AMD extends Socket AM5 support through at least 2029; AM4 refuses to die

The 5800X3D returns at $349, while the 7700X3D debuts at $329.

See full article...

 

[DanNeely]

Unless AMD is somehow not able to sell all the 9000 series X3D parts they're making I really don't understand why they're resuming 5800x3d production. The limiting factor for x3D chips is the stacking process from TSMC used to make them. It's a legacy product and AMD is already buying ~100% of its output, so the number of x3D chips they can make is fixed and every 5000 series part they make is one fewer 9000 series one.

The 7700 is at least theoretically explainable if they're new old stock that failed to meet the specs needed to sell as a 7800 but were stable as lower clocks.

 

[norton_I]

So there is nothing technically stopping AMD from releasing versions of Zen 4 or Zen 5 on AM4

It would need a significant redesign of the IO die to replace the memory controller with a DDR4 controller. So yes they could, but it would be a significant technical project and new silicon in order to make an underwhelming product starved for memory bandwidth.

 

[RickyP784]

In the modern world of system-on-chips, the socket mostly determines the type of memory. So AM4 == dual channel DDR4, AM5 == dual channel DDR5. It also determines the number of PCIe lanes available, including thunderbolt / USB4 if directly connected to the CPU. It includes other things too, but those are the most obvious ones when comparing desktop CPU sockets between generations.

It also can dictate the fastest PCIe speed, but there is more room for change here. When AM4 launched with Zen+ it supported PCIe3, but Zen2 upgraded to PCIe4. AM5 supports at least PCIe4, but with the right motherboard and chipset supports PCIe5.

Ok, but is that a function of the hardware using the socket or a limitation of the socket pinout itself?

 

[RickyP784]

The socket is mostly the interface that connects the CPU to memory, power, and I/O. So there is nothing technically stopping AMD from releasing versions of Zen 4 or Zen 5 on AM4, it is mostly a business decision. While others focused on the memory and I/O aspects that AM5 introduces, a big one is a much higher power limit that the top end AM5 CPUs actively exploit to run a lot faster.

Is the higher power draw limited by the socket (quantity or utilization of the pins) or by hardware detection of an AM4 vs AM5 CPU?

 

[ERIFNOMI]

Unless AMD is somehow not able to sell all the 9000 series X3D parts they're making I really don't understand why they're resuming 5800x3d production. The limiting factor for x3D chips is the stacking process from TSMC used to make them. It's a legacy product and AMD is already buying ~100% of its output, so the number of x3D chips they can make is fixed and every 5000 series part they make is one fewer 9000 series one.

The 7700 is at least theoretically explainable if they're new old stock that failed to meet the specs needed to sell as a 7800 but were stable as lower clocks.

Because the 5000 series is AM4 and RAM costs an arm and a dick right now.

 

[cameron2]

So, dumb QUESTION. Does the socket introduce any performance or new features or is it just a hardware plug that prevents sufficiently new CPUs from working in older boards?

Asked another way: is AM5 inherently faster or better than AM4, or is it just "better" cause the hardware that requires it is better?

AM5 is inherently wider and faster with respect to memory, and as you increase the number of cores, memory becomes a natural bottleneck. For most games, this doesn't matter much, but for multi-threaded business apps built in C# or Java or JavaScript (e.g. "Node"), it will make a significant difference. For AM5 with the 6xx chipset, you generally want DDR5 6000, and with the 8xx chipset, you generally want DDR5 6400. (If you're a gamer, there's some advantage around DDR5 8800 on the 8xx chipset, but not worth the crazy prices.)

We're in agreement on all your points. To the OP's point - having dual channel RAM or multiple memory paths on a slower RAM can end up being faster than less channels on a faster chip-speed RAM. Parallelism for the win.

So - never put one stick of memory in a machine - use two of half the size - it'll be faster. If you want to get the most performance out of your motherboard - fill all the RAM slots. If you want to leave room for expansion (without discarding old RAM) - fill half the RAM slots.

Dual channel for the win, of course! But you do not want to fill up all the RAM slots -- for the same amount of RAM capacity, you're significantly faster at only one DIMM per channel, and your motherboard where tell you which DIMM sockets to populate if you only have 2 DIMMs (one per channel) to get the best perf and stability. I have a lot of AMD machines now (I run a software startup), but on my main gaming machine I run 64GB (32GB per DIMM), and for work I have e.g. a test rig with 256GB (4x 64GB per DIMM).

 

[Barleyman]

4 min boot time in Windows 10? Was it a spinning disk, running Norton Antivirus scan at boot, or something else?

Something just seems off here. Windows boot that bad means something is slowing it down...

