
|
|
#
¿
Sep 27, 2022 06:45
|
|
|
Combat Pretzel posted:Hmmm, Asus seems to list ECC support everywhere. I guess that's a lie, IIRC this thread earlier? Yeah I don't know what is going on with Asus, when I made my post last page the detailed tech specs for Asus x670/e motherboards didn't even list memory details but they're back up now. Now even the ROG boards show a fairly standard "... ECC and Non-ECC,Un-buffered Memory*" with a disclaimer that the non-ecc supports on-die ecc. I guess I'll be waiting and seeing, could just be that they're working out the kinks; I've had very, very buggy Asus bios revisions on new products. For what it's worth my Asus x399 ROG board has working ECC but without working error reporting, at least in Linux. When it hits an uncorrectable error the entire system locks up, which is at least better than nothing and let me (eventually) realize I had a bad DIMM. Level1techs is the channel that claimed it wasn't working in one of their videos, and they care a lot about ECC, virtualization, and the other stuff I care about so hopefully they'll go back to it later. I'd go with asrock because of their historically good iommu groups and ecc support but their boards just have the worst physical or electrical slot layouts imaginable. Even then none of the Asus boards have a simple x8/x8/x4 set of slots - best they have is x8/x8/x2 and x8/x4/x4 - I'm fine if the third is off the chipset and slower but these limitations are bizarre, especially the board that can do x16 or x8/x4 but not x8/x8.
|
|
#
¿
Sep 28, 2022 00:05
|
|
|
I'll definitely be waiting for benchmarks if there's something going on with heterogeneous cores, especially phoronix benchmarks in Linux. Without some weird performance characteristics it'll probably be the 7950x3d. But heterogeneous cores are the main reason I'm not really considering Intel this generation since I want to give it at least another couple of years for Linux + qemu/kvm support to shake out. Though with the terrible motherboard selection any additional time I can wait to give them time to release better SKUs the better.
|
|
#
¿
Jan 5, 2023 08:01
|
|
|
https://www.phoronix.com/review/amd-ryzen9-7950x3d-linux Linux benchmarks including a lot of non-gaming benchmarks. Not quite the slam dunk I was hoping for but judging on current performance, not any ideas of them fixing the scheduler later, I think I'll go for it. I can pin my VM to one ccx or another for gaming which is where the biggest gains still are.
|
|
#
¿
Feb 27, 2023 18:45
|
|
|
I don't think anyone considering the 7800x3d should be impulse buying the 7950x3d right now. The 7950x3d should only really cannibalize 7950x sales with both 7950x variants positioned as low-end HEDT parts.
|
|
#
¿
Feb 27, 2023 19:32
|
|
|
I ended up buying/Amazon renting a 7600 just so I could get my system up and running to test the other components before their return windows close. Board is the Asus b650-creator, which has the least-bad pcie slot layout, with 128GB of micron ECC. 4800mhz/cl40 (stock) is achievable with 64GB but all four sticks defaulted to 3600 and wouldn't work at 4800. I haven't put in the effort to tighten up the timings or push the clocks too hard until I get my 7950x3d since it'll have a different memory controller, but just setting it to 4200 works and is good enough for now. Training after bios changes is ~2 minutes, but without changes it boots extremely quickly. I sure do miss proper HEDT platforms but Intel's stuff would add another $1500 to my costs and I don't trust Threadripper after how AMD screwed over TR4 owners.
|
|
#
¿
Mar 20, 2023 21:33
|
|
|
Well I finally got my 7950x3d. I could have had it earlier but my local store kept lying about when they'd have it in stock, so I put in an order on Amazon before newegg got restocked and just left it at that. Good news is I can add the asus proart b650-creator to the list of motherboards with fully functional ECC implementations including working error reporting. Bad news is even with a better IMC that can boot 4x32GB at 4800mhz it doesn't seem to be completely stable above 4200. A short memtest run didn't catch the errors, and it seems correlated with heavy CPU usage, but I can see corrected errors down as low as 4400mhz and across at least three of my DIMMs, making it unlikely to be bad memory. Maybe boosting the voltage to the IMC would help but I don't want to deal with that right now. Without ECC I might never have noticed since I ran a quick memtest and was running for several hours apparently stable at 4800mhz before the first uncorrected error crashed the system and I started digging around. I've also somehow gotten lucky and not run into any of the Asus weirdness this generation.
|
|
#
¿
Apr 19, 2023 06:40
|
|
|
I'm on the one Asus board (proart b650-creator) without a new BIOS update for this overvolting issue. On the other hand, I checked all my voltages and unlike the reddit users reporting 1.3-1.4 CPU/CPU SoC voltages in both bios and hwinfo, I see completely normal 1v and 1.1v values. I still want to see a new BIOS update but, to my best understanding, it seems like my board isn't overvolting the CPU. The x670 proart board did get a bios update so I can't speak for that, but I would guess the proart boards tend to be one step more "workstationy" than all the other consumer/gamer boards so aren't chasing benchmarks as much. Will still be watching for bios updates over the next couple of days but I might be in the clear. This might explain my lacklustre memory results compared to other people able to get 4800mhz stable at 128GB, if my SoC was running at 1.1v and I was comparing it against SoCs running at 1.35v. e: and a bunch of users reporting memory overclocks no longer working after updating their boards, which also tracks. Desuwa fucked around with this message at 20:14 on Apr 22, 2023 |
|
#
¿
Apr 22, 2023 20:12
|
|
|
That video also has damage in the same area on a non-x3d cpu and it seems like the cause is likely to be high SoC voltages from EXPO settings.
