|
what do you think about ZFS and v-cache? I know ZFS loves having more metadata and ARC available and I could see it lowering average latency on SSDs/etc due to immediate availability of metadata. I know ZFS kind of has a problem with that due to being optimized in the days of spinners and maybe a big ol cache would help. I can't find any benchmarks though. also what do you think about optane as a metadata-only L2ARC to extend that further?
|
#
¿
Mar 8, 2023 05:38
|
|
|
|
| # ¿ May 21, 2024 03:41 |
|
|
Khorne posted:The part that annoys me most is that a 7950x3d will work better than a 7800x3d or 7950x for my use case. However, it's a product that is only appealing because they didn't release the product that best fits my use case. I'm sitting out until zen5 due to it. I was listening to MLID's (yeah, I know) most recent podcast tonight and he had an xbox studio engineering lead on and one of the things that he said was that he personally would rather have the dual-cache as well, but also that he thinks he would have uses for both clock-optimized ("7950XT") and the cache-optimized ("7950X3D") variants even just within his own studio... compiling loves X3D but there's also other stuff that is just number crunching (probably CAD work/graphics design). He thinks the X3D is really a fence-sitting SKU that satisfies neither camp, he thinks most people have a strong preference and they know whether they want cache or clocks and most people don't want the SKU that sits in the middle. And most people would probably pay more for a fully-clock-optimized XT with two good CCDs or two cache CCDs. And my take is, that leaves the possibility of heterogeneous as a fence-sitting base tier that is OK at both but lower-cost than fully optimized either way. (the other interesting nugget was, his opinion is don't buy the 7900X3D, because effectively that's a pair of 6-core chips and his opinion is we're about to shift into next-gen titles hard and 6-core chips are going to "begin screaming" as he put it. I think that really applies equally to the 7900X non-3D, the CCD barrier is kind of a cliff and games don't like suddenly having big latency jumps moving between CCDs, it's the same stuff you're talking about with discord audio crackling. So like, 7900X is a little iffy to begin with but a premium-priced dual-6C chip is an idiot trap... and really I have not encountered a single person who thinks 7900X3D is a good SKU as currently imagined.) Obviously you can Process Lasso processes to whichever CCD you like, but it'd be nice if Ryzen Master did that automatically. Ryzen Master can "help" windows scheduler too, doesn't have to ride totally on Windows, yeah we all know Microsoft are fuckups on the scheduler and I don't see that changing. To clarify one thing I said: I'm not sure heterogeneous will go away forever, I think there's a good chance XT and (fully-)X3D SKUs will become the premium models because that's a way to juice a little more ASP out of the silicon they have, and there's clearly customers for both products. But I could see heterogeneous becoming the base-tier for 7900X and 7950X especially with a poo poo-tier bin on the cache die (although again, that becomes a separate binning stream) but there's also arguments against doing that, it's more expensive than the current X configuration (although maybe if you're cranking out+stacking tons of cache it becomes cheaper if it's basically the default) and still loses to the current X configuration. But paradoxically maybe making the base tier shittier actually makes the premium models look better, it's a decoy product, nobody has to actually buy it, it just funnels people to the actual SKUs. I guess it doesn't make a ton of sense but who knows, big cache is a game-changer for some stuff, other people would love the flamethrower 6 ghz all-core 7950XTXX, and it's a way to push ASPs at a time when companies are desperate for that, I think fully-optimized versions are going to be available soon regardless of what happens with heterogeneous. On that note: Phoronix did a "for the things it helps, how much does it help?" piece. This is obviously a bit of a problematic number to try and wrap your head around, because it's inherently cherrypicked, it's "how big a number once we drop everything that's below average". But basically for those tasks the 7950X3D is around 18.5% faster (geomean) than the non-X3D model. Paul MaudDib fucked around with this message at 09:16 on Mar 12, 2023 |
|
#
¿
Mar 12, 2023 02:31
|
|
|
Threadkiller Dog posted:So my 6x2 5900x is already bad for gaming huh. no, especially if you got a good deal on it or have been enjoying it for a lot of years. but in a value-oriented market a premium midrange SKU at $550 or whatever is a tough sell when you could go cheaper with a 7900X or do the premium thing right with a 7950X3D. and the lifespan will be shorter than someone paying full price for a 12C in 2023 probably would want, if you don't want to do 8C for gaming then do 16C. the dev's thesis there is basically that hardware requirements are about to break loose a bit again as they cut PS4/XB1 legacy titles loose. And he might not be wrong either, we're pretty far into this "next-gen", it won't be all that many years until PS6 and a new xbox, and there's like, a total of 4 next-gen exclusives so far or something. (why do they call it ps5? less than 5 exclusive titles  )ok a few more have announced now. Paul MaudDib fucked around with this message at 10:12 on Mar 12, 2023 |