Hey, let's not poke holes on "cool story, bro" Linux war story, shall we? It's perfectly possible that Linux would dramatically outperform Windows on 32GB kit, if, for example, the OP just forgot to add that HDD was swapped to SDD in this story.

Most likely your suspicion of something hinky going on at the boot time is correct, or perhaps this "4 minutes" just happened to include windows update or something...

 

[Pixy Misa Mk II]

I doubt it. The AI MAX chips are mainly in laptops and a few SFF PCs aimed purely at AI enthusiasts, if they were truly their path forward they'd be trying to put them in larger desktops where improved cooling would allow for crazy power using. AM6 will happen, but the current market fuckery means we probably won't see anything serious until 2028 or later.

Well, AM6 implies DDR6 memory, and that's not expected until 2028 anyway.

But AMD has said (in discussion with Framework) that it's not practical to run the AI Max chips with regular DIMMs, making it a problem for general-purpose desktops. Not sure exactly why that is, since their Threadripper and Epyc CPUs have anything from four to twelve memory channels.

 

[Pixy Misa Mk II]

Unless AMD is somehow not able to sell all the 9000 series X3D parts they're making I really don't understand why they're resuming 5800x3d production. The limiting factor for x3D chips is the stacking process from TSMC used to make them. It's a legacy product and AMD is already buying ~100% of its output, so the number of x3D chips they can make is fixed and every 5000 series part they make is one fewer 9000 series one.

Memory prices and customer demand.

People have been paying $500 in the used market for 5800X3D chips to upgrade existing AM4 computers, rather than building a whole new system.

 

[conifernoggin]

I know its never going to happen, but as long as the AM4 is going to stick around a little longer, I have been wishing for a new AM4/DDR 4 chipset or a revised motherboard design to fix the USB connectivity issues that I (and many others) have experienced from day one, in my case with two versions of ASUS AM4/DDR4 - X570 and B-550-F ROG, the CPUs are 5950x, 5900x. I believe MSI, ASUS and Gigabyte boards all suffered these issues.

The X570 & B-550-F ROG boards USB problems were only slightly mitigated by downgrading the PCIe Gen 4 capable buss settings to Gen 3, disabling C states, and in Device manager turning off the USB hub OS power management settings, and also deleting the USB host controllers - this last fix leaves you looking at your monitor without any mouse or keyboard to shut down with, but upon reboot the host controllers load back in. BIOS updates never made any appreciable difference.

The B-550 board is about 11 months old, the x570 board is about 3 years now. Win 10/Win 11. DRAM and drives were purchased well before the shortage began last fall.
Since problem is USB, both keyboard and mouse will occasionally stutter or stop functioning or sometimes the keyboard takes the last character typed, and runs with it until Esc makes it quit the infinite repeat.

OT: All of this was on top of new installs of Win 11 and it made for a rough few months before I was able to sort out the Win stuff from the AMD issues.

I also have a 9 YO B350M-E with Ryzen 7 5700x & Win 10 that did not have this issue, but it is limited to 32GB and 4 SATA connections.

 

[Marlor_AU]

Because the 5000 series is AM4 and RAM costs an arm and a dick right now.

Exactly. Plenty of us have large quantities of perfectly serviceable DDR4 from times when it was extremely cheap. During my various builds over the past decade, "future proofing" on RAM was a no-brainer, so I often went entirely overboard. Between my various PCs here, I have 192GB of perfectly good DDR4 RAM (mostly DDR4-3200).

This is paired with processors that are getting a bit long in the tooth. An upgrade to a new DDR4/AM4 motherboard and a 5000 series processor is a very affordable way to upgrade and be set to ride out the RAM price bubble. I've just done this with one of my PCs, and I certainly can't complain about the performance – I'll be set for quite some time.

 

[norton_I]

Ok, but is that a function of the hardware using the socket or a limitation of the socket pinout itself?

I'm not sure exactly what you mean?

DDR4 isn't compatible with DDR5 in any way. It uses different voltages, different signalling, breaks into subchannels that DDR 4 doesn't have, uses significantly different clock frequency, and so on. If you for some reason made a Zen5 CPU that could physically fit in an AM4 socket it wouldn't work in any existing AM4 motherboard, since those motherboards would have DDR4 dimm sockets and be designed for DDR4 memory. The zen5 memory controller wouldn't be able to talk to the memory. You would need a new motherboard which in turn wouldn't support older CPUs.

