|
Kazinsal posted:e: I guess more than anything it seems to me that everone's implementation of firmware-ish-level H.264 encoding on GPU dies has been just abysmal and that's the actual problem here. I think he means hardware encoders. They're proprietary implementations that come with fixed presets and bitrates. Nvidia Pascal architecture has acceptable quality and Turing improved upon that to the point where it is quite useable. I'm not up to date on AMD quality but it was pretty bad with Polaris. I don't use Plex but if it supports NVENC then I'd build a system with the cheapest Turing card with full NVENC support (GTX 1660? — not sure) unless you want to throw a lot of cores at it for software encoding.
|
#
?
Nov 3, 2019 15:09
|
|
|
|
| # ? May 30, 2024 10:25 |
|
|
D. Ebdrup posted:Yep, QuickSync uses special chips that're included in the iGPU of the Intel chips, so while it doesn't use GPGPU execution units like nvenc, it does need the iGPU to be on the die. Makes sense! It's not that the extra $100 is a deal breaker - but honestly it seems like most of the latest generation seems to be moving more towards no integrated graphics because of how standard having dedicated GPUs are these days. DrDork posted:Yup, and GPGPU encoding also tends to look like poo poo in general. Video encoding is one of the areas where you still really want to be doing everything on the CPU. I am much too lazy and low effort to search and buy used off of something like eBay. eames posted:I think he means hardware encoders. They're proprietary implementations that come with fixed presets and bitrates. That's correct. The GTX 1050 TI that I have does fully support the NVENC codec. It's an extra card I have laying around (bought it as a backup when my previous died and I had to wait on the RMA) so if I can reuse it, awesome. I can go ask this in the GPU thread too. Doh004 fucked around with this message at 15:29 on Nov 3, 2019 |
|
#
?
Nov 3, 2019 15:23
|
|
|
eames posted:I think he means hardware encoders. They're proprietary implementations that come with fixed presets and bitrates. Plex supports NVENC and I’ve had a really positive experience with it. Just set the bitrate modestly higher than you would do for a CPU encode and enjoy.
|
|
#
?
Nov 3, 2019 15:24
|
|
|
The ASICs that accelerate encoding like those in quicksync and nvenc make trade offs of precision, complexity and features in exchange for speed and power efficiency, pure software encoding can be set to make no quality/precision/complexity or supported feature trade offs but will incur a huge performance penalty in exchange. If the differences in implementation actually translate to something detectable by the average human in a high bit rate archival encode is debatable. It is more likely that pure software encoding can achieve the same quality output with fewer bits instead of its output being universally better than hardware encoding if you just throw as many bits as it needs to look good.
|
|
#
?
Nov 3, 2019 15:40
|
|
|
Doh004 posted:I am much too lazy and low effort to search and buy used off of something like eBay. It's your money/choice and all, but you can pick up E3-1270v3 or similar for $75 all day every day. How many streams and what else is this server going to be doing, anyhow? If it's just one or two non-4k streams at a time, even a simple i3-9100F would be easily able to handle the load.
|
|
#
?
Nov 3, 2019 15:47
|
|
|
DrDork posted:It's your money/choice and all, but you can pick up E3-1270v3 or similar for $75 all day every day. At most, probably 4 1080p streams at once if all of my family members are watching something at once from their homes while we're watching something in our living room. This isn't for a huge group of people or anything.
|
|
#
?
Nov 3, 2019 15:52
|
|
|
You'll probably be fine, then. I'm assuming that at least two of those streams would be on your internal network, and thus not need transcoding at all--simply serving up a stream without processing is trivial for modern CPUs.
|
|
#
?
Nov 3, 2019 16:03
|
|
|
DrDork posted:It's your money/choice and all, but you can pick up E3-1270v3 or similar for $75 all day every day. Yeah I just checked and two years ago I paid 300 shipped for a used HP box with an e5-2650v2 and 32gb of ecc ram. All I did was plug a bunch of 8tb wd reds and an unraid USB stick in. It's worth mentioning that I am far too stupid to deal with Linux for more than ten minutes and I have had zero problems with my unraid box.
|
|
#
?
Nov 3, 2019 16:50
|
|
|
Silicon Lottery published their 9900Ks binning stats 100% 5.0 GHz 31% 5.1 GHz 3% 5.2 GHz
|
|
#
?