|
#
¿
Mar 12, 2023 09:26
|
|
|
BlankSystemDaemon posted:Yep, Stanley Pain got it in one. again maybe I'm missing it but why would having a device-managed subset of L1ARC being superfast not help? Or is the problem that with a LRU-style cache eviction mechanism (or something similar) that the fact that metadata and actual data both live in L1ARC? But if there's actual metadata cachelines wouldn't those still be accessed frequently compared to actual data lines and thus stay in L3 better and benefit from a bigger cache? like i'm viewing it from a "the bigger the cache, the less you have to stop the world and go out to memory or beyond" and ZFS seems like it could benefit from stopping the world less and holding more metadata close to the core where it can be accessed quickly. It's not that it needs to be directly managed by ZFS itself, it's that the working set of metadata lines is probably quite large and de facto that's your working set, ZFS has to walk the tree to find other blocks. It's inherently the (tree-)pointer-chasing style of algorithm that a big L3 is supposed to benefit, right? Isn’t it faster to have your next pointer you want to chase cached in? Paul MaudDib fucked around with this message at 10:49 on Mar 12, 2023 |
|
#
¿
Mar 12, 2023 10:18
|
|
|
K8.0 posted:Either your TV is a piece of poo poo or you're doing something very wrong. It's not any harder to get your PC to work with a TV than with a monitor, aside from TV-end stuff that you most certainly need to do for a console anyway. There's no difference between the output from a PC and the output from a console. The only thing that might be harder on PC is controlling when and how the TV wakes and sleeps. my (taichi ultimate+9900K, not AMD) and EVGA 3090 refuses to show any display output on my LG C1 until it gets into windows. It also hangs when trying to restart. So there's no way to get into BIOS with my C1  not that big an issue so I haven't taken the time to debug it
|
|
#
¿
Mar 14, 2023 03:46
|
|
|
Anime Schoolgirl posted:AMD probably will never do an AM5 Athlon 3000G successor especially considering how expensively set up the AM5 chipsets have to be why not simply rebrand the X300 chipset for entry-level Zen4 products  (actually while this is a meme, the joke is X300 means literally "no chipset", all AMD processors are SOC and are perfectly capable of running without one, technically this also means that there's no technical reason you can't put a X470 or whatever on an AM5 chip. They're just the world's most expensive IO expander card, in stark contrast to Intel they play no role in the processor bringup and you literally can just put an X470 on AM5 and it will work fine, there is absolutely nothing special about an "AM5 chipset" other than it having PCIe 5 capability (maybe). See also TR4 vs TRX40 and how that was bullshit when they maintained 3 generations of socket compatibility with SP3.) (naturally, Asrock has slapped one on a PCIe card  )I have to imagine that both AMD and Intel employ a full-time Asrock Wrangler whose job is just to stay on top of whatever crackhead thing their engineering department is up to this week and yoink them offstage if they start to act out in public Paul MaudDib fucked around with this message at 07:16 on Mar 17, 2023 |
|
#
¿
Mar 17, 2023 06:49
|
|
|
Arzachel posted:I need gently caress all for IO. Gimme a cheap chipset-less itx board you cowards! googled for "X300 mini-itx" and wouldn't you know it...
|
|
#
¿
Mar 17, 2023 20:29
|
|
|
5800X3D has a 65W TDP which is the lowest tier of "enthusiast" processors (technically things like the athlon 200GE slot in below it though). to be blunt it's the same power as a 3600 and it's really not demanding in any way so if your board can't handle a 5800X3D it's just not fit for purpose in the DIY space. it's a wonderful thing that you can get leading performance out of something with bargain basement-tier VRM and cooling requirements and near-total insensitivity to memory clocks. your board would have to be absolute loving garbage to actually struggle with it. go ahead and throw a 5800x3d onto an A520 board or whatever, it will be just fine
|
|
#
¿
Apr 6, 2023 07:24
|
|
|
BlankSystemDaemon posted:I heard of the Alveo U30, but the thing is that unless you were a hyperscaler, you genuinely couldn't buy it, no matter how many money your threw at Xilinx - so it never seemed worth looking into it. there's some weird products in this niche that aren't likely to cross consumer radar assuming they're even purchasable at all, like this one that was basically 3 skylake-R CPUs (with socketed SODIMM memory, and crystal well EDRAM side cache) https://www.youtube.com/watch?v=wnf6NwTgPZ0 but yeah I'm also extremely interested in the new AMD thing as well. It would be cool for doing VDI with accelerated streams as well... although the things AMD is advertising this for is, uh, going to be considerably less popular: Cygni posted:The previous generation card, the Alveo U30, doesnt have windows drivers and you can't buy it at retail. This isn't a consumer product. From the AT Article: you missed the best slide:  so, basically, the immediate application is gambling and home shopping network type stuff lol. maybe camgirl porn? but "micro-transaction revenue" is how they're framing it.