The socket definition mostly isn't about the physical connector, it's what signals go over those pins That's what defines "AM5". In the case of AM5, the socket also has more pins with the move from PGA to LGA. This allows for higher peak power delivery and better signal integrity on the high speed signals. But that's not the tricky part. You couldn't make an adapter for AM5 to AM4 even if you limited it to lower TDP chips.

 

[GTechnician]

And just as I was about to give them a trophy for showing up. I guess they won't even be getting that now.
Not giving the option to upgrade is no option.

 

[GTechnician]

Exactly. Plenty of us have large quantities of perfectly serviceable DDR4 from times when it was extremely cheap. During my various builds over the past decade, "future proofing" on RAM was a no-brainer, so I often went entirely overboard. Between my various PCs here, I have 192GB of perfectly good DDR4 RAM (mostly DDR4-3200).

This is paired with processors that are getting a bit long in the tooth. An upgrade to a new DDR4/AM4 motherboard and a 5000 series processor is a very affordable way to upgrade and be set to ride out the RAM price bubble. I've just done this with one of my PCs, and I certainly can't complain about the performance – I'll be set for quite some time.

Unfortunately w are on DDR5 now and need the added m.2 slots the new boards have.

 

[eldakka]

It would need a significant redesign of the IO die to replace the memory controller with a DDR4 controller. So yes they could, but it would be a significant technical project and new silicon in order to make an underwhelming product starved for memory bandwidth.

I doubt they'd need to do that.

They could probably pair a zen4 CCD with a zen 3 I/O die. All the I/O dies use Infinity fabric to connect the CCD to the I/O die. Sure, it might take some Infinity protocol-type tweaking, but probably very little.

But the architecture of the zen 4 was designed around the increased bandwidth available on DDR5. Therefore many of the enchancements made to things like wider/deeper pipelines, instruction issue units, etc. probably can't be kept fed by the slower memory controller, so while it would work, you'd lose much of the generational IPC gains. Sure, especially with an X3D you might be able to keep some of those generational IPC gains, but still a zen4 CCD being fed by a zen3 DDR4 I/O die wouldn't be nearly as impressive as zen 4 CCD + zen 4 DDR5 I/O die, therefore there is likely little point, certainly not enough to justify the RnD and the setup costs of a production line to make such a product.

 

[DarxKull]

Or that it's a limited run for which economies of scale don't apply.

I wish somebody would do an investigation of how extra cache affects tasks such as running a ton of VMs on consumer hardware, although ideally you'd want it on both CCDs for that, which doesn't apply to the older X3D chips.

Its a limited run to make scalpers loads of cash. No one is going to get one of these from a first party retailer

 

[StuiWooi]

Back before the RAM-pocalypse, when I was speccing out an AM5 build, I wished there was a 7700X3D, I think I had the intention of pairing with a 7700XT GPU because I liked the way the numbers paired.

Now neither are are current and it's probably more expensive but I'm still tempted to revive the idea...

 

[SeanJW]

What the heck was that computer doing?!? My old 8gb Win10 I5 computer I built in 2014 booted into Windows in about 30 seconds. I am still running on the same copy of Win10, now on a Ryzen 5 5500 with 16gb and its boot time is still about 30 seconds. Quite honestly, those times are being conservative, I wouldn't be surprised if the actual boot times are 15-20 seconds. Both systems are running on SSDs.

"Boot time" is a bit nebulous. Is it "to the login prompt" or "to the desktop and ready to go"? Because if it's the latter, the shit that's not from Microsoft can end up taking a truly hideous amount of time loading for no good reason... I've watched Windows start, then Nextcloud slowly appears....then the Bitwarden app....eventually the prompt for my SSH private key password... If you manually start them, they're all ready almost immediately, but during startup, they pace themselves....

Same thing under Linux or macOS? (literally - Nextcloud, Bitwarden, an SSH key prompt etc), under any of the desktop environments I've tried... much smoother and things are just ready to go promptly.

Couldn't for the life of me figure out why. I expect because I didn't use Windows regularly it decided that the boot period was the perfect time to catch up madly, but so much crap loads under umbrella services, you couldn't pin it down.

 

[Dr.Kamiya]

The switch to AM6 is supposed to come with DDR6 right? I take this to mean that either the transition is delayed or they expect DDR6 to be extremely expensive, probably even more than when DDR5 was the shiny new thing.

 

[CrystalCowboy]

I'm curious if there's any subtle different between the new 5800x3d and the original one.

And as for anyone on the fence with an am4 computer and an older CPU, I upgraded from a 3700x to a 5800x3d and it was a very noticeable leap. I'm not surprised they did this at all. I thought that when they originally stopped making them that they'd do another run. And with the price of RAM, it's a no brainer.

Subtle differences. Anything you would notice? I don't know.

