|
https://twitter.com/VideoCardz/status/1636331643945320450?s=20 lol
|
#
?
Mar 16, 2023 13:18
|
|
|
|
| # ? May 13, 2024 05:33 |
|
|
I know it’s a workstation / server card so power is kind of a different beast but lol at the processor alone drawing more than a regular 15-amp breaker could provide.
|
|
#
?
Mar 16, 2023 14:20
|
|
|
Not if you live in a country with decent electricity standards (3000w on 13a sockets, or 3680 on 16a)
|
|
#
?
Mar 16, 2023 14:30
|
|
|
The fact they can torture the silicon to that extent without turning to slag is pretty amazing.
|
|
#
?
Mar 16, 2023 15:12
|
|
|
From the comments quote:The old phrase 'full pot' will go away soon and be replaced with 'full hose'
|
|
#
?
Mar 16, 2023 16:42
|
|
|
KKKLIP ART posted:I know it’s a workstation / server card so power is kind of a different beast but lol at the processor alone drawing more than a regular 15-amp breaker could provide. Why yes I need a NEMA 5-30R installed just for my computer.
|
|
#
?
Mar 16, 2023 16:48
|
|
|
Beef posted:The fact they can torture the silicon to that extent without turning to slag is pretty amazing. It’s “only” 1.05W/mm^2. The minimum for a welder in about 10W/mm^2
|
|
#
?
Mar 16, 2023 16:49
|
|
|
Pablo Bluth posted:Not if you live in a country with decent electricity standards (3000w on 13a sockets, or 3680 on 16a) Nah, Americans are perfectly able to pull out the full 220V when they need it. These might just be sour grapes but for most purposes I'd prefer to have a 110V line that's less likely to kill me if I touch it.
|
|
#
?
Mar 16, 2023 16:50
|
|
|
Eletriarnation posted:Nah, Americans are perfectly able to pull out the full 220V when they need it. These might just be sour grapes but for most purposes I'd prefer to have a 110V line that's less likely to kill me if I touch it. Bad 220v wiring sounds a lot cheaper than a sound retirement strategy in 21st century America.
|
|
#
?
Mar 16, 2023 17:08
|
|
|
On the other hand, we don't have a culture of electrocuting ourselves with male-to-male extension leads...
|
|
#
?
Mar 16, 2023 17:41
|
|
|
Eletriarnation posted:Nah, Americans are perfectly able to pull out the full 220V when they need it. These might just be sour grapes but for most purposes I'd prefer to have a 110V line that's less likely to kill me if I touch it. To be fair plenty of countries with the higher voltages have RCDs/GFCIs protecting the entire house, all the circuits.. not just random ones in the bathroom or kitchen Plus socket/plug designs that don't directly expose you to mains voltage every time you insert and remove them vv I'm just saying that those things somewhat reduce the risk of you getting a shock in the first place HalloKitty fucked around with this message at 20:19 on Mar 16, 2023 |
|
#
?
Mar 16, 2023 20:04
|
|
|
I mean, yeah, those safety features are all nice as well. I was just making a statement on the voltage standard, not all the electrical standards.
|
|
#
?
Mar 16, 2023 20:15
|
|
|
hobbesmaster posted:It’s “only” 1.05W/mm^2. The minimum for a welder in about 10W/mm^2 I thought you'd misplaced a decimal in your calculation, but nope, 1908 mm2 for Intel's top end CPUs  That's over 50% more silicon for 36 fewer cores than Epyc.
|
|
#
?
Mar 16, 2023 22:53
|
|
|
HalloKitty posted:To be fair plenty of countries with the higher voltages have RCDs/GFCIs protecting the entire house, all the circuits.. not just random ones in the bathroom or kitchen And in the UK you’d be able to blow all the fuses in your new house because they still build ring circuits in the 21st century.
|
|
#
?
