|
Genoa looks to still be 8 core chiplets based on the wireframe, and not 3d compute stacked. Which meshes with the rumors. The Zen 4c cores in Bergamo are certainly interesting. They mentioned multiple times that they would support the same instructions as the "big" cores in Genoa, which means its gotta be a cache cut because there is... basically nothing else on those dies, lol. So I guess AMD will give you a spectrum of parts with either monster caches or monster compute depending on your workload. Pretty interesting. All the rumors point to both Zen4 and Sapphire Rapids being a massive leap, so this gonna get wild.
|
#
?
Nov 8, 2021 19:51
|
|
|
|
| # ? May 17, 2024 05:05 |
|
|
Pablo Bluth posted:I thought the idea behind this type of chip was to trade peak IPC/freq and the power that drinks, for a design that uses the power efficiency to pack in a large eough number of cores to increase the total throughput of the chip? Yeah, you're right. The article says the same thing but in two ways at different points, and I misinterpreted what they meant when I got to the first instance. The recap at the end of the article makes it clear that the tradeoff is clocks in favor of power efficiency.
|
|
#
?
Nov 8, 2021 19:59
|
|
|
Zen4c is basically "little" cores, but Zen4 is a "big" core but still somewhat smaller/lower performing than Zen3. (edit: AMD is now confirming it's Zen4c, the c is for "cloud optimized") It's the same math that Intel is doing with Alder Lake - it's a tradeoff of how much you chase single-thread performance vs how many of them you can fit onto a die. If it's 10% slower but you can fit 50% more of them onto a die (33% smaller) that's a win for MT. There also is the consideration of efficiency. Recently, CPUs have heavily favored "relaxed pitch" approaches, where you trade a little bit of density for higher clocks. Obviously there are tradeoffs here - not just power, but also if you make them too big then you start having design problems where signals don't propagate across the core quickly enough, like clock jitter and so on, and you end up having to lengthen the pipeline to account for that, etc. Anyway, what Apple has showed is that maybe designing for 3-4 GHz isn't the best idea and they should be designing for 2-3 GHz instead - go with high-density libraries, keep them small and efficient, but then spend the same or greater area in total to keep the total performance the same and achieve much higher IPC. It didn't really break into the public consciousness until A15 that "oh poo poo apple's big cores are better than a 5900X and pull 3W per core peak" but Apple has been on this path for years and I'm sure AMD took note much earlier than the general public. The gotcha is going to be whether x86 really lets you achieve those kinds of IPC given what a giant mess the instruction set is - Apple has freakishly deep reorder buffers and the enabling factor there is not just the decoder having fixed with instructions but also the very high orthogonality of the instruction set. ARMv8 is designed from the ground up, by Apple, for Apple (David Kanter was corrected on this in an interview with an Apple engineer - Kanter tried to pull the "you mean ARMv8 was designed in collaboration with ARM", the engineer corrected him - ARMv8 is literally Apple's spec, Apple told ARM how they wanted it and ARM published it) to enable those super deep reorder buffers, and you can't necessarily do that on x86 to achieve the super-high single-threaded IPCs. They did that for a reason, if the existing ISAs had been adequate they wouldn't have. Oh and ARMv8 also has twice the number of registers which helps a ton with register renaming and reordering - can't do that without another x86 spec revision like x86-64. I think the ST battle cannot be won - Apple is pushing at least triple the IPC of x86 designs on a per-thread basis. Yes, there's workarounds like instruction cache for the downsides of x86's decoder/etc - but those workarounds largely have already been tried and I strongly question whether there's enough juice left there to get to triple the IPC. Surely if that were possible, it would have been advantageous to do it - Zen already is server-first and it would have been hugely advantageous there. But what you can do is SMT (perhaps even more than 2 threads per core - SMT4 or SMT8), and you can do little cores as a workaround to increase the MT areal efficiency of your design. Both of those approaches would let them fight Apple on servers and other purely-MT workloads even if they can't compete with single-threaded IPC - and I think that's most likely what Jim Keller meant with his "x86 isn't dead" speech. I really don't see how they can possibly win the single-threaded battle. But I guess we'll see. Paul MaudDib fucked around with this message at 21:15 on Nov 8, 2021 |
|
#
?
Nov 8, 2021 20:33
|
|
|
For all the discussion around X64 baggage, all the experts who ought to know (like Jim Keller) always seem to come back with 'it's not a problem'.
|
|
#
?