|
|
#
¿
Apr 7, 2023 23:00
|
|
|
BlankSystemDaemon posted:That's a loving hilarious daughterboard; I bet Intel were thinking "this is how we're gonna make big bugs from our Scalable Video Technology" the choice of crystal well is kind of weird at first but I was looking through phoronix and video encoding (x264 and x265) heavily benefit from v-cache so I bet video encoding a bunch of streams in parallel needs a pretty decent working set even with iGPU-based encoding. video encoding is a lot more intense than people give it credit for, and the idea of CPU encoding your stream while you game has always been a little... dubious
|
|
#
¿
Apr 7, 2023 23:38
|
|
|
BlankSystemDaemon posted:That's a loving hilarious daughterboard; I bet Intel were thinking "this is how we're gonna make big bugs from our Scalable Video Technology" SVT is software encoding, I think the idea here would be to turn the CPUs into iGPU video encode offload accelerators, literally using CPUs just for the GPU perf (and throwing a bunch of side cache at it for a dedicated scratch space). Or at least using client/laptop quad-cores is not gonna be that fast, probably within similar reach of a config you could build with normal server CPUs (albeit probably a more expensive server, but I doubt that would have been a cheap product either). It's somewhat interesting that a number of companies have made a run at the hardware video encoding market for a number of years. Rabb.it had custom servers with a bunch of jetson tk1s inside a 1U with some NUCs controlling them (arise, methenar). Intel made this video encode board. Google reportedly commissioned AMD to make RDNA1 HEVC really good for Stadia, because they had the same business need too. Now Xilinx/AMD have another take on it (without any attached GPU compute this time, to drop costs per stream). I guess it's a very common need to just encode some video quickly for whatever use-case, video-encode-as-a-service tailored to some segment. I kinda thought that had already been fairly democratized, it's not like it hasn't been a thing NVIDIA has addressed for a while. the Grid series were basically designed for these sorts of things (VDI, teleconferencing, etc) even 10 years ago, and more recently a t4 card is a very cheap buy-in for a hardware nvenc encoder with turing quality. And hell every phone SOC (raspberry pi etc) has come with video encode for a long time. It's interesting that apparently a quality-optimized encoder running modern codecs at the lowest possible cost ($1600 is fairly cheap for server given what it does) makes a relatively modern ASIC on probably a decent node (6nm? 5nm?) a viable proposition. I guess AMD see a big future in AV1 content delivery/encoding, and in all fairness it could result in tangible savings for companies like Google (your youtube upload never watched? now it's squished to av1 for space savings), although of course at a certain scale they can also commission whatever ASICs they want (custom 16nm or even 6nm isn't all that insane if you're a hyperscaler). I wonder what market penetration looks like for AV1 decode these days. Based on the slide Cygni linked, it sounds like maybe it's also grid compute stuff, video encoders for your PS Now and stadia and other streaming services. But again the question there is what you're attaching them to that doesn't already have an encoder - consoles do, dGPUs do, phones do, almost every PC does. Is your quality enough better to warrant the cost than just having the client's laptop do it (if they're cheap, then welp quality sucks, sucks to be you)? If it's just pure VDI, isn't having the dGPU portion of the card an advantage to accelerate the desktop? Twitch/onlyfans datacenter re-encoding, or youtube archival data compression - sure I guess, but how big is that market? It's a real cool product, I'd love to have one for turbonerd reasons, but it's a mildly perplexing product in terms of market fit, like who does this product fit that wouldn't just buy a bunch of quadro A4000s or T4s instead? twitch and teleconferencing I guess? I guess the answer may just be that this was developed before Xilinx was acquired and these are the pieces of the product they had in 2019 or 2020 or whatever. Paul MaudDib fucked around with this message at 02:44 on Apr 8, 2023 |
|
#
¿
Apr 8, 2023 02:05
|
|
|
Combat Pretzel posted:Put a BMC on all boards. Intel did that a long time ago, it’s called vPro
|
|
#
¿
Apr 17, 2023 17:10
|
|
|
Perplx posted:my board has a com port header doing nothing, just send the whole boot log over that like linux does, you could get real detailed, like tell you what dimms were detected and the speeds and voltages of everything, hell make it interactive linux usually supports actual TTYs still. That's why virtual terminals are like TTY7 or whatever. I saw someone printing a neofetch recently (  ) it would actually be funni to see a real TTY running modern unix.