Mar 16, 2023 23:15
|
|
|
I know someone who wants to know if a 13900k would be good for doing video editing work. The catch is they tell me they do this in Handbrake. On Linux. Don't ask me why. The problem is I can't seem to find a definitive answer for how well Linux' scheduler has adapted to the P/E core split, with most of the results I've seen describing it as a "work in progress" but that could mean anything and the most recent I saw was still late 2022. Does someone here know?
|
|
#
?
Mar 17, 2023 02:23
|
|
|
If you are using a distro with updates in the last few years, it will be fine. https://www.phoronix.com/review/intel-core-i9-13900k Caveat being I don't have personal experience with it.
|
|
#
?
Mar 17, 2023 02:42
|
|
|
power crystals posted:I know someone who wants to know if a 13900k would be good for doing video editing work. The catch is they tell me they do this in Handbrake. On Linux. Don't ask me why. The problem is I can't seem to find a definitive answer for how well Linux' scheduler has adapted to the P/E core split, with most of the results I've seen describing it as a "work in progress" but that could mean anything and the most recent I saw was still late 2022. Does someone here know? Handbrake isn't an editing tool, more of a compression tool. As someone who uses it regularly, I've got some thoughts. A 13900 actually has more cores than most common Handbrake jobs can use effectively! Handbrake runs one job at a time, and video encoding has some limits on parallelism. You can run multiple Handbrake instances to work around this, but it's clunky. Handbrake is open source so I'm going to poke around and see if it's feasible to make one Handbrake run multiple jobs from the queue at the same time instead. Also, video encoding is really vector heavy, and the E cores have pretty mediocre vector performance. That usage case sounds like a job for the AMD 7900 or even 7950X, both of which have better all-core vector performance. The encoders are also starting to use AVX-512, which is a big advantage that the AMD chips have, but the Intel 12th and 13th gen Core products don't. Intel 11th, and laptop 10th (and even some 8th!) gen stuff had it, Intel are giant fuckups that have confused their product line. If you look at benchmarks for video encoding, you'll see that a 13900k can run 1 job the fastest or trades blows with the 7950X, but if you've got a lot of encoding to do and run 10+ jobs at once with fewer threads for each, the AMD chips are just plain faster. A little counterintuitive, but I'd go AMD for that. If you want to see how bad Intel hosed up with AVX-512 and how impactful it can be, look at workloads that use it:  That said, AVX-512 is just a 10-20% boost to video encoding right now, there hasn't been a ton of effort put in because it's not ubiquitous and also Intel's implementation on server chips used to be really power hungry and made the CPU down clock significantly. Edit again: I checked Handbrakes documentation, and they still say that it scales poorly past 6 cores, so if you've got 24 cores you would absolutely want to run 4+ Handbrake instances or just use ffmpeg instead, with 1 job per core. Twerk from Home fucked around with this message at 04:16 on Mar 17, 2023 |
|
#
?
Mar 17, 2023 04:04
|
|
|
ConanTheLibrarian posted:I thought you'd misplaced a decimal in your calculation, but nope, 1908 mm2 for Intel's top end CPUs golden cove is something like 3x the transistors of zen3. the reason AMD isn't doing a full-on e-core is that they don't need it - their p-cores are already relatively compact and space-efficient, they just are going to squeeze the last 10-20% by using high-density libraries and microcoding AVX-512 support and call it a day. Zen5 will probably be bigger but Zen4c is perfectly adequate as a space-optimized e-core, zen3 is only something like 1.5-2x the transistors of Gracemont but it also has SMT which Gracemont doesn't. That's fine imo. (btw Intel's actual scaling factor is closer to 3 e-cores per p-core, not 4 - the reason the number "4" comes up is because it's a 4-core CCX, but a 4-core e-core CCX is actually slightly larger than a p-core, despite that being the marketing.) it will be interesting to see whether Intel continues the development of the Cove line - I think the long-term future of Intel's server line lies in Sierra Forest not Sapphire/Emerald/Granite Rapids. And the question is, if you're Intel and you're cleaning out the "take a penny" dish in the cafeteria to keep the lights on, do you pay to develop two whole parallel lines of products? I kinda wonder if they will just milk the last out of the Golden Cove line and call it quits. Not that heterogeneous won't continue into the future, it's entirely possible they re-develop a Mont-based p-core and throw some smaller Mont e-cores. But in a backhanded way, heterogeneous may be less about big.little on the desktop and more about letting Intel start developing the necessary pieces to move forward without having to throw everything away. Monts have a 4-core CCX, so you have a ringbus with 4 cores per stop instead of 1 core per stop. And you can get the Mont series into actual mainstream consumer products and get software optimized for it without having to dump it into the market and be like "we're doing something totally different, re-optimize everything right now plz". Even with Atom NUCs and similar, I'm not sure that's the same thing as having them on your mainstream client platform, Intel has got the Monts into (almost) every desktop they sell for the last 18 months. And then Sierra Forest starts that on server in another 2 years or so. That's the start of the Merom-style pivot. Paul MaudDib fucked around with this message at 19:43 on Mar 17, 2023 |
|
#
?