Nov 3, 2019 23:19
|
|
|
I worked for a stint in digital video broadcast, distribution, and encoding. Nobody that’s really competitive at commercial scale does GPU based encoders of any sort. Netflix, for example, is running map reduce jobs on chunks of video at a time and encodes at like 200x real-time as a result. For some others that put out content that airs simultaneously with live broadcast like most network and premium TV shows, there’s a ton of industry standards (or more correctly, lack thereof) that keep some of these systems from adopting stuff live (HBO Now doesn’t support Real Time with Bill Maher for example). The typical point of stuff like NVEnc and QuickSync is for video teleconferencing on low end laptops to keep them from sucking up battery and where quality doesn’t matter terribly much. NVEnc silicon hasn’t changed that much for like 6 years and commercial vendors are not supporting it anymore due to how little nVidia has kept it maintained and because video encoding standards are so much more advanced these days.
|
|
#
?
Nov 3, 2019 23:57
|
|
|
eames posted:Silicon Lottery published their 9900Ks binning stats One thing about their binning statistics for the 9900k that strikes me as amusing is their 4.8 GHz all core model has an AVX frequency of 4.6 GHz. Stock performance of a 9900k is 4.7 GHz all core with no AVX offset, literally 100% of the samples should be able to do 4.7 GHz all core AVX.
|
|
#
?
Nov 4, 2019 00:02
|
|
|
eames posted:Silicon Lottery published their 9900Ks binning stats However they are doing it at way better voltages. eg 5.1 GHz on the 9900K is 1.312V, on the 9900KS it's 1.287V. So 0.025V lower for the same clocks. That means even easier cooling - not that 1.312V is a big problem after delid, but these will run even cooler. Also, if you don't pussyfoot around with the voltages like SL is doing, then you can undoubtedly ramp those clocks up even more. I would guess that 100% of those chips will do 5.1 at 1.312V or slightly above. I would guess the majority will do 5.2 at 1.37 or so. The golden 3% there will probably do 5.3 and stay at 1.4v or under.
|
|
#
?
Nov 4, 2019 03:42
|
|
|
I'm not in a position to argue against that as I’ve never owned a SL binned CPU but they certainly are not known to be conservative on voltages. They test voltages right at the edge of what their QVL cooling systems are capable of. Something like a NH-D15 doesn't even make it to that list. All of their advertised overclocks are also done with unrealistically slow RAM (max. DDR4-2800 for 4 DIMMs, max. DDR4-3200 for 2 DIMMS) and faster RAM does effect core overclocking results negatively in my experience, not just due to higher heat dissipation from VCCIO/VCCSA. I'm sure there are people out there that think that $1200 for a guaranteed 5.2 GHz 9900K with -2 AVX Offset is a good deal. I'm not one of them but I wouldn't be surprised if they sold a bunch of 5.3/5.4 GHz examples privately to HFT people for much, much more than that.
|
|
#
?
Nov 4, 2019 09:14
|
|
|
eames posted:I'm not in a position to argue against that as I’ve never owned a SL binned CPU but they certainly are not known to be conservative on voltages. They test voltages right at the edge of what their QVL cooling systems are capable of. Something like a NH-D15 doesn't even make it to that list. Well, their cooling systems must have gotten shittier since last week, because they were willing to test up to 1.312V on the last batch of chips and now they're only willing to test 1.287V, on the same product. edit: the more salient point there (from another poster, not from SiliconLottery themselves) is that they test at a -2 AVX offset and yeah, AVX is going to be tougher and tougher to push at 5+ GHz. But they don't have to keep one AVX offset across the whole lineup... they list AVX and non-AVX bins separately on their historical binning statistics... they could leave AVX at 4.9 and keep stepping up the non-AVX clocks where possible (i.e. increasing AVX offset on the top bins). But SiliconLottery's claim that they need lower voltages than the K series because of thermal limitations doesn't pass the sniff test. Paul MaudDib fucked around with this message at 18:45 on Nov 4, 2019 |
|
#
?
Nov 4, 2019 18:31
|
|
|
Their 5.2ghz are at 1.325, what are you on about? If they could hit 5.3 and have it be sufficiently stable they could charge like $600 more per processor than they are for 5.2, I don't think they aren't trying.
|
|
#
?
Nov 4, 2019 18:45
|
|
|
what is the technology that Intel uses for inter-core communication? I've absorbed a bunch of information about Ryzen's Infinity Fabric, and I'd like to know about how Intel does it, or is the design so different that there's no analogue?
|
|
#
?
