Comments Locked

52 Comments

Back to Article

  • NextGen_Gamer - Tuesday, August 16, 2022 - link

    With regards to pricing details, there are two rumors for when the CPUs/AM5 platform will actually go on sale. One puts it on Sep 15, the other on Sep 27. If it is Sep 15, then I'm guessing we will hear how much everything will cost. That is only a couple of weeks after the presentation after all, and preorders (if they exist for them) would start even sooner than that.

    I'm also guessing we will get a “together we advance_Gaming” event sometime in October, or late September, for the RDNA-3 cards :)
  • meacupla - Tuesday, August 16, 2022 - link

    Pricing leak rumour is that 7000 series will be more expensive than a core/thread equivalent 5000 series. But I guess we'll have to see how hard intel punches with 13th gen.

    Hopefully DDR5 compatibility on 7000 series will be good with faster RAM sticks, as Intel has already shown that DDR5 is very slow outside of 2x16GB sticks on 2 slot mobos.

    Still waiting for an announcement of what the 600 series chipset will look like. The last rumour I heard was dual chipset?
  • erotomania - Tuesday, August 16, 2022 - link

    X670E (for Extreme) will be dual chipset for more I/O. X670 and B650 (or whatever the equivalents end up being this gen) will be standard single chipset.
  • James5mith - Tuesday, August 16, 2022 - link

    Like a return to old school northbridge/southbridge style designs? What would the second chipset be for in this case? Old school northbridge's were the location of the memory controller in the past. Now all of the northbridge functionality has been absorbed on-die or on-package into the CPU itself.
  • erotomania - Tuesday, August 16, 2022 - link

    Today's "Northbridge" is basically a fancy I/O hub, so two of them just increases I/O on the mobo. PCIe mainly, possibly USB and SATA.
  • at_clucks - Thursday, August 18, 2022 - link

    Yesterday's southbridge is today's northbridge.
  • erotomania - Tuesday, August 16, 2022 - link

    A) Together we advance_PCs
    B) Together we underscore PCs...in advance.
  • WaltC - Tuesday, August 16, 2022 - link

    You say in the article: "In a brief press release sent out this morning, AMD has announced that they will be delivering their eagerly anticipated Ryzen 7000 unveiling later this month as a live stream."

    "Live stream" says to me that it's a live event, which will be streamed live, as opposed to per-recorded.

    Yet your article near the end says, "Meanwhile, AMD’s press release does not mention whether the presentation will be recorded or live." I mean, doesn't "live stream" kind of give it away?

    This is an important release for AMD--not just a refresh, etc. So I imagine the 7:00 pm ET is also due to the live event.
  • Ryan Smith - Tuesday, August 16, 2022 - link

    "Yet your article near the end says, "Meanwhile, AMD’s press release does not mention whether the presentation will be recorded or live." I mean, doesn't "live stream" kind of give it away?"

    AMD has called previous recordings livestreams as well. If it's broadcast on YouTube and it's not uploaded as a recording (that is, it's in "live" mode and you can't skip ahead to the end), then AMD considers that a livestream.
  • abufrejoval - Tuesday, August 16, 2022 - link

    There don’t seem to be too many secrets left. Some things I’d like to know: any additional features with regards to control flow integrity? Shadow stacks were supported on Zen 3, but the forward branching extensions were missing: have they been added?

    Any news on per VM memory encryption being carried forward to client CPUs?

    I am mostly a bit disappointed by the desktop “chipsets”. Since the Zen 4 CCDs are shared between servers and desktops, they obviously are designed at least for CXL 1.1. And just like the Infinity Fabric that means a lot of flexibility what protocol you speak on which bundles of lanes.

    Except that it doesn’t carry forward into the desktop space.

    I understand that using half of a full IOD as X570 “chipset” was expensive in terms of power and chip cost, but it was genius in having at least potentially IF between the two IODs. I’d just love to see an extension of this perhaps even with these next-gen IODs talking CXL, too.