Mar 17, 2023 04:09
|
|
|
Openbenchmark has tests of ffmeg performance which should be directly readable to handbrake. https://openbenchmarking.org/test/pts/ffmpeg 13900k seems to be king of x264, while AMD overtake it for x265. I will confess to not understanding what the different scenarios are actually testing.
|
|
#
?
Mar 17, 2023 07:36
|
|
|
Pablo Bluth posted:Openbenchmark has tests of ffmeg performance which should be directly readable to handbrake. different compression algorithms (DivX vs HEVC basically). HEVC is much more intensive to compute but achieves higher compression especially at lower bitrates (although supposedly the quality ceiling is lower at very high bitrates). HEVC really benefits from AVX-512 which Intel completely hosed up, where x264 either doesn't or doesn't have enough vector intensity to really get good speedups out of it (it also doesn't benefit as much from higher threadcounts etc - x264 is just easier essentially). Since HEVC/H265 is very very patent-encumbered, nobody adopted it for streaming, but the  movie community has adopted it (and I think bluray is HEVC normally too). It's good for anime in particular, very very good compression ratios and it doesn't suffer from the loss of very fine detail that can sometimes occur with HEVC.
|
|
#
?
Mar 17, 2023 07:51
|
|
|
Paul MaudDib posted:Since HEVC/H265 is very very patent-encumbered, nobody adopted it for streaming, but the Microsoft charges users a dollar for it: https://apps.microsoft.com/store/detail/hevc-video-extensions/9NMZLZ57R3T7
|
|
#
?
Mar 17, 2023 08:07
|
|
|
I understand x264/x265, its the live/on demand/platform/upload/popular test scenarios that aren't explained by openbenchmark. Managed to find them explained here: http://arcade.cs.columbia.edu/vbench-asplos18.pdf
|
|
#
?
Mar 17, 2023 08:42
|
|
|
hobbesmaster posted:It’s “only” 1.05W/mm^2. The minimum for a welder in about 10W/mm^2  
|
|
#
?
Mar 17, 2023 11:03
|
|
|
Broke: switching from gas stoves to induction stoves Woke: using your Core 2 Duo as a hot plate
|
|
#
?
Mar 17, 2023 11:18
|
|
|
Encrypted posted:it's still 105W/cm^2 though still lower power density than a stock 12900k at full load in prime95 lol (250W at 2.15 cm^2 die size)
|
|
#
?
Mar 17, 2023 11:41
|
|
|
hobbesmaster posted:And in the UK you’d be able to blow all the fuses in your new house because they still build ring circuits in the 21st century. That's why there's a fuse in the plug, but yeah, rings can be dodgy if there's a break in them, then you just have a pair of undersized radials. But it did save copper and let people put resistive heaters all over their draughty houses HalloKitty fucked around with this message at 11:49 on Mar 17, 2023 |
|
#
?
Mar 17, 2023 11:45
|
|
|
Twerk from Home posted:Handbrake isn't an editing tool, more of a compression tool. Yeah I was confused at that too but that's what I was told. Thanks for all the info, guys!
|
|
#
?