AMD ‘had to re-engineer’ the Ryzen 7 5800X3D for a re-release — 10th Anniversary Edition chip had ‘a whole body of engineering work’ put into it

... Despite the re-release looking like a simple product spin-up, AMD’s David McAfee, VP and general manager of Radeon and Ryzen, says that “a whole body of engineering work” went into the re-release, as the original bonding process TSMC used for the Ryzen 7 5800X3D was no longer available...
“It completely changed the characteristics of how those two pieces of silicon are bonded together and how they were stacked together, and so when that first-gen facility really kind of went offline, then it meant there was a whole, you know, body of engineering work that had to be done to understand if we could even migrate the 5800X3D to the new, second-generation stacking process,” McAfee said...

 

[norton_I]

The switch to AM6 is supposed to come with DDR6 right? I take this to mean that either the transition is delayed or they expect DDR6 to be extremely expensive, probably even more than when DDR5 was the shiny new thing.

Yes. "Extending support for AM5" means "we think ram shortages will delay consumer DDR6 availability at least a few years"

 

[sing.electric]

I doubt it. The AI MAX chips are mainly in laptops and a few SFF PCs aimed purely at AI enthusiasts, if they were truly their path forward they'd be trying to put them in larger desktops where improved cooling would allow for crazy power using. AM6 will happen, but the current market fuckery means we probably won't see anything serious until 2028 or later.

My point about the AI Max (Strix Halo) wasn't about what the parts that have been released can do, it's about the fact that even though it uses Zen 5, it's got a "sea of wires" interconnect that replaces AMD's infinity fabric, but now, 18 months after Strix Halo came out, we haven't seen traditional desktop processors released with the tech (which helps drive some of the efficiency gains we see on the AI MAX parts vs say, a 9800X3D or whatever - it isn't just running at clockspeeds that hit a better part of the efficiency curve).

For AM6 to happen, AMD would have to decide that it was worth engineering a new platform for a market that even in 2028 could very well be starved for affordable parts (especially since new systems would typically also need a discrete GPU).

 

[EduSantos]

Is the higher power draw limited by the socket (quantity or utilization of the pins) or by hardware detection of an AM4 vs AM5 CPU?

Very much a physical socket design issue: in order to safely draw additional current you have to increase the area (or the pins would melt). You can do that by either using larger, wider wires (standard electric installation solution), or in the realm of semiconductors, by using more pins. Additionally, you are going to need more voltage regulators and better motherboard power delivery, which is pretty much part of the overall spec (and is one of the reasons why motherboards for newer sockets tend to cost more).

 

[EduSantos]

It would need a significant redesign of the IO die to replace the memory controller with a DDR4 controller. So yes they could, but it would be a significant technical project and new silicon in order to make an underwhelming product starved for memory bandwidth.

"Starved" is a very strong word. Keeping pipelines fed is very much a function of the internal core design and cache structure, in order to minimize latencies, and internal cache bandwidth matters a lot more. Zen 4 is not that different from Zen 3 in a lot of the internal aspects: https://chipsandcheese.com/p/amds-zen-4-part-2-memory-subsystem-and-conclusion

 

[EduSantos]

I doubt they'd need to do that.

They could probably pair a zen4 CCD with a zen 3 I/O die. All the I/O dies use Infinity fabric to connect the CCD to the I/O die. Sure, it might take some Infinity protocol-type tweaking, but probably very little.

But the architecture of the zen 4 was designed around the increased bandwidth available on DDR5. Therefore many of the enchancements made to things like wider/deeper pipelines, instruction issue units, etc. probably can't be kept fed by the slower memory controller, so while it would work, you'd lose much of the generational IPC gains. Sure, especially with an X3D you might be able to keep some of those generational IPC gains, but still a zen4 CCD being fed by a zen3 DDR4 I/O die wouldn't be nearly as impressive as zen 4 CCD + zen 4 DDR5 I/O die, therefore there is likely little point, certainly not enough to justify the RnD and the setup costs of a production line to make such a product.

The Zen 4 design is a lot closer to Zen 3 than most people think, regarding memory performance: https://chipsandcheese.com/p/amds-zen-4-part-2-memory-subsystem-and-conclusion

Sure, DDR5 has 50% more total bandwidth. But by the time you need that you already stalled the pipeline.

 

[CrystalCowboy]

AMD is considering a potential Ryzen 5 9600X3D — company says six-core Zen 5 X3D chip 'maybe something we look at doing... later this year'

 

[evan_s]

The switch to AM6 is supposed to come with DDR6 right? I take this to mean that either the transition is delayed or they expect DDR6 to be extremely expensive, probably even more than when DDR5 was the shiny new thing.