|
|
#
¿
Apr 17, 2023 18:07
|
|
|
FuturePastNow posted:just validates my decision to never overclock anything it seems this might have come simply from setting "XMP" (Expo is AMD's equivalent for DDR5, like DOCP for DDR4) and not touching other voltages directly. Asus may have done some weird "moving memory IO voltage also moves VSOC" and iirc also a third and they're all geared 1:1 by default. And 1.5v or even 1.35v is really loving high for VSOC. Rational values are like 0.85v to 1.25v, so like, you're fine at JEDEC but anything past that might step up to Bad Things. on the other hand the burned pins are definitely in VCore area. Which wouldn't be a VSOC thing I guess. Finally you've got seemingly nonfunctional thermal limits. chips shouldn't be unsoldering themselves regardless of voltage death at all. that means thermal limits were breached significantly for a very prolonged period of time - you can soak the thing at 100 or 110c forever. if that's a 150w soldering iron dumping heat into your die it'll eventually melt the solder yeah, and the die getting that hot for that long means it must have significantly breached power limits for a bit and it kept going despite package temps being super-hot, like 200C (at which point you might get MOSFET runaway death anyway). And it's got a weird lumpy pattern, it didn't melt evenly. it's difficult to tell right now how all of these fit together, there may be multiple separate issues, or just one/they're related, or they may actually be causal. For example, a runaway thermal hotspot might lead to a lumpy desolding thing, and maybe that hotspot is where the burn happened (there are a lot of vcore pad areas!). Or maybe these are unrelated and asus just hosed up thermal monitoring and is melting cpu lids off but the memory controller is being zapped to death by a separate bug. I mean we'll see what the final diagnosis is but the damage pattern really looks like Asus hosed up and released a quarter-baked bios that does some super questionable things with voltage, and lets some voltages go way higher (perhaps functionally unlimited) than they should. Does it lead to some weird edge-case where a core runs away or the cache die runs away etc? I think there's a serious possibility that it's actually zapping quite a few more chips than X3D but X3D just happens to be a delicate flower that dies more easily (the first desoldering CPU was a 7900X). I wonder if those users sustained damage, there have been a few weird reports of memory controller stability loss (but it's always hard to know how seriously to take that with the state of DDR5 stability) even just from undervolting (!), or people who lost a decent amount of max boost, etc. But boost algorithms get patched and people have placebo effect/etc. Or they did the paste worse, or the paste pumped a bit, or the memory retrained and sucks a bit now, etc. Boost algorithm is sooo complex now and all kinds of things affect it. A few Gigabyte boards have failed too, so it's hard to know how seriously to take the Asus pattern. Or gigabyte could have hosed up too, and perhaps others. It's always hard to figure out what's internet pattern-sleuthing and what's just boston-marathon sleuthing. It is unironically always super hard to know whether there are real patterns, or people that aren't reporting it because "mine's not an asus", or people doing things wrong (touching pads etc), or marketshare differences etc. Crowd-sourced failure reporting loving sucks as a diagnostic tool unless it's professionally verified (eg, GN contracts a failure lab). one other thing to ponder - the cache die is on a customized nodelet with optimized voltage profiles/cells for big SRAM blocks, the SRAM is actually denser on Zen's v-cache dies than you can get on the "logic-optimized" variants. so it's entirely possible that the cache die is quite a bit more delicate in the "doesn't like 1.35v or 1.5v pumped into it" sense and maybe it's got easier MOSFET failure etc. But like, the cache die is not necessarily quite the same nodelet as a 6nm logic die. They very well may be serious about the "no for reals you can't do more than X voltage" for like, actual valid technical reasons, and even modestly pushing the bound like Asus accidentally (may have done) might cause big problems. Like they might be right at the limit of "what we need to drive the CCD with big boosts" and "what works with this cache nodelet before it blows up", independently of the heat/power problem. 1.5v may be like, a lot for this nodelet, they may be right there with the 1.35v official limit. You're not allowed to play with voltage so nobody knows, and this may be the answer, there is no voltage headroom on the cache die before things release the smoke. AMD is absolutely a prime candidate for something like DLVR/FIVR themselves. Cache die really kinda needs its own rail, and ideally it should be the kind of response and precision that onboard voltage regulation gets you rather than an external VRM rail because it's a delicate loving flower. That frees up the CCD core voltage to do its own thing to work with boost. Paul MaudDib fucked around with this message at 06:23 on Apr 25, 2023 |