Nov 8, 2021 21:06
|
|
|
mdxi posted:Next year EPYC Genoa arrives on TSMC 5nm, with 96 Zen4 cores, DDR5, PCIe5, and 1.25X performance uplift Seems a bit underwhelming considering Zen4 benefits from both architectural improvements and a better node. Just enlarging the L3 cache of Zen3 is supposed to increase performance by 15%. Did AMD give any sort of indication when in 2022 Zen4 would be released?
|
|
#
?
Nov 9, 2021 00:21
|
|
|
The whole big-little-adjacent thing isn't rumored to happen until zen 5, where the 4c cores are the little cores to zen 5's big cores. So far the other components of the rumors have born out (with some slightly altered names), so we'll see.
|
|
#
?
Nov 9, 2021 00:29
|
|
|
ConanTheLibrarian posted:Seems a bit underwhelming considering Zen4 benefits from both architectural improvements and a better node. Just enlarging the L3 cache of Zen3 is supposed to increase performance by 15%. Anandtech posted:and >1.25x silicon performance over the regular N7 it uses. When asked if this was a specific statement about core performance, AMD said that it wasn’t, and just a comment on the process node technologies.
|
|
#
?
Nov 9, 2021 00:35
|
|
|
If there's a flavour of Zen4 runs a little less hot and I can slap it in my current X570 I'd be tempted to swap out my 5900X to be honest.
|
|
#
?
Nov 9, 2021 00:43
|
|
|
I'm hoping those multichip GPUs start making their way to desktop at some point.
|
|
#
?
Nov 9, 2021 00:43
|
|
|
ijyt posted:If there's a flavour of Zen4 runs a little less hot and I can slap it in my current X570 I'd be tempted to swap out my 5900X to be honest. Zen 4 supposedly will be AM5 only, sry
|
|
#
?
Nov 9, 2021 00:50
|
|
|
Cygni posted:Zen 4 supposedly will be AM5 only, sry I cry
|
|
#
?
Nov 9, 2021 00:50
|
|
|
Dry your tears, you're getting an LGA socket!
|
|
#
?
Nov 9, 2021 02:32
|
|
|
come on AMD bump up a Ryzen 5 to be an 8-core chip
|
|
#
?
Nov 9, 2021 03:40
|
|
|
ijyt posted:I cry There is always zen3d which is rumored to be 15% faster
|
|
#
?
Nov 9, 2021 06:00
|
|
|
Zedsdeadbaby posted:There is always zen3d which is rumored to be 15% faster That's just the number AMD said themselves, and we should know by now not to trust official benchmarks/performance comparisons.
|
|
#
?
Nov 9, 2021 08:08
|
|
|
Pablo Bluth posted:I'd file that under 'vague statement that doesn't tell us much about the final answer that matters'. Anandtech had a bit more detail - the slide the 1.25x figure was from referred to the 4c cores specifically (good), and AMD said 1.25x didn't mean actual perf (  ).
|
|
#
?
Nov 9, 2021 09:20
|
|
|
Dr. Video Games 0031 posted:That's just the number AMD said themselves, and we should know by now not to trust official benchmarks/performance comparisons. Hence rumor
|
|
#
?
Nov 9, 2021 12:47
|
|
|
Good overview from servethehome https://www.youtube.com/watch?v=ZEDKNtt-erk Those Computational fluid dynamics speedups from the 3d cache look sweet.
|
|
#
?
Nov 9, 2021 20:19
|
|
|
Strangely enough, the fact that Milan-X is available right now in Azure has almost flown under the radar. For memory-bound applications, the benefits are massive.
|
|
#
?
Nov 9, 2021 20:53
|
|
|
Number19 posted:I'm really curious to see what this means for the Zen4 Ryzen lineups, since a 96 core Genoa pretty much confirms a new 12 core CCX. There's literally a schematic of a 12 die CPU being shown in the presentation. 12x8 = 96? See top of this page. What am I missing? --edit: I can't count. Combat Pretzel fucked around with this message at 22:52 on Nov 9, 2021 |
|
#
?
Nov 9, 2021 22:50
|
|
|
Combat Pretzel posted:Uh, what? I mean what the core counts for Zen4 based Ryzens will look like now that there's going to be a 12 core CCX
|
|
#
?
Nov 9, 2021 23:56
|
|
|
I got that much. If that thing shown off as art in the presentation is supposed to be Genoa, it's made from twelve 8-core CCDs, tho.
|
|
#
?