|
|
#
¿
Apr 24, 2023 23:55
|
|
|
Not sure I'd count EXPO as overclocking when things are advertised at those speeds. The manufacturers love to call it overclocking as a way of shifting responsibility when things go poorly (like this debacle) and being able to advertise things they can't guarantee.Lowen SoDium posted:So does the 7950x3D have this same issue on those bioses? I'd say any BIOS from the last couple of days is probably good, but we're still operating on anecdotes and guesswork. Board vendors have been pushing out BIOS updates to limit CPU voltages for x3d, and it does seem to be tied to voltages, but there have only been a few actual failures to go work from. As far as I'm concerned, if you have any AM5 CPU I'd make sure to grab any new BIOS revisions and also manually confirm the voltages are reasonable, x3d or not.
|
|
#
¿
Apr 25, 2023 02:39
|
|
|
Asus at least are rolling out another round of BIOS updates that extend the 1.3v limit to non -3d cpus.
|
|
#
¿
Apr 28, 2023 23:18
|
|
|
Based on my experience getting 4x32 to work, I would not trust 4x48 at any reasonable speeds without monitoring for ECC errors. Memtest should pick up errors but it can sometimes miss CPU-side errors especially if you end it early while trying to dial in an overclock.
|
|
#
¿
May 9, 2023 01:01
|
|
|
Subjunctive posted:What did you get 4x32 to run at? Without boosting any voltages or playing with timings, 4000mhz has been stable without any corrected errors logged. Even 4200 gave me one after a few days of actively monitoring it. Once all this voltage nonsense settles I might give it another 100mv to see if I can't get it stable at 4800. It boots just fine at 4800 but eventually an uncorrected error will cause a reset. Since I haven't played with voltages I don't really know how well 4 ranks per channel scales with voltage, but 192GB at 6000mhz does not sound believable to me. Even if it appears stable I wouldn't actually trust it myself, and I assume whoever was claiming it wasn't running ECC.
|
|
#
¿
May 9, 2023 05:56
|
|
|
I'd been running 1.0.0.7c for a while, and while it did increase stability it's still not perfectly stable with 4x32GB at 4800mhz. It's been about three weeks since I bothered to upgrade and clocked the memory at 4800mhz (was running 4000 before) but I got a corrected error today. This is close to how it behaved at 4200 before the update - going multiple weeks without any errors - so I'm assuming it'll be stable at 4600 or 4400, so there is an improvement but I can't help being a little disappointed. On the other hand I did leave everything else, including voltages, on Auto and I don't think I got the "right" memory chips that keep zen 4 happiest, so maybe tuning or buying the right chips would have helped.
|
|
#
¿
Oct 8, 2023 05:10
|
|
|
While the new threadrippers don't look completely horrible they're a year late - and there's still the chance of them releasing threadripper X3D in six months. That and them actively loving over threadripper owners before multiple times, with them allowing everyone to believe TR4 would be supported like AM4 and then only giving sTRX4 a single generation despite promises otherwise. The one thing that'd really get me interested is the amount of PCIe lanes, but even that seems unnecessarily exacerbated by the horrible slot layouts of AM5 boards. 24 pcie lanes isn't great but if you allow dropping to 8 lanes to each GPU plus the chipset providing some, there should be plenty for nvme and additional pcie cards, but the AM5 motherboards just suck. The non-PRO prices don't look nearly good enough considering you cannot count on AMD giving it more than one generation of support. If I were to have gone HEDT this generation, I'd have gone the Intel route months ago and have gotten better ddr5 speeds out of it.
|
|
#
¿
Oct 20, 2023 04:40
|
|
|
Especially with desktop systems, running overclocked RAM pushing the limits of stability, I am not convinced RAM errors are that rare. All the studies have been of servers running ECC in-spec. Even when the RAM itself is fine the memory controller can be unstable, and that can pass memtest for extended periods. I made the questionable choice of going with 4x32GB ddr5 this time around and even with BIOS updates I got ~weekly corrected errors (across all sticks) as low as 4400mhz and was able to pass short memtest checks at 4800. While I'm pushing the capacity that Zen 4 is willing to tolerate, I assume a lot of these just-barely-stable overclocks with two DIMMs are also not actually that stable on the memory controller side. I've also had two DIMMs go bad after multiple years of stability even without overclocks. One of those was ECC which was at least easier to conclusively narrow down.
|
|
#
¿
Feb 29, 2024 13:04
|
|
|
|
| # ¿ Apr 24, 2024 03:34 |
|
|
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
|
|