News AMD's AM5 Platform Launches With Only DDR5 Support for Ryzen 7000, Dual-Chipset Design

[wifiburger]

AMD's vcache could well reduce effects of ddr5's higher latencies. However, this also means it will benefit less from faster ddr5 as well.

5800x3d memory latency is worst vs 5800x and that's because the bigger the cache the more cycles are required to get date in/out.

 

[cryoburner]

No problem. AMD will continue to sell Zen3 cpus for ddr4 modules. Those who are willing to pay premium for GPU and CPU are not afraid of DDR5 pricing!
But lets first see how pricing in general develop in this year...

Yeah, I wouldn't be surprised if AMD primarily targeted the high-end with Zen4 initially, while continuing to sell Zen3 on AM4 for more budget to mid-range systems, much like we saw with the Zen3 launch. That may be a part of why they held off on releasing value-oriented Ryzen 5000 CPUs until now, almost a year and a half after the higher-end models came out, as they may plan to keep them around for a while. It might be more than a year from now before they start offering sub-$300 processors for AM5, by which point DDR5 pricing may be down to a much more manageable level where it could make more sense in mid-range systems.

This might also help with availability, as AMD has limited manufacturing capacity. With Zen3, even when restricting their initial lineup to higher-end parts, availability was a problem for a number of months following launch due to high demand. This wasn't as much of an issue when AMD's fastest parts were slightly behind Intel's, at least in terms of lightly-threaded performance, but industry-leading performance can push demand to a level they seem to have trouble supporting. If the 7000-series manages to outperform Raptor Lake, or at least outperform it at certain tasks, like gaming, then a similar situation could occur, and that would only be made worse if they tried to spread their chiplets across a full lineup.

 

[InvalidError]

adopting a new MB, cpu, & DDR5 prices at once limits the userbase.

New socket to correct AM4's shortcomings and accommodate new standards, so you were never going to avoid needing a new motherboard no matter what. Nobody is forcing you to upgrade to Zen 4 out of the gate either. You can always wait it out a year or two, let AMD, its board partners and DRAM manufacturers figure out all of the platform and logistics issues before buying in. I prefer buying at the end of a major platform cycle where most significant bugs have most likely been sorted out than at the beginning where most of the bugs are yet to be discovered.

Bleeding-edge tech usually isn't for the budget-sensitive people, especially in the recent age of mid-range boards costing $70-100 more than they used to due to much tighter board quality requirements for PCIe 4.0/5.0, USB 3-gen2(x2)/4, TB3/4, 3200+MT/s memory busses, tighter CPU VRM requirements from modern CPUs that can boost to 200+W, etc.

If the cost of buying into all-new stuff bothers you, then buy into current-gen stuff good enough to skip the next few generations. Skipping the awkward transition phase from DDR4 to DDR5 is another reason I chose to get an i5-11400 last year.

 

[drivinfast247]

If we make it that far, fall '23 will be nice. AM5, DDR5 and Win11 should have a majority of the bugs worked out and hopefully at decent prices. But I won't hold my breath.

 

[mdd1963]

And this is the same debate we had 6 years ago with DDR3 vs DDR4. Servers benefit greatly, desktops don't.

To be fair, we can't (or should not) ignore the additional 10-20% gains in average FPS in quite a few games upon using DDR5-6400...

 

[ThatMouse]

No one will be buying it this year. 5900X was over $550 when it came out and it took over 6 months to even find them. Look at it now.,

 

[jgraham11]

Considering Ryzen chips are chiplets, it would be great if Zen4 was available on both AM4 and AM5 sockets because the io chiplet would dictate the memory controller. Wouldn't that just destroy this argument and AM4 would continue to live!

 

[Kamen Rider Blade]

As long as Zen 3 / AM4 / DDR4 remain available in the channel while Zen 4 / AM5 / DDR5 sorts itself out, everything should be fine in the long term.

This is just DDR5 early adopter tax.

 

[InvalidError]

Considering Ryzen chips are chiplets, it would be great if Zen4 was available on both AM4 and AM5 sockets because the io chiplet would dictate the memory controller. Wouldn't that just destroy this argument and AM4 would continue to live!

AMD had a hard time getting Zen 2 to work on AM4 (guessing it has to do with the pinout causing layout clashes between Zen 1/1+ chips and Zen 2 chiplets which required a few extra substrate layers to work around of) and almost ended AM4 there. I doubt it wants a rehash of all the backlash it got from heavily restricting official Zen 3 compatibility to make things easier on itself. Easiest way to achieve that is sticking to Zen 4 being AM5 only.

This is just DDR5 early adopter tax.

As it has always been. Though in previous DRAM transitions, server and HEDT sales quietly bit that bullet first about a year ahead of mainstream, lowering the mainstream pain quite considerably. This time around, every market segment wants the new stuff at the same time, which compounds the pain.

 

[Alvar "Miles" Udell]

To be fair, we can't (or should not) ignore the additional 10-20% gains in average FPS in quite a few games upon using DDR5-6400...