Nov 10, 2021 00:03
|
|
|
Oh duh. Why did I see it as a 12 core CXX? Yeah in that case I doubt anything changes at all. Too bad, they really could have a stuck a knife into Intel by upping the core counts. Maybe not though: multi core compute is doing pretty well these days. We need more single thread and power efficiency gains again.
|
|
#
?
Nov 10, 2021 01:59
|
|
|
What in the gently caress is this poo poo? I guess kiss easy availability of the V-Cache Ryzens come Q1/22 goodbye? I wish this cryptobullshit would finally loving die.quote:According to a report in the UK’s Bitcoin Press, part of the reason it’s so hard to find a current-gen AMD CPU for sale anywhere is because of a crypto currency named Raptoreum that uses the CPU to mine instead of an ASIC or a GPU. Apparently, its mining is sped up significantly by the large L3 cache embedded in CPUs such as AMD Ryzen, Epyc, and Threadripper.
|
|
#
?
Nov 12, 2021 20:26
|
|
|
Color me doubtful that a budding shitcoin with a $31m market cap is having any meaningful impact on availability of CPUs. Ethereum, which does definitely have at least some impact on GPU availability, has a market cap of $550,000m.
|
|
#
?
Nov 12, 2021 20:33
|
|
|
yeah that article is going for crypto outrage clicks
|
|
#
?
Nov 12, 2021 20:41
|
|
|
I have no idea. Whenever someone starts talking about cryptocoins, it sounds like the teacher from the Peanuts is talking to me. I don't even get why there's a million and one different ones to begin with.
|
|
#
?
Nov 12, 2021 21:07
|
|
|
Combat Pretzel posted:I have no idea. Whenever someone starts talking about cryptocoins, it sounds like the teacher from the Peanuts is talking to me. I don't even get why there's a million and one different ones to begin with. because if you start your own coin, you can give yourself 10% “premine” (of the total coins to ever exist) for your essential role as the “ideas guy” who copied and pasted some other coin and changed the name, and then a year later you can cash out a couple hundred grand and retire to Belize
|
|
#
?
Nov 12, 2021 21:10
|
|
|
Paul MaudDib posted:retire to Belize Bit drastic
|
|
#
?
Nov 12, 2021 21:32
|
|
|
technically it's puerto rico because of the tax breaks for cryptodorks
|
|
#
?
Nov 12, 2021 21:34
|
|
|
Why is the B550 usually recommended for a 5600X over an X570, when for Intel CPUs everybody (i.e., goons) say to get the higher end board (Z570 over the B560)? (both naming conventions are weird to me) Is it because you need more power to get the full potential of an Intel chip/higher TDPs?
|
|
#
?
Nov 14, 2021 04:02
|
|
|
a b550 is a very cost effective way to get a decently performing modern motherboard. it also uses passively cooled VRMs instead of the typical actively cooled x570s which is really only mandatory for beefier chips and overclocks and comes with some drawbacks. a 5600x does not require much, iirc it's a 65w tdp. you'd go x570 if you wanted more cores or overclocking headroom or needed a bunch of m.2 slots or whatever. no idea on intel, maybe something to do with how they rotate the socket so often
|
|
#
?
Nov 14, 2021 04:27
|
|
|
Note there's an actual X570S chipset now that doesn't have the active cooling requirement of the X570 chipset anymore. For Intel you used to take the "Z" boards because those were officially supporting the unlocked K-CPUs for overclocking and higher speed XMP RAM profiles. This doesn't really matter anymore because IIRC you can also run "unsupported" XMP RAM on B-chipsets and overclocking headroom is way down anyway (since the manufacturers are heavily binning their chips), and you don't need Z-chipsets to unlock power limits. The main difference is now in how many optional features, RGB and unnecessary (and sometimes necessary) power stages they put on boards for what price. On the AMD side it was never necessary to buy "X" mainboards to use the full potential of your CPU and RAM like that. Of course the X(370, 470, 570) mainboards often still had more features and better voltage regulation for overclocking, at a higher price. With X570 they eg. have the faster chipset interconnect for stuff like additional GPUs or more NVMe so you only need one if your build uses more than one GPU and a single NVMe (PCI-e 4.0) SSD and SATA drives. orcane fucked around with this message at 05:12 on Nov 14, 2021 |
|
#
?