Nov 5, 2019 06:03
|
|
|
gradenko_2000 posted:what is the technology that Intel uses for inter-core communication? I've absorbed a bunch of information about Ryzen's Infinity Fabric, and I'd like to know about how Intel does it, or is the design so different that there's no analogue? "Ring bus" https://www.anandtech.com/show/3922/intels-sandy-bridge-architecture-exposed/4 https://pcper.com/2017/06/the-intel-core-i9-7900x-10-core-skylake-x-processor-review/3/ (at least for the consumer chips, workstation/server chips have moved to a 2D mesh routing. https://www.anandtech.com/show/13124/the-amd-threadripper-2990wx-and-2950x-review/4 As core counts increased, Broadwell server chips used a pretty crazy dual-ringbus arrangement that Intel abandoned with the Skylake-X/Skylake-SP generation. https://www.tomshardware.com/reviews/intel-xeon-e5-2600-v4-broadwell-ep,4514-2.html Basically ringbus is very fast and very power efficient, for limited core counts, but obviously the latency increases with higher core counts (as the number of hops between "farthest" nodes increases), which is why Intel moved off it. Zen is actually faster within a CCX, but slower between CCXs. On top of that there are also differences in cache topology - Intel is a Single Big Inclusive Cache which favors gaming as well as some per-thread-performance-sensitive applications (single-threaded isn't quite the right concept), while AMD is actually a separate cache per CCX (although they might be moving away from this next generation), and a victim cache (only stuff that isn't actively used gets to L3 and can be snooped easily). Also if you are interested in this stuff you should read Agner Fog's "Microarchitecture", everything you ever wanted to know about low-level CPU design. https://www.agner.org/optimize/microarchitecture.pdf Paul MaudDib fucked around with this message at 22:53 on Nov 5, 2019 |
|
#
?
Nov 5, 2019 06:30
|
|
|
Here's a video about the arcane magic that is EUV lithography https://www.youtube.com/watch?v=f0gMdGrVteI
|
|
#
?
Nov 5, 2019 22:45
|
|
|
2 years ago:mobby_6kl posted:Is ARM on servers the new Linux on desktop thing I keep hearing every year? 1 year ago: Rastor posted:Hah, apparently I had set a reminder to myself for one year after this post. FunOne posted:Probably. I think ARM on the Desktop is coming soon(tm). The eternal rumor is that the progress Apple is making on their A00x chips will translate into real laptops at some point in the near future. Well, it's another year gone by. Chromebook sales are booming. quote:According to the latest report from The NPD Group that tracks retail sales in the U.S., consumers are buying into the Chrome OS ecosystem at an accelerated pace. In fact, overall sales of non-Chrome OS notebooks were down 6.1% while Chromebooks enjoyed a 22% YoY growth spurt. ARM on the Desktop:... still hasn't really come, a year later. At least, whatever Apple's doing is still only rumors. Microsoft made a bit of a splash with their SQ1 processor for the Surface Pro X, but it is reviewing poorly. And ARM on the server... yeah I think Intel has successfully staved off any competition there. AMD's x86 options are the competition there. Should I set the alarm again... ?
|
|
#
?
Nov 8, 2019 16:09
|
|
|
Yes.
|
|
#
?
Nov 8, 2019 16:14
|
|
|
Geekbench score of a Dell AIO with a 10C/20T Comet Lake CPU leaked https://browser.geekbench.com/v4/cpu/14904070 The score is useless because of clockspeed and RAM but Q1/2020 sounds about right.
|
|
#
?
Nov 8, 2019 19:27
|
|
|
https://blogs.intel.com/technology/2019/11/ipas-november-2019-intel-platform-update-ipu/quote:Intel is heavily invested in both industry collaboration and in conducting security research into our own products. As a result, while we are addressing 77 vulnerabilities this month, 67 were discovered internally through our own testing, validation and analysis. We believe that assigning CVE ID’s and publicly documenting internally found vulnerabilities helps our customers to accurately assess risk, prioritize, and deploy updates. By the time you are reading this blog post, mitigations for many of these issues will have already been propagated throughout the ecosystem through the IPU process. At the same time, the external researchers who reported the remaining issues to us have all been good partners in working with us on coordinated vulnerability disclosure (CVD). Intel plz. I remember when six or more CVEs in A YEAR was unusual, but 77 in a month? SwissArmyDruid fucked around with this message at 03:25 on Nov 13, 2019 |
|
#
?
Nov 13, 2019 03:22
|
|
|
By my read it’s just thinking harder about errata most wouldn’t look at anyway? A lot of errata have a fantastically small audience, BIOS/(efi whatever) authors before the ucode patch is loaded. It still needs to be documented for them but it’s not a cloud attack vector.
|
|
#
?
Nov 13, 2019 06:21
|
|
|
The release also includes a fix for a new erratum. The fix will reduce performance by 0-4% depending on the workload. Source: https://www.intel.com/content/dam/support/us/en/documents/processors/mitigations-jump-conditional-code-erratum.pdf My good old 8700K still has 20% higher single thread performance than the leaked 3950X if Geekbench is any indication but it doesn't look like that'll last at this rate.  e: phoronix provides benchmarks and games are also affected. It's not horrible (though hosting companies using PHP may disagree) but the performance losses add up if one of these gets released every year. eames fucked around with this message at 13:00 on Nov 13, 2019 |
|
#
?