    There are 7 groups of PCIe 5.0 x4 coming off the SoC and just imagine if these 7 groups were to be slots on a passive backplane. Add-on cards would then have an IOD switch at their base and fan out to say 4-8 M.2 slots, perhaps with a bit of USB strewn in to fill the slot’s bracket. Obviously, it could also be one or several 10/40/100 Gbit NICs or CXL. And CXL includes RAM, accelerators, TPUs, DPUs, you-name-it.

    The current ‘solution’ which has one set being dedicated to USB4, another for the ‘South bridge’, daisy chained or not, is just lame! And a complete waste.

    In this world of NVMe storage, any networking below 10Gbit/s is just insane. We’ve been crippled by Gbit NICs since the last millennium and this has just got to go! The Aquantia/Marvell ACQ113 has been around for some time now, yet you have to shell out an extra €400 to have that €20 chip included on a mainboard?

    This ASmedia daisy chain setup may be extremely cheap and commodity (surely it won’t be sold that way), but it’s utterly wasting the potential of Zen 4 hardware.

    AMD has an opportunity to completely redefine the desktop and entry/edge level server market and deliver a level of flexibility and expandability that is dizzying, but it’s letting it slide through their hands. Neither Intel nor ARM can currently match that and it’s one of those once-in-a-lifetime opportunities that won’t come back.

    AMD, please wake up and deliver a new entry level mainboard architecture. No need to change AM5 or the SoCs, only a complete revamp of the mainboards will suffice.

    You can start by hooking another ASmedia chip to the ‘NVMe’ or ‘USB4’ lane bundles, daisy chain or not. Put those on a slot and you can cool and use them on a tower chassis.
  • ballsystemlord - Tuesday, August 16, 2022 - link

    I'd also like to see more PCIe ports.
  • meacupla - Wednesday, August 17, 2022 - link

    did you mean PCIe lanes?
    24x PCIe 5.0 lanes is equivalent bandwidth to 48x PCIe 4.0 lanes
    That's enough to run two 5.0 NVMe SSDs, and two GPUs, and running 2x GPUs is a dead feature anyways.
  • lilkwarrior - Wednesday, August 17, 2022 - link

    Running 2x GPUs is extremely common for productive computer users, especially high-end Ryzen/Intel CPUs and the respective motherboards.

    Even Apple unapologetically supports mGPU rendering out of common sense. It's game developers who haven't emphasized it; in turn pure gamers haven't had an incentive to care but that audience likes to talk as though high-end computing hardware is primarily for them when most of that audience can hardly comfortably afford or go above entry-level motherboards and CPUs.

    12VHPWR standard is specifically designed to enable single GPU–especially dual GPU die ones–the power to replace two GPUs with one + one cable to connect them. This allows four-way mGPU rendering to be replaced by two GPUs and so on.

    This is common knowledge in the AI and creative professional industries .
  • meacupla - Wednesday, August 17, 2022 - link

    okay, but 24 lanes of PCIe 5.0 has the bandwidth for that, so the problem is what?
  • schujj07 - Wednesday, August 17, 2022 - link

    Unless you have a PCIe switch the added bandwidth won't allow you to turn 24x PCIe 5 lanes into 48x PCIe 4 lanes.
  • Threska - Wednesday, August 17, 2022 - link

    "This is common knowledge in the AI and creative professional industries ."

    That's about the only two, since the current setup is dealing with transient and bulk power issues.
  • ballsystemlord - Wednesday, August 17, 2022 - link

    No, PCIe ports.
  • abufrejoval - Wednesday, August 17, 2022 - link

    That touches on something completely unclear to me so far: can PCIe switches actually translate/manage the trade-off between PCIe generations and lanes?

    I'm not sure how much PCIe is packet based or connection based: it seems to have elments of both.

    But say you have a NIC that uses 4 lanes of PCIe 2.0 to deliver 10Gbit/s Etherent: could a PCIe switch chip translate that to a single PCIe 4.0 lane, which evidently provides the bandwidth (and vice-versa)? I naturally assumed it could (network mentality, I guess), but currently I'm very much in doubt.

    Obviously going backward from PCIe 5.0 all the way to PCIe 1.0 is no issue, nor is using fewer lanes than a slot might provide. But can "store-and-forward" switches translate/trade between versions and lanes?