Mar 17, 2023 12:31
|
|
|
Pablo Bluth posted:Openbenchmark has tests of ffmeg performance which should be directly readable to handbrake. That's what I said about benchmarking one encode at a time, but that isn't how someone with lots of encodes to do would actually run the tool. If you want the highest encoding throughput, you're going to run 1 job per core with encoder settings to use only 1 thread each. You get much better total throughout that way. Heck, you actually get marginally better compression in the end too. Using multiple threads to encode not only scales less than linearly for performance, it makes the compression worse because each frame / block doesn't have all possible data available from previously frames. Hell, assuming that you're not RAM or I/O limited my experience has been that hyper threading can squeeze out another 5-10% so you can actually do between 1-2 jobs per core. I wouldnt go all the way to 2 jobs per core unless the system was non-GUI and doing literally nothing else though.
|
|
#
?
Mar 17, 2023 14:59
|
|
|
Paul MaudDib posted:different compression algorithms (DivX vs HEVC basically). HEVC is much more intensive to compute but achieves higher compression especially at lower bitrates (although supposedly the quality ceiling is lower at very high bitrates). HEVC really benefits from AVX-512 which Intel completely hosed up, where x264 either doesn't or doesn't have enough vector intensity to really get good speedups out of it (it also doesn't benefit as much from higher threadcounts etc - x264 is just easier essentially). Intel expects you to buy a Xeon Scalable to do h265 - something they've been quite honest about for at least 5 years. Pablo Bluth posted:I understand x264/x265, its the live/on demand/platform/upload/popular test scenarios that aren't explained by openbenchmark.
|
|
#
?
Mar 17, 2023 15:29
|
|
|
BlankSystemDaemon posted:Nah, it's much more simple than that. gently caress that, my money is on the 13900k outperforming Xeons at twice or 3x the price. AVX-512 can't make up for the gigantic clock advantage and also 8+16 cores is a lot of cores in the first place. Edit: Related to that, Hetzner is offering server rentals with 13900 or 13500 processors and ECC RAM: https://www.hetzner.com/press-release/neue-dedicated-server-2023/ This is going to gently caress up Intel's market planning badly. Benchmarks and my own initial testing shows that a single 13900 outperforms two $1200 Xeon 4316s on a lot of stuff. Twerk from Home fucked around with this message at 15:49 on Mar 17, 2023 |
|
#
?
Mar 17, 2023 15:46
|
|
|
Paul MaudDib posted:golden cove is something like 3x the transistors of zen3. the reason AMD isn't doing a full-on e-core is that they don't need it - their p-cores are already relatively compact and space-efficient, they just are going to squeeze the last 10-20% by using high-density libraries and microcoding AVX-512 support and call it a day. Zen5 will probably be bigger but Zen4c is perfectly adequate as a space-optimized e-core, it's something like 1.5-2x the transistors of Gracemont but it also has SMT which Gracemont doesn't. That's fine imo. The Cove core seem like a dead-end to me too. It's gotten too big and power hungry and it doesn't look like anything fundamental could be done to it (though I'm no expert so). If Crestmont can repeat the Tremont->Gracemont improvements, it would be within like 30% of today's top Cove performance at much lower cost and power (in most scenarios, on average, etc). Still not enough to compete on the top end but would make for a much more sensible product for 90% of users. There are more reviews of the Gracemont CPUs starting to come out that you can kind of get a picture of how the models compare, here's the N95 (with N100 reference from past review): https://www.youtube.com/watch?v=g7HWTwAyGNY I'm pretty confused by the power situation. My N5100 has a TDP of 6W and will not go over like 3.5-4W package in any single core workload (1.5w of it is base idle power). Yet here he's seeing a difference between 20W and 30W PL1!?    And somehow changing the power limit, despite boosting the performance, doesn't actually impact the draw at the wall very much so it ends up matching the N100 in CB while using 10 fewer watts  Might be just sketchy measurements though.
|
|
#
?