Nov 13, 2019 08:30
|
|
|
SwissArmyDruid posted:https://blogs.intel.com/technology/2019/11/ipas-november-2019-intel-platform-update-ipu/ There's also http://tpm.fail/ which we've known about for almost a month.
|
|
#
?
Nov 13, 2019 11:21
|
|
|
TAA/RIDL, too.
|
|
#
?
Nov 13, 2019 13:40
|
|
|
So is this critical affecting Server/Cloud environments or "my gaming PC" as well? What would the scenario be?
|
|
#
?
Nov 13, 2019 13:50
|
|
|
Mr.PayDay posted:So is this critical affecting Server/Cloud environments or "my gaming PC" as well? What would the scenario be? From what I understand Zombieload/RIDL are mostly about data leaks via untrusted code. In *theory* your CPU could leak the content of your unlocked password manager window while you visit a site with a malicious javascript ad. All you can do is keep your computer up to date. OS, BIOS, Browser. Adblocker won't hurt either. The multiple research teams that found these exploits suspect that there are many similar variants of this exploit that will be discovered over the years. That's why they strongly suggest disabling Hyperthreading which makes these attacks significantly faster and easier. Obviously that's a major performance/efficiency loss depending on the workload. Intel is trying to fix this on a more granular level (new steppings, microcode, software updates) but they can only fix what they find or what gets disclosed to them. I suspect Intel would've phased out HT for consumers with the 10th gen if it wasn't for AMD. e: there's also a bug in the managment engine that allows access to the machine over the (local?) network. That was fixed via a firmware update which typically comes bundled with BIOS updates. eames fucked around with this message at 16:01 on Nov 13, 2019 |
|
#
?
Nov 13, 2019 15:58
|
|
|
eames posted:In *theory* your CPU could leak the content of your unlocked password manager window while you visit a site with a malicious javascript ad.
|
|
#
?
Nov 13, 2019 18:29
|
|
|
Or you can just disable javascript and make browsing the internet infinitely quicker and easier
|
|
#
?
Nov 13, 2019 19:15
|
|
|
Disable html for maximum ease of use.
|
|
#
?
Nov 13, 2019 20:13
|
|
|
Intel outside for shitposting in Something Awful then?
|
|
#
?
Nov 13, 2019 20:17
|
|
|
https://mdsattacks.com
|
|
#
?
Nov 13, 2019 20:17
|
|
|
https://zombieloadattack.com is another one
|
|
#
?
Nov 13, 2019 21:14
|
|
|
Recently got a notification from our Dell reps that some orders might be delayed in the coming months due to shipping delays from Intel. eames posted:That's why they strongly suggest disabling Hyperthreading which makes these attacks significantly faster and easier. Obviously that's a major performance/efficiency loss depending on the workload. Heh, I don't even have hyperthreading on my Haswell i5. 
|
|
#
?
Nov 14, 2019 01:48
|
|
|
Somebody on a different forum pointed out that Intel released new microcode for security mitigations yesterday and the AMD 3950X embargo got lifted today. Technically they released the fixes before this round of benchmarks got published but practically nobody took the time to update their 9900K(S) machines to redo the benchmarks. 
|
|
#
?
Nov 14, 2019 17:14
|
|
|
eames posted:Somebody on a different forum pointed out that Intel released new microcode for security mitigations yesterday and the AMD 3950X embargo got lifted today. The 3950X pretty much destroys any Intel offerings at price/performance.
|
|
#
?
Nov 14, 2019 17:27
|
|
|
....I think Intel was better off not shooting themselves in the foot when they were still ignoring AMD. This performance anxiety they've suddenly contracted a case of is gonna be ugly for the next few years.
|
|
#
?
Nov 14, 2019 17:53
|
|
|
ratbert90 posted:The 3950X pretty much destroys any Intel offerings at price/performance. Intel sweeps gaming still, which is the point of the 9900KS and the primary reason for sales in DIY PCs, so I would be interested to see if the microcode hit performance at all.
|
|
#
?
Nov 14, 2019 20:16
|
|
|
|
| # ? May 30, 2024 10:25 |
|
|
eames posted:The release also includes a fix for a new erratum. The fix will reduce performance by 0-4% depending on the workload. 0-4%... Or 20%... https://twitter.com/damageboy/status/1194751045454454784
|
|
#
?
Nov 14, 2019 20:31
|
|