|
#
¿
Apr 25, 2023 04:33
|
|
|
kliras posted:game of telephone caveats and all that Again, the caveat with this theory being that VCore is the only voltage the cache die gets, and none of those voltages have anything to do with VCore. VSOC and VDIMM aren’t fed to the cache. The burned pads consistently in the VCore region are suggestive that it is VCore, and it happening on X3D chips in particular (which have an obvious reason to be sensitive towards VCore) makes sense along those lines. If it’s caused by zapping the VSOC then it would damage regular CPUs at an equal rate. (which I have said is likely a distinct possibility too but the failures do seem to point towards X3D being hit at a higher rate - but again that also can be explained by enthusiasts having more X3D chips than the public as a whole. Crowd-sourced failure reports suck rear end, they’re just all we have until someone gets a failure lab report.) Again, it’s possible there are actually multiple issues or failure modes here, for example if AMD’s reference AGESA implementation (it seems) doesn’t enforce any voltage or thermal limits and there’s also this other issue around the VDIMM/VSOC Paul MaudDib fucked around with this message at 15:11 on Apr 25, 2023 |
|
#
¿
Apr 25, 2023 14:39
|
|
|
gigabyte dropped new bioses too. I think this is a “update your bios” moment, i wouldn’t sleep on it regardless of brand or processor. If the VSOC thing is true it could be killing non-X3D too and it seems to have been part of AMD’s reference implementation.
|
|
#
¿
Apr 25, 2023 17:48
|
|
|
hobbesmaster posted:What’s funny is from the standpoint of someone that has only personally tuned/OC’d AM4 even the normal VSoC settings on zen4 seem insane. Yes, the idea of 1.3v being normal is kind of crazy, AM4 is more like 0.8-1.1v right? I get the impression this Io die is really really iffy, AMD has always been a bit crappy compared to intel (apart from like, X99) but this time they can’t even run all the slots populated etc and intel didn’t have those problems even on their first gen DDR5 stuff (or X99!). If there ends up being a new IO die next gen it’ll be interesting to see how it does.
|
|
#
¿
Apr 25, 2023 19:11
|
|
|
It’s all about thermal deltas. If the hot side is X degrees hotter than the cold side at steady state, for a cooler of a given quality (delta-t over watts? Or whatever) you’ll move Y watts. Any given cooler can move more watts at a higher delta-t, and vice versa. Like the reason blower coolers run hot is they just need a really high delta-t because they need a high temperature differential to move any heat because they suck. Now if you really want to push maximum cooling what you need is not a radiative cooler but a heat pump  https://m.youtube.com/watch?v=185VQ_d2bYU (h/t Swiss Army Druid forever ago, I believe)
|
|
#
¿
Apr 26, 2023 06:12
|
|
|
HalloKitty posted:Strapping a freezer to your cpu used to be a "mainstream" thing, in that you could run out and buy a Prometia or Vapochill case ready to go. the novel thing is that this is the difference between an A/C unit and a heatpump. this is a really direct way to move heat from one side to the other - it's a mechanical peltier. And the other side gets really hot, but it's air, so, if you can design around expelling a giant jet of 400C air then awesome! over a million rpm
|
|
#
¿
Apr 26, 2023 08:49
|
|
|
Combat Pretzel posted:Scuttlebutt is that AGESA 1.0.0.7 fucks with memory speeds, dropping things to DDR5-4400. Not sure if it’s on profiles where things are kinda automagic, or if it overrides entirely manual settings, too. AMD’s ostensibly working on v1.0.0.9. I rather doubt 192GB was stable at DDR5-6000, and if it ever was it would require absolutely titanic VSOC voltage which is exactly the thing that was killing chips
|
|
#
¿
May 8, 2023 22:25
|
|
|
Cygni posted:7800X3D already $30 off at microcenter, with $20 motherboard credit. Think AMD's sales are probably real painful at the moment. please pop off if you see the discounts get real serious or anything, I'm not opposed to jumping the gun on my "waiting until 2nd gen" if AMD wants to cut their own throat. like if the memory bundle gets extended or whatever that'd be sick anyone besides asus got thunderbolt motherboards, ideally thunderbolt+10gbe? I'm sure asus will be fine eventually but right now there aren't a lot of alternatives to the ProCreate X670E Paul MaudDib fucked around with this message at 05:03 on May 13, 2023 |