    And yeah, GPUs aren't just used for games. Some are still used for mining and evidently in that context the CPU/GPU bandwidth is of no matter, which is why they often operate only with a single PCIe lane for the transfer of metadata.

    And also yes, I've been sticking two V100 into a few servers for machine learning purposes, where the halving the bandwidth between CPU/GPU didn't make things much worse already, because you'd better make sure to fit your models inside the GPU memory anyway: host RAM is the new type in machine learning...
  • DanaGoyette - Wednesday, August 17, 2022 - link

    Since the link between chipset and SOC is just PCIe anyway, I'd love to see them make add-in PCIe cards with those chips on them, exposing most of the useful IO of the chipset. Imagine if you could get a bunch of SATA ports, and the 10Gbe MAC if it even still exists, all on one card...
  • spaceship9876 - Tuesday, August 16, 2022 - link

    Do we know whether this comes with built-in 2.5gbps ethernet support like intel's latest chips?
  • shabby - Tuesday, August 16, 2022 - link

    One can only hope, 1gbps ports need to die a quick death. It's the slowest port on a mobo now.
  • ballsystemlord - Tuesday, August 16, 2022 - link

    But most Americans do not even have 1/2 Gb speed internet. So there's no point. :cry:
  • ballsystemlord - Tuesday, August 16, 2022 - link

    Yes, there's an actual US Gov report that says this.
  • abufrejoval - Wednesday, August 17, 2022 - link

    Not all traffic is North-South, there are people with LANs.

    I've had 10Gbit networking even in my home-lab for 10 years and only got Gbit Internet last year.
  • shabby - Wednesday, August 17, 2022 - link

    Have you heard of home networking?
  • Qasar - Tuesday, August 16, 2022 - link

    unless 2.5+ gig switches become less expensive, and more widely available, this is probably not going to happen any time soon
  • abufrejoval - Wednesday, August 17, 2022 - link

    There are 8-port NBase-T switches capable of 1/2.5/5/10Gbit available for €350 with taxes, e.g. from Netgear and TP-Link...

    ...as globally as Amazon and just a click away.

    I have a Buffalo 12-port that's no longer sold and a NetGear 8-port that also comes in a 5-port variant. I resorted to Noctua fans to make the Buffalo 'home-lab compatible', the NetGear was easy on the ears from the start.

    10GBase-T started at around 10 Watts power consumption per port 10 years ago, because the modulation required on the copper wires to push that bandwidth was at a similar level of complexity as the last generation of telephone modems. But with NBase-T and Green Ethernet efforts that's been pushed down to around 3 Watts on active ports for 10Gbit/s, much less for 2.5Gbit/s. The result is switches that have a fan just in case someone is pushing 10Gbit/s on all ports without any Green Ethernet, that will remain silent or inaudible most of the time.

    ~€50/port is obviously more than Gbit-switches today, but those have just become dirt cheap.

    And it's still less than 50% per port of what a 10Gbit add-on NIC will cost you. Once switch port cost sink below NIC cost, you should stop complain about switch cost. I know that wasn't the case for years, but please note that the tide has changed.

    Of course, I don't expect that an Aquantia ACQ113 on-board chip will cost anywhere near the €90 they are still charging for its ACQ107 predecessor add-in NIC.

    I believe RealTek 2.5Gbit NICs are about €1 in volume yet sell at €25 in their USB3 incarnation: there is a significant overhead for any SKU and packaging a NIC, which is why it's so infuriating they don't just add the ACQ113 and be done with it: it will just handle anything from 100Mbit to 10Gbit and has great support on all operating systems.

    Something that no longer applies to anything Intel is selling in the 1-10Gbit Ethernet space today: they have moved as far away from 'one driver for all' as you can imagine in your worst nightmares.