Nov 14, 2021 04:51
|
|
|
orcane posted:You still don't need one if your build only uses a GPU and a single NVMe SSD eg. which easily fits into the features provided by B550 boards. To clarify, the B550 boards provide only one set of PCIe 4.0 NVMe lanes coming from the CPU. The chipset NVMe lanes are PCIe 3.0, which is slower, but for practical purposes makes no difference for the vast majority of workloads. So if you are running multiple 4.0 NVMe drives, then the X570 is necessary in that case to get the full 4.0 bandwidth to each drive.
|
|
#
?
Nov 14, 2021 04:57
|
|
|
if I'm not mistaken, Z590s only get that one m.2 4.0 lane (and zero if you have a tenth gen cpu)
|
|
#
?
Nov 14, 2021 05:02
|
|
|
https://www.techpowerup.com/288911/asus-and-gigabyte-enable-amd-ryzen-5000-series-processor-support-on-a320-chipsets AM4 still proving to be the king of sockets
|
|
#
?
Nov 14, 2021 05:08
|
|
|
Even when DDR5 has proliferated, AM4 will still be around. Not sure if it will be like the Core2Duo office box, slap a cheap GPU in it for a usable machine circa 2012 but yeah they'll keep crankin.
|
|
#
?
Nov 14, 2021 07:19
|
|
|
Rinkles posted:Why is the B550 usually recommended for a 5600X over an X570, when for Intel CPUs everybody (i.e., goons) say to get the higher end board (Z570 over the B560)? (both naming conventions are weird to me) Until the 500 series, the only intel chipset worth a drat for performance was Z because you couldn't use overclocking on non-Z boards. That included both CPU and memory "overclocking", aka running your memory at the XMP frequency printed on the box. You were restricted to 2666 or 2933, depending on the CPU (more expensive CPUs "supported" the higher frequency). Now how much this really mattered is a good question, because in recent history Intel CPUs have not gotten nearly as much performance boost from high-clocked memory, totally unlike Ryzen where it's critical. If your DDR4 3200 was running at 2933 it probably wasn't a performance tragedy, but it didn't feel good. Also only the Z was able to do a 8x/8x split to the 16x GPU lanes coming from the CPU, for SLI or whatever other purpose. SLI hasn't mattered in forever but it's another limitation of non-Z (which is still extant). So then there's the knock-on effect of that situation: informed DIYers bought nothing but Z chipset boards, everyone said "buy a Z board". So the only people who bought lower tiers had major budget limitations, were building an office PC or something, or were clueless idiots. This meant that the OEMs made nothing but budget trash for those chipsets. That's changing now, but even for the upcoming 600 series I expect there will be a 50/50 ratio on B660 still. As opposed to B550 where 90% of the models are at least acceptable and only a few are trash.
|
|
#
?
Nov 14, 2021 13:52
|
|
|
B560 boards are decent, though. The advice of only getting Z boards based on memory speed and overclocking restrictions is kinda outdated now. B560 supports XMP and you can mess with the CPU power limits too, I believe. You cannot mess with your multipliers, but overclocking in 10th and 11th gen CPUs is kind of worthless anyway, so no real loss there. So it's mostly just down to getting the board with the features you want, and avoiding ASRock in general (many of their intel boards are straight trash). However, B660 is sounding like it could be a big step down in terms of total PCIe bandwidth available, so if you want oodles of m.2 drives and IO and poo poo, you'll probably want to stick with the Z690 boards. With AMD, B550 is enough because X570 just doesn't bring enough new features to the table. B550 does everything you might want with regards to memory support and overclocking, and the X570 only changes the chipset m.2 slot from PCIe gen 3 to gen 4, while also adding support for some extra USB slots, but that's about it. The higher-end X570 boards are also the ones with the best VRMs, but that doesn't matter for normies.
|
|
#
?
Nov 14, 2021 14:20
|
|
|
|
| # ? May 17, 2024 05:05 |
|
|
Dr. Video Games 0031 posted:B560 boards are decent, though. There are a whole lot of them which do not meet Intel's optional maximum power delivery spec, which has real results on Rocket Lake chips. The ones that are actually good are not a big discount from Z590 boards, just because the VRMs needed to run those CPUs are so expensive. And the bottom half of B560 are far worse than the bottom half of B550s. Like if you were getting an i5 and didn't plan on removing power limits to run it at 250w, yeah there are plenty that are "decent". But a B460 that limited memory to 2933 was also "decent" by that standard.
|
|
#
?
Nov 14, 2021 15:13
|
|