More AMD Ryzen 9000 "Zen 5" Desktop Processor Details Emerge

[TumbleGeorge]

ZEN 5 main series Q4 2024 as usual. Probably in same month with Windows 12.

 

[Chaitanya]

What is proper integration to you?

No need for add-on card with propietary cable connected to motherboard, right now there are couple of high end X670 boards with USB 4 on board but they are far too expensive for me.

 

[TheLostSwede]

No need for add-on card with propietary cable connected to motherboard, right now there are couple of high end X670 boards with USB 4 on board but they are far too expensive for me.

Well, your wish will be fulfilled, I just hope you know what you're wishing for.
It will still be the same as integrated Thunderbolt on motherboards now, i.e. that you need cables from the graphics card to the I/O shield inputs to run graphics over the USB4 interface.

 

[rv8000]

Huuuuh?

DDR5 is already quad channel (4x 32bit instead of 2x 64bit) and its already been shown multiple times by different people that going over 6000 gets you almost nothing for Zen4.

Zen4 is much more sensitive to latency than bandwidth. That is why once people get to 6000 they focus on getting the timings as tight as possible. Going any faster usually gets you almost nothing and getting tighter timings is incredibly difficult so its not worth it.

Going big on cache also usually doesn't get you huge gains across the board (as has been shown with 7800x3D) so I'm not sure where you're getting the idea its 'cache starved' from. Cache does help with some games a fair degree but even with gaming its not providing the gigantic 100%+ gains across the board to justify 'cache starved'.


The x3D chips hardly get any benefit in most games from going over 6000.

Some people have gotten 7800x3D's to 7800 or 8000 DDR5 and benched it and gains are super minimal. The giant cache makes memory overclocking mostly moot.

This is a good thing!

You can buy the cheap RAM, or use what you already got, instead. OC'ing RAM is pretty pricey and even on Intel getting things stable at high clocks with low timings is still incredibly difficult. Many people have had to lower their standards for what they consider 'stable' to get DDR5 8000 working on either vendor.

Single CCD Zen4 chips gain little to nothing going past 6000 (staying in 1:1), dual CCD chips are a different story when it comes to any software that can leverage the bandwidth increase.

Beginning at around 7600 c34-c36 to 7800 c36-c38, dual CCD chips can match gaming performance achievable by 6000-6400 setups and get the added bandwidth benefits in any other software than can take advantage of it (primary trade offs being memory training and finding a mobo that can actually run it boot to boot).

Sadly there are few viable motherboards that can run 2:1 frequencies above 7600 reliably. The Gene is pretty much your only option due to AIBs refusal to make 2dimm boards available (no tachyon release, no apex, no unify).

 

[Chaitanya]

Well, your wish will be fulfilled, I just hope you know what you're wishing for.
It will still be the same as integrated Thunderbolt on motherboards now, i.e. that you need cables from the graphics card to the I/O shield inputs to run graphics over the USB4 interface.

I dont need USB 4 for the hubs with video functionality, just need it for CFx-B card reader so I can do without passing the video through controller.

Single CCD Zen4 chips gain little to nothing going past 6000 (staying in 1:1), dual CCD chips are a different story when it comes to any software that can leverage the bandwidth increase.

Beginning at around 7600 c34-c36 to 7800 c36-c38, dual CCD chips can match gaming performance achievable by 6000-6400 setups and get the added bandwidth benefits in any other software than can take advantage of it (primary trade offs being memory training and finding a mobo that can actually run it boot to boot).

Sadly there are few viable motherboards that can run 2:1 frequencies above 7