I don't think AMD has officially announced anything yet but that would be the logical thing and what everyone assumes will happen. New socket will come with new DDR6 support, possibly more PCI-E lanes directly off the CPU or higher minimum PCI-E gen, better connection to the chipset and what ever other little tweaks AMD decides to throw in there.

 

[gosand]

Perhaps you wanted to say that extra RAM above what you need isn't critical.

But, yeah having sufficient RAM is more important the CPU speed, your graphics card or anything else - just try running a computer with 2 GB of RAM these days.

Obligatory footnote - I'm not saying that extra RAM won't be used or won't be useful - any OS will gladly use it for caching.

I run Linux, and am on AM4. I ran 8GB for many years, and upgraded to 16GB about 8 years ago. I was never really able to use it all. I once tried, and was able to use about 13GB. I upgraded my RAM to 32GB a couple of years ago just because I was building out my son's PC and 16GB of DDR4 was so cheap. Now is a different story, people are even gouging on ebay just because they can.

I had a Ryzen5 5600G processor die on me (mobo too, not sure who fried who) but I actually downgraded to a 5500. At the time I got it for $85. I can max out the cpu when trancoding videos or something, but in general it is quite great for everything else. I do play some games, but nothing too intensive. If anything I could stand to upgrade my GPU, it's a Radeon 5600XT.

I honestly don't do anything that CPU intensive, and wonder who does. Audio/video production or compiling I could see, but I don't know if people are just chasing the latest and greatest because newer is better. My machine runs fast, I can't quite figure out what I would use more CPU or RAM for anyway.

 

[evan_s]

I run Linux, and am on AM4. I ran 8GB for many years, and upgraded to 16GB about 8 years ago. I was never really able to use it all. I once tried, and was able to use about 13GB. I upgraded my RAM to 32GB a couple of years ago just because I was building out my son's PC and 16GB of DDR4 was so cheap. Now is a different story, people are even gouging on ebay just because they can.

I had a Ryzen5 5600G processor die on me (mobo too, not sure who fried who) but I actually downgraded to a 5500. At the time I got it for $85. I can max out the cpu when trancoding videos or something, but in general it is quite great for everything else. I do play some games, but nothing too intensive. If anything I could stand to upgrade my GPU, it's a Radeon 5600XT.

I honestly don't do anything that CPU intensive, and wonder who does. Audio/video production or compiling I could see, but I don't know if people are just chasing the latest and greatest because newer is better. My machine runs fast, I can't quite figure out what I would use more CPU or RAM for anyway.

The 5500 is basically the 5600g with the iGPU disabled so since you have a dGPU it's not much of a downgrade. On the CPU front some of it is just more is always better. For games shader compilation is one area where those faster CPUs with lots of cores can be handy. Steam on Linux tends to mitigate some of that but some games can have pretty lengthy shader compilation steps every time they update and they update fairly frequently. For actual gameplay faster CPUs are often about higher max framerate but if you aren't targeting high framerates and/or are already hitting your max monitor refresh rate you are probably ending up GPU limited anyway.

 

[DRJlaw]

Unless AMD is somehow not able to sell all the 9000 series X3D parts they're making I really don't understand why they're resuming 5800x3d production.

Because if you've got an AM4 platform with 32GB or more of DDR4 and anything between a Ryzen 1000-series and a Ryzen 5300-series, and stretching it to the 5500-series, the cost of the CPU is less than the price of the RAM that you need to move to an AM5 platform.

My daily driver has 128 GB of ECC. I'm not in the market for this because I splurged on a 5950X when I got it at the beginning of the pandemic, and don't game more than I use VMs and software transcoding, but I've still been interested just a bit.

 

[DRJlaw]

The switch to AM6 is supposed to come with DDR6 right? I take this to mean that either the transition is delayed or they expect DDR6 to be extremely expensive, probably even more than when DDR5 was the shiny new thing.

AMD's more recent socket changes have tended to align with changes in DDR RAM support, since things like PCI-e lane speed can be downgraded in silicon to match the motherboard capability and upgraded (with a different motherboard and trace layout and/or redrivers) to match the CPU capability across the same socket interface (AM4 went from PCI-e 3 to PCI-e 4 support). The DDR standards have tended to have different electrical interfaces in addition to different signalling rates, so there has to be a physical change to the arrangement and number of socket pins used.

 

[CrystalCowboy]

Intel’s next-gen LGA1954 socket will support Nova Lake, Razor Lake, and beyond — finally an Intel socket that outlives its CPUs

Where do they get these ideas?