Basis and source for that claim?

When TH published their DDR5 vs DDR4 test back in November, the difference between DDR4-3200 and DDR5-6000 was marginal (3%) based on 7 games. TechPowerUp's review of the same month shows an even narrower gap of less than 1% between DDR4-3600 CL16 and DDR5-6000 CL36 in their 10 game test.

So yes, for the mainstream market and I would go so far as to include the lower end of the enthusiast market, there's every reason to ignore DDR5 especially when the performance differences shrink to zero as you approach 3840x2160.

 

[George Stamper]

Our sources have confirmed that AMD's AM5 platform for the 5nm Ryzen 7000 'Raphael' CPUs will only support DDR5 at launch.

AMD's AM5 Platform Launches With Only DDR5 Support for Ryzen 7000, Dual-Chipset Design : Read more

Nice Product

 

[saltweaver]

So yes, for the mainstream market and I would go so far as to include the lower end of the enthusiast market, there's every reason to ignore DDR5 especially when the performance differences shrink to zero as you approach 3840x2160.

Shouldn't game engines be rewritten in a memory efficient way to get bigger performance gap between DDR4 vs. DDR5?

 

[escksu]

5800x3d memory latency is worst vs 5800x and that's because the bigger the cache the more cycles are required to get date in/out.

Not alot worse... AMD managed to keep the latency increase to a bare min.

 

[InvalidError]

Shouldn't game engines be rewritten in a memory efficient way to get bigger performance gap between DDR4 vs. DDR5?

If your game engine or any other code's focus is performance, first thing to do would be designing around cache sizes to minimize dependence on system memory wherever possible. That would explain why most games show minimal gains from running on anything faster than DDR4-3600-16.

I think it was HWUB who simulated an i5 with 25MB and 30MB of L3$ by matching clocks and disabling cores on i7s and i9s where they concluded that a hypothetical i5 with 30MB of L3$ would perform the same as the i9 in most games. Put otherwise: most of the i9's gaming performance gains can be directly attributed to its larger L3$.

AMD popularized 32MB L3$ size, game developers responded by optimizing their newer engines around having 20-30MB of L3$ available and now CPUs with less than 24MB of L3$ are limping along in those.

 

[RareAir23]

Interesting piece about what's to come for people in the early days of socket AM5. The DDR5 memory market is indeed quite expensive and for those who gave forecasts about greater availability and lower price in earlier posts? I second what you said. As for me and where I stand personally? I just finished my upgrade to a Ryzen 7 5800X CPU just last fall. I still think I'm sitting pretty with the current specs of my build where it stands so I'll hold off for a while. Out!

 

[gruffi]

It's the right decision to support only DDR5. It makes platform validation a lot easier. Which should also reduce platform costs. And it will improve the longevity of the platform.

Zen 4 will start with the premium models anyway. You won't get a 100 EUR/USD model at launch. So, people who buy Zen 4 will also be willing to spend more in new memory. That wasn't a big deal when the market switched from DDR3 to DDR4. And it won't be a big deal with DDR5. Especially because DDR5 prices are constantly going down. We will see significant improvements with DDR5 in the upcoming months. For low cost systems AMD will continue to offer AM4 and Zen 3 based models. Which are still fine.

 

[abufrejoval]

I am a bit lost on whether Raphael will be an APU-only design or Zen3 style chiplets (or both again).

I would still assume that AMD just can't/won't do twenty different die variants, since they're not Intel.

Their APU chips covered an awful lot of ground rather well with Cezanne, I've played with it as 5800U at 15/28 Watts to a 5700G at 65 Watts and will admit that it's hard to get 2x benefit from my 16-core 5950X with real-life use cases. And for those who really need that, someone will pay EPYCally.

Beyond a certain number of cores your (CPU) code base is less likely to be nicely threaded (or it is already vectored), so memory access patterns from decoupled code become so spread out you'll need more independent channels (assuming AM5 will still be limited to two channels for economy) or much bigger caches (5800X3D) to avoid stalling on RAM access. The bandwidth advantage of DDR5 won't translate to better random access performance.

I'm not sure we'll actually see a 16-core AM5 chip when they could use the area for things that sell in higher volumes. E.g. to put an extra set of DDR5 channels for the iGPU frame buffer there on the die carrier somewhat Apple M1 style while still maintaining two external channels for soldered RAM or DIMMs. That way they could get a total of 4 channels of RAM for iGPU use, while two channels would go off die carrier for normal code/data expandability.

Depending on just how dense/big such DDR5 RAM stacks might be, you might even be able to expand this again Apple style to a dual APU on single die-carrier system, with only one set of external RAM and I/O going off die carrier to maintain the AM5 socket constraints. There is a lot of surface area on the AM5 socket and now some extra Wattage, too. I count on AMD to surprise us with something smart.