    And it doesn't even buy you features for virtualization or relevant offload engines... I can't think of any reason why anyone would chose an Intel i225 over an ACQ113 on technical merits alone.
  • Qasar - Wednesday, August 17, 2022 - link

    never seen those switches for sale here, the only switches i have been able to find here that are 2.5gig or faster, are a minimum of $800 and i think it was only a 5 port. i would need at least 8 ports. or they are SFP+ based, which adds to the cost over all.
  • spamaway - Thursday, August 18, 2022 - link

    Take a look at what Mikrotik sells. Their CRS309-1G-8S+IN is passively cooled, has eight SFP+ cages, and has an MSRP of 269 USD. The CRS305-1G-4S+IN is also passively cooled, has four SFP+ cages, and an MSRP of 149 USD.

    I have been using several of their CRS326-24G-2S+IN switches for months now, and they work great. I highly recommend buying their hardware.
  • Qasar - Thursday, August 18, 2022 - link

    never heard of that one. and needs sfp+ so same with the ones here, thats adds to the cost.
  • spamaway - Friday, August 19, 2022 - link

    10gtek sells 10GBase-SR SFPs for ~16 USD per, and are usually less than that if you buy two, five, or ten at a time.

    Unless you insist on 10GBase-T, decent SFP+ modules are cheap. (OM3 or OM4 fiber is pretty cheap, too.)
  • Qasar - Friday, August 19, 2022 - link

    10gtek ?? never heard of them either.
    10g sfp+ modules are $60, and rj45 10g is $100
    at least here, 10g is still too expensive. ive been looking at getting a gigabit switch that supports LACP properly for now. the one i have doesnt seem to support link aggregation fully
  • schujj07 - Wednesday, August 17, 2022 - link

    1GbE is plenty fast for home user management/Internet. The only place in the home that 1GbE becomes a problem is if you have a NAS and need to retrieve a large amount of data. If all you do is use your NAS for say streaming 4k movies a 1GbE connection is plenty.
  • abufrejoval - Wednesday, August 17, 2022 - link

    I find this attitude both infuriating and condescending!

    The whole x86 server ecosystem depended on the tenet, that desktop hardware could be reused for servers, undercutting and ultimately killing near everything Unix and RISC.

    And it's just what I want to continue to exploit.

    So I build HCI clusters of x86 systems using RAIDs, SSDs and now NVMe and Gbit Ethernet just isn't a match to more than a single drive of spinning rust.

    The network fabric needs to match the bandwidth of your storage at least, and width CXL it moves to the bandwidth of RAM.

    10Gbit Ethernet is already far too slow, vis-à-vis 20Gbit USB 3.2 or 40 Gbits Thunderbolt 3.
  • schujj07 - Thursday, August 18, 2022 - link

    99% of people are fine with GbE. You are not a typical use case. You could look into getting some 56Gb Mellanox ConnectX-3 NICs on ebay and a used switch for it. Lots of enterprises are getting rid of their 10/40 setups and going 25/100 or faster. The old stuff you can get at a good price.
  • erotomania - Thursday, August 25, 2022 - link

    No. Just no.
  • ZoZo - Wednesday, August 17, 2022 - link

    Intel's latest desktop chips come with a built-in Ethernet controller. They still require a separate controller such as the I225-V.
  • ZoZo - Wednesday, August 17, 2022 - link

    I meant, they *don't* come with one. (is there really no editing comments here?)
  • ballsystemlord - Wednesday, August 17, 2022 - link

    No editing for any of us. It's actually the only bad thing about AT's comment section.
  • Flunk - Wednesday, August 17, 2022 - link

    No Ethernet on chip.
  • Silver5urfer - Tuesday, August 16, 2022 - link

    The important infact most important feature of all is how stable is IODie now. Ryzen was a failure in that and my research, their Ryzen 5000 has this instability issue even if you barely touch Curve Optimizer and or that DRAM Freq above 3600MHz. Plus the BIOS has no proper documentation like Intel's Datasheet. I don't expect to have datasheets since many AMD users will run at bone stock. But the IOD stability (USB, WHEA and general machine response) it will take some time since it has to be in the market just like Intel's BS LGA1700 ILM failure. I just want to see if they make any special announcement for that piece.

    As for the chipset, Techpowerup mentioned the CPU has PCIe5.0 link speed to the PCH but the Chipset downgrades it to PCIe 4.0 that's a bummer since the lanes are not like Intel so it has to depend on PCIe but with generation downgrade that means the I/O won't be much high, esp those PCIe SSDs you all want to run and etc.