|
#
¿
May 13, 2023 05:01
|
|
|
Cygni posted:Yeah one of the key pieces of advice for life happiness I can offer is: do not install any of your motherboard manufacturers software under any circumstances. clean install krew with my no Windows Binary Platform privilege
|
|
#
¿
May 13, 2023 07:48
|
|
|
Taima posted:Wait am I not supposed to have PBO on my 7800x3d either? oh HEH HEH this guy doesn't understand the difference between PBO and PBO2 well you see 
|
|
#
¿
May 13, 2023 10:32
|
|
|
Core “complex”? Actually I find it quite shrimple
|
|
#
¿
Jun 10, 2023 07:29
|
|
|
stacking happens after binning, and there are no defects that are going to happen during/after stacking that are going to leave you with a functional chip in any respect, if it fails during stacking it's dead. just a reminder but all epyc v-cache chips are 8-CCD with 768MB of cache on them. so for example a 7473x is 16C, and that means they're stacking v-cache on 2-core CCDs. Since AMD makes 16, 24, 32, and 64c models, there are epyc binning streams for 2c, 4c, and 8c CCDs, but not 6c ones like 5600x would use, since there is no 48C epyc v-cache SKU. if there are too many 5800X3D chips made, or too many epyc 8C dies stacked, they could be binning down 8c chips into 6c ones for salvage - just like they turned 7900X into 7600X and 7950X into 7800X. CPU sales (especially consumer sales) are miserable overall and AMD is having to shuffle inventory around to the SKUs that are actually selling. Or they may have excess space on the assembly line and hey, we're selling every 5800X3D we can without dropping prices too far, add a new 5600X3D sku instead. but it's not a case of core harvesting or clock harvesting or epyc defect harvesting in general. it's rare for a chip to fail that late, we are talking sub-1% in general, but when they do fail the whole chip is dead at that point. Same for the dual-CCD 7600Xs/etc - there is no meaningful amount of 7900X that fails during packaging and results in a viable IO die and CCD, such that two random youtubers get one within a week. at the current price it's pointless, but if it drops to $200 there's a market for that. It'd be an esports champ. the other possibility is since 5800X3D is getting so far below MSRP that AMD is going to slot a new SKU and stop dropping 5800X3D so much, so it may be a case of "your choices are now 5600X3D at $250 or 5800X3D at $329" rather than $250-vs-$280.
|
|
#
¿
Jun 14, 2023 22:07
|
|
|
hobbesmaster posted:That feels alarmingly close to the kind of thing that accidentally starts WWIII in a Clancy novel. There are some fish that are too big to gently caress with. When it becomes a geopolitical hot-potato you unlock it and do your best to disappear into the woodwork and hope nobody's too mad. Like colonial pipeline, the russian group apologized and announced they'd had a sudden desire to improve their manners, were unlocking everything they'd ever done, and would disband as a group. Some wild swing in gas prices during COVID economic disruption year 2 was exactly the opposite of what anybody needed, and there's also nobody on the planet who's not affected by TSMC disruption let alone the possibility of losing a month of production or w/e.
|
|
#
¿
Jul 1, 2023 02:34
|
|
|
orcane posted:Paul??? 
|
|
#
¿
Jul 3, 2023 22:38
|
|
|
Dr. Video Games 0031 posted:MicroLED displays already have internal refresh rates of 12,000 Hz or more. And there are quarter-inch 5,000-DPI MicroLED displays in labs currently, so I figure it's just a matter of time before we end up reaching the goal of a "perfect" display where both the resolution and refresh rate are perceptually infinite. The only limitations on resolution and refresh rate will be the source hardware and available signal bandwidth. The point wouldn't be to actually output at those absurd resolutions but to be able to scale from any arbitrary resolution to the display's resolution with no perceptible scaling artifacts. It's probably 10 - 20 years before this ends up on our desks though. imo at that point you basically forget the idea of line scanout and just go for a minimum-error "compression" codec, and bear in mind that at 48gbit/s or whatever you can get a very low error minimum. but the idea to me there would be you draw macroblocks instead of lines, or you choose the lines that yield the minimum-error encoding. you update the perceptually-optimal parts of the image. and again, I think that fits very interestingly with the idea of variable-rate spatial+temporal sampling with DLSS3+. You can perceptually sample that edge where the enemy is emerging at like 10,000fps and update those pixels on the absolute hotpath etc. Like I haven't looked into the tech at all but just as described, being able to draw specific bits at super high refresh, or specific lines at super high refresh, is potentially super cool. If you can do that, why draw a beam-chasing pixel stream instead of a compressed video stream format with block/macroblock updates? one of my favorite videos, this does it with just line encoding on an IBM PC. https://www.youtube.com/watch?v=MWdG413nNkI https://trixter.oldskool.org/2014/06/19/8088-domination-post-mortem-part-1/ (but it's 48gbps, and writing macroblocks instead of lines) Paul MaudDib fucked around with this message at 08:04 on Jul 4, 2023 |