Intel has always been far too comfortable throwing in lots of die area without monetization (they never really charged for the iGPU and felt much too comfortable about deactivating cores and blocks) and AMD's kicker with Zen has been to make every transistor count. To include a backward oriented DDR4 RAM interface as a migration compromise is just too much of a waste in transistors and AM5 socket specs when that is designed for long term.

The APU design isn't driven by the desktop, even if it might be driving the vast majority of desktops. It's driven by notebooks and there the power consumption vs. bandwidth optimizations of the new RAM generation, even if that won't be cheapest for a couple of months. And on laptops the DDR5 DIMM price penalty may not be nearly as big, because specialty packages with volume contracts may drive that down. And if the DDR5 bandwidth enables everything 3D you need in FHD on the iGPU, that price bump just won't be felt there.

I too am a PC or desktop guy, but even if the PC was the original technology driver for both ISA servers and notebooks, the split between the two has become so large, the desktop on its own is no longer big enough to justify its own technology, but will have to decide between which of the other two it's build on.

And with 8 cores at 5GHz as the new base line, that's a lot of desktop for notebook tech.

 

[abufrejoval]

On the dual chipset design:

With an extra four lanes of Infinity fabric available from the socket, adding another "South bridge" becomes an easy 'LEGO' option, that might appeal to buyers of luxury style main boards, especially since (affordable) stand-alone PCIe switch chips have pretty much disappeared.

With PCIe v5 the issue that you can't allocate based on bandwidth but are tied to lanes becomes much bigger and therefore the need for switches to provide some flexibility. It's been pretty painful giving 4 PCIe lanes to a 10Gbase-T Ethernet card when a single lane of PCIe v4 should do it (and Marvell finally seems to sell the chips).

I also wonder if it doesn't make more sense to pipe the Infinity Fabric through a "PCIe v5 x4" slot and put the 'chipset' switch chip there along with say 4 M.2 slots and some USB4/TB ports out back.

And why not do that for both 'chipsets' while we're at it ;-)

More flexibility and on-demand expansion as well as a more compact design and easier cooling.

But then issues with validation, support, mixing vendors and recycling being bad for business...

 

[KananX]

Great, no chaos and a straight path of DDR, I don’t like the Intel approach anyway. The only downside is that premium DDR5 could be prohibitively expensive and that’s it. The question is how much the speed difference will be between low end and medium to high end DDR5 kits. That will decide if it’s price effective or not.

 

[InvalidError]

The only downside is that premium DDR5 could be prohibitively expensive and that’s it.

Even low-end DDR5 currently costs about twice as much as decent DDR4 while often causing worse performance due to increased latencies.

Performance gains from faster (and far more expensive) DDR5 are slim to none in most cases, which is a horrible proposition for performance per dollar. DDR5 is currently a no-go for people for whom cost-efficiency is a concern. We're likely still 15-18 months away from DDR5 making sense for the mainstream.

The first generation going into a new memory standard is practically never worth the massive early adopter tax.

 

[KananX]

Performance gains from faster (and far more expensive) DDR5 are slim to none in most cases, which is a horrible proposition for performance per dollar. DDR5 is currently a no-go for people for whom cost-efficiency is a concern. We're likely still 15-18 months away from DDR5 making sense for the mainstream.

It is absolutely unknown what the performance gains are for DDR5 in AMD platforms. AMD and Intel CPUs aren’t identical, maybe they were comparable 12-15 years ago.

The first generation going into a new memory standard is practically never worth the massive early adopter tax.

Depends on how good Zen 4 is and again, how DDR5 prices are and how it will perform. I suspect Intel won’t be a better deal, I bet Zen 4 will be faster and Intel mainboards are more expensive, which slims down any money you could save on DDR4 compared to Ryzen 7000 with DDR5.

In the end it’s a big “we will see”.

 

[InvalidError]

It is absolutely unknown what the performance gains are for DDR5 in AMD platforms.

Faster memory mostly comes into play when all caches miss. Since AMD will have the cache advantage at least until Raptor Lake launches, I'd expect Zen 4 to be even less memory-sensitive than Alder Lake provided Zen 4 has enough IF : DRAM clock flexibility to allow people to run fast IF (~1.9GHz on Zen 3) without locking themselves into premium DDR5 bins.

 

[KananX]

Faster memory mostly comes into play when all caches miss. Since AMD will have the cache advantage at least until Raptor Lake launches, I'd expect Zen 4 to be even less memory-sensitive than Alder Lake provided Zen 4 has enough IFRAM clock flexibility to allow people to run fast IF (~1.9GHz on Zen 3) without locking themselves into premium DDR5 bins.

If it’s like that (if), you can just buy the cheapest DDR5 kit, and then it’s not that expensive compared to DDR4, manageable.

 

[Charogne]

I love the "ddr5 cost too much at 256$", says the gamers playing on their 2000$+ covid priced GPU...

 

[giorgiog]

Well, that's going to make me choose a 12700k or a 13700k (upgrading from a 1950X.) There's just no reason to pay the DDR5 premium.