    For the meat, the ST performance and how MT scales will it beat Intel in ST or not, and how the Cores are varied now and esp the Voltage. Ryzen rammed 1.4v through all Ryzen 5000 and they barely can get to boost clocks on top of the stupid issues with IODie, X3D had 1.3v but it has all blocked to control and tune, so a worthless chip for people who want to tinker.

    DRAM IF how is it gonna work, same like Intel's 2 gear BS or does AMD want to step to next level by having 1:1 clock speed link ? EXPO is their new XMP rival gotta see how it works in action.
  • Qasar - Tuesday, August 16, 2022 - link

    " their Ryzen 5000 has this instability issue even if you barely touch Curve Optimizer and or that DRAM Freq above 3600MHz " 4 ryzen 5k comps here, no issues at all with them since i bought them, so there could be other reasons for these "issues " my 5900x was running 3600mhz ram too.
  • DanaGoyette - Wednesday, August 17, 2022 - link

    Hell, even with stock clocks, I've had instability issues with multiple CPUs (3700X and 5800X), multiple motherboards, multiple sets of RAM (with ECC, even), multiple GPUs (5700XT, RMA replacement 5700XT, and Radeon Pro W5700), and multiple power supplies.

    I still get WHEA errors sometimes (PCIe or CPU Cache), and the GPU driver sometimes dies or drops the application (causing FF14 to crash, or Minecraft Bedrock to reload). It happened all the time with the Ryzen 5800X, and less often once I switched back to the 3700X.

    If I'd known the 5800X was going to be such a pain, I would've stuck with the 3700X, and gone with a Xeon D for the server machine I had moved the 3700X into. It's honestly making me consider going back to Intel next time.
  • Qasar - Wednesday, August 17, 2022 - link

    the i guess the 4 comps i have, must of been the few good ones. * shrug *
    heck the x99 based comp i also use, still has issues.
  • schujj07 - Friday, August 19, 2022 - link

    I have built a 2400G, 2600, and 3600 Ryzen desktop systems without a single one having an issue. Where I work I am the person who has designed our entire data center upgrade and laptop upgrade. All of those run Epyc or Ryzen and never had an issue either.
  • meacupla - Monday, August 22, 2022 - link

    In my experience, the vast majority of stability issues on AM4 platform are the result of insufficient memory voltage. Or rather, the crappy RAM kits that work fine on intel, don't work on AM4, unless the voltage is upped to 1.5V. and there are a lot of crappy RAM kits out there.
  • Silver5urfer - Tuesday, August 23, 2022 - link

    DDR4 1.5v means you want at-least 32GB 16x2 4000MHz Dual Rank B-Die C15 speeds, and Ryzen cannot handle that fast memory Ryzen cannot even go above 3600MHz without getting issues and stability. Not even Alder Lake can that fast Gear 1, last one which can do it without gear ratio was Comet Lake 10th gen 10900K/KF.

    Pumping 1.5v and expecting stability not going to make anything better. Ryzen's weakness is IODie, AMD fked up big time. They must fix it on 7000 series or else history will repeat. Beautiful numbers but IRL tons of issues.
  • GeoffreyA - Thursday, August 18, 2022 - link

    I wonder, have there been similar issues with Cezanne?
  • erotomania - Thursday, August 25, 2022 - link

    "Your research"? Dying to read that...
  • erotomania - Thursday, August 25, 2022 - link

    "...but with generation downgrade that means the I/O won't be much high"

    But you sure are
  • Flunk - Wednesday, August 17, 2022 - link

    Wow, AMD has really been executing lately. Props to Lisa Su, her entire executive team and the hundreds of engineers who make this possible.
  • spamaway - Friday, August 19, 2022 - link

    10gtek sells 10GBase-SR SFPs for ~16 USD per, and are usually less than that if you buy two, five, or ten at a time.

    Unless you insist on 10GBase-T, decent SFP+ modules are cheap. (OM3 or OM4 fiber is pretty cheap, too.)

Log in

Don't have an account? Sign up now