|
#
¿
Jul 4, 2023 07:48
|
|
|
people generally need to suck it up and dwi, 7000 series is getting cheap enough it's pretty much just better now. Getting 7600X or whatever is really fine too. Especially now that 4x32GB+ appears to be getting stable at reasonable clocks on gear 2. that's fine, especially with the 7800X3D. with some of those skus being APU cutdowns probably, people should carefully consider the implications of the particular SKU they're buying in terms of pcie lanes (some may be x8) and cache configurations, reduced cache on APUs will hurt games somewhat. I, again, think 7600X bundle (often less than 7600) and 7800X3D are the ones that make sense for most people. 7800X3D is a value winner that helps drop your VRM power while boosting performance, get that Asrock Lite whatever. I used to still think AM4 would be better for UDIMM server builds because of 4x32 easy compatibility and I think AM5 is probably good enough to not matter now, (for gamers) especially with the ram bundle deals. otoh, zen5 is going to be a big deal when it comes out in probably end of next year or w/e. legit big improvement, zen4 was conservative to port onto DDR5 and zen5 is the big rewrite. it should be fairly significantly better, although I think they're taking the same big/little dive a little bit and gaming enthusiasts still want 8C X3D all-p-core type skus. Paul MaudDib fucked around with this message at 08:00 on Jul 23, 2023 |
|
#
¿
Jul 23, 2023 07:55
|
|
|
https://www.youtube.com/watch?v=1nu-GKq58Og minisforum appears to have stepped it up. this is a really nice looking mini pc, all metal, quiet, upgradeable, but using LM in the cooler so don't open it. 2x usb4/thunderbolt allow either dock/display expansion or some interesting networking possibilities. the triangle networking (with 2-port) is a very common strategy for small clusters in general. works with 2-port enterprise ethernet adapters or infiniband adapters in homelab servers too, just direct-connect your way to success. QSFP/QSFP28 copper DAC or SFP+ copper DAC cables are cheap, as are QSFP28 to SFP+ breakout cables.
|
|
#
¿
Jul 30, 2023 09:22
|
|
|
repiv posted:naughty dog switched over to AVX2 for the PS5 branch of their engine and didn't bother adding an SSE fallback to the PC version of TLOU1, it won't run without AVX2 support pre-Haswell/pre-Ryzen is coincidentally also when NVMe support started to be rolled out, and you can't play the game without NVMe storage either. it's easy to justify not writing that pre-AVX2 path when the users mostly can't play anyway for other reasons. (technically you can use an adapter card but that's either going to cut your GPU lanes in half, or be on a chipset slot that stands a decent chance of being like pcie 2.0x1 or at best pcie 2.0x4, perhaps literally 1.1x1 speed. And while you technically don't need actual BIOS support (especially if you are not booting from the SSD/have a SATA drive for a system drive) from what I remember the Oprom thing is pretty clunky and a lot of drives don't ship with oproms anymore cause nobody actually uses them anymore. I honestly have no idea if oproms are even viable anymore for boards without native support.) Paul MaudDib fucked around with this message at 00:44 on Aug 9, 2023 |
|
#
¿
Aug 9, 2023 00:40
|
|
|
Twerk from Home posted:Wait what, games actually require NVMe storage now? TLOU at that? Wow, I assumed that SATA SSDs would just make for slightly slower load times. whoops I was thinking Ratchet and Clank and mixed up the studio (nixxes, not naughty dog) but yeah ratchet and clank doesn't run well on SATA SSDs. it's still loads better than a HDD, but there's definitely some stuttering in john's stream on the "minspec PC" https://www.youtube.com/watch?v=I71u7j1Izww
|
|
#
¿
Aug 9, 2023 01:32
|
|
|
Klyith posted:The other question to ask is: how many things are actually bandwidth limited by 2 channels of memory on a normal desktop CPU? ... Yep. I honestly think this was part of the calculus for keeping quad-cores for so long too. that's what the business world wants as a "decent but not extravagant workstation" (not HEDT) in the 2012-2017 era. the drive for higher cores wasn't there, and if you were a prosumer the HEDT platform was a lot cheaper and more accessible.