Mar 17, 2023 16:00
|
|
|
Thunder Bay x86 + Movidius VPU SOC, which turned out to (theoretically) be Arm A53 + VPU, now cancelled before production. https://www.phoronix.com/news/Intel-Thunder-Bay-Cancelled Just mentioning this here because it's so weird to swerve from "we'll make an integrated x86 computer vision solution", to (one supposes) "oops turns out that Arm works better", to nothing.
|
|
#
?
Mar 17, 2023 17:25
|
|
|
Llamadeus posted:It's widespread in streaming, just not in a browser. It's still the de facto standard for 4K and HDR on Netflix, Amazon, etc. but not for Twitch or YouTube thanks to the prohibitively expensive royalties Chrome/Firefox/etc would have to pay per browser download. Huh. That's bullshit, Microsoft used to supply them for free. The store page was titled "Get HEVC Video Extensions From Device Manufacturer", I just checked the old bookmark I had, and it's been deprecated.
|
|
#
?
Mar 17, 2023 17:47
|
|
|
mdxi posted:Thunder Bay x86 + Movidius VPU SOC, which turned out to (theoretically) be Arm A53 + VPU, now cancelled before production. VPU has to be one of the most useless acronyms in computing right now. It’s impossible to know what a “VPU” refers to even if you have the context of a vendor. For what was last publicly known about Thunder Bay: quote:This week brought initial patches for bringing up the Thunder Bay SoC and initial board support under Linux. The patches confirm this new Intel Movidius SoC uses Cortex-A53 CPU cores with the Movidius VPU and that so far are actually two variants to this new SoC. The Thunder Bay "full" configuration has four clusters of four A53 cores per cluster and four VPUs. The Thunder Bay "Prime" configuration has four clusters of four A53 cores but only two VPUs along with less memory. The memory support comes down to 8GB + 8GB + 4GB + 4GB for the Thunder Bay "full" and 8GB + 4GB for the "prime" configuration. Off the shelf A53clusters + gpu + some other vector/video/other unit are made by… well, everyone.
|
|
#
?
Mar 17, 2023 17:47
|
|
|
hobbesmaster posted:VPU has to be one of the most useless acronyms in computing right now. It’s impossible to know what a “VPU” refers to even if you have the context of a vendor. The P stands for Pee.
|
|
#
?
Mar 17, 2023 18:39
|
|
|
hobbesmaster posted:Off the shelf A53clusters + gpu + some other vector/video/other unit are made by… well, everyone. Yeah, reading through what was known about it, sorta seems like the niche they were targeting got filled, including by Intel themselves with the built in "AI" accelerator blocks in CPUs and their dGPUs. Keem Bay was 2 years late to market and is still unobtanium, and looking at the specs, it was pushing 7.1 TOPS in ResNet-50. Thats like... a Jetson Nano? I'm no expert in this field though so im just talkin out my rear end.
|
|
#
?
Mar 17, 2023 18:58
|
|
|
It feels like there’s maybe one or two hyperscalers buying it for one or two apps, but the relevant swengs got their promo to principal for rolling it out and the next gen perf differentiation wasn’t there anymore
|
|
#
?
Mar 17, 2023 23:58
|
|
|
itshappening.gif https://twitter.com/PGelsinger/status/1638234206613667842
|
|
#
?
Mar 21, 2023 18:47
|
|
|
Clearly it's time for Raja and Phil Harrison to jointly found a startup that will innovate failure in the graphics and gaming spaces. Nobody else has their experience in loving up the things they're supposed to be experts at!
|
|
#
?
Mar 21, 2023 19:30
|
|
|
|
| # ? May 13, 2024 05:33 |
|
|
He’s supposedly going to do some gaming AI thing, which I can only assume means he’s gonna train an AI to automatically produce even more or those insane “caught my husband cheating” mobile games. https://twitter.com/rajaxg/status/1638236242537250816?s=46&t=cj0103ZhjrzDsJf6X4zTug
|
|
#
?
Mar 21, 2023 19:37
|
|