|
|
#
¿
Aug 22, 2023 04:56
|
|
|
Anime Schoolgirl posted:https://www.youtube.com/watch?v=raf_Qo60Gi4 this game is so hosed that it seems premature to draw conclusions. wait for the first 5 patches and 2 gameready driver cycles and see what happens. in fallout 4, creation engine liked memory bandwidth, and cache size, and having SSD available (NVMe probably mandatory now). maybe v-cache is just more than Creation can utilize but that's also a somewhat surprising/counterintuitive finding in a game as badly programmed as that, with the history of that. maybe it's preferring the larger L1/L2 cache of alder and raptor over previous gens, and also the generally higher bandwidth? does 5800X3D/7800X3D do good in ARMA3/Fallout 4? Those are sort of the bellwether memory bandwidth titles I'm thinking of and I'm curious whether X3D helps in that situation. Def does in counterstrike etc. the "GPU busy" metric is also amazing and it's cool to see it be rolled out. it'd also be so supremely hilarious to go back and see who was getting tons of driver overhead in X thing. Witcher 3, Crysis 3, etc. Paul MaudDib fucked around with this message at 04:36 on Sep 5, 2023 |
|
#
¿
Sep 5, 2023 04:34
|
|
|
Pablo Bluth posted: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. this was always true of threadripper, though? it's always been a cut-down epyc, and development costs have always been "minimal" and has always benefited from riding on epyc's coat-tails, but they still half-assed it anyway
|
|
#
¿
Oct 19, 2023 16:40
|
|
|
$1500 isn't an increase at least, but it's still ultimately 4x R5 7600 on a package, with half the memory channels cut off, etc. That should be a $700 product, not $1500, but it's "what the market will bear". And then you'll have the $1500 motherboard for it. And again, regardless of it being what AMD thinks the market will bear, the fact is that the epyc boards and chips are literally better and cheaper for most people, so this is another "waste of silicon" launch. the server market "won't bear" $1500 motherboards and produces very good 8-channel motherboards kitted out with PCIe lanes and poo poo for $700, but then we apply the gamer tax and threadripper is $1500 for a board with much less capability. Chips are cheaper too, once the used market gets into gear (which is of course why they've gone after that with platform lock/etc) https://www.asrockrack.com/general/productdetail.asp?Model=GENOAD8X-2T/BCM#Specifications in the server market, these are $931 on provantage right now (the predecessor on SP3, ROMED8-2T, was $600). AMD wants you to pay $1500 for the threadrippper-lite or $2-3k for the full threadripper equivalent. it's just pretty much a non-starter if you care about value Paul MaudDib fucked around with this message at 16:58 on Oct 19, 2023 |
|
#
¿
Oct 19, 2023 16:48
|
|
|
Subjunctive posted:Jealous that they can fill all their DIMM slots, unlike my 7800X3D! have you tried again after the new microcode? supposedly memory is a lot better now. Kivi posted:Now if any of you US goons were willing to do a customs fraud and buy and send me a https://www.newegg.com/asrock-rack-romed8-2t/p/N82E16813140044 as that's also the lowest that board has been. oh gently caress, I knew about the mITX one but I didn't know romed8-2t was on sale... 😭 time to go look up epyc X and F sku pricing again... edit: they also have a dual-socket version on sale too. https://www.newegg.com/asrock-rack-rome2d16-2t-amd-epyc-7002-series-processors/p/N82E16813140060 BlankSystemDaemon posted:Sure, you can do cpusets/pinning/affinity, but then you're not using all those threads you bought the CPU to have access to, so what's the point? sure but the REAL question is whether you can do one socket with F skus and one socket with X, make your own big.little Paul MaudDib fucked around with this message at 20:24 on Oct 21, 2023 |
|
#
¿
Oct 21, 2023 20:21
|
|
|
give me a strix halo APU for the threadripper-lite platform (siena?) you cowards
|
|
#
¿
Nov 8, 2023 22:55
|
|
|
|
| # ¿ May 21, 2024 03:41 |
|
|
Tuna-Fish posted:Parametric yields are a thing. There is no way 90%+ of vCache Zen3 dies can do the 5800X3D spec, if they could, you would assume you could find a lot of golden samples that could do much more, there's just too much variability for that. That's why you have lower-tier SKUs that have lower clocks, so you can sell everything you can make. CoWoS (oWo) stacking comes after binning. you know how many chips can do it before you stack them. there are also very few stacking problems that leave you with a fully functional die other than a single core, tbh AMD maintains binning streams for 2C, 3C, 4C, and 8C stacks. They have never released a 6C Epyc X3D stack - that would be a 48C Milan-X. Doesn't exist.
|
|
#
¿
Nov 20, 2023 07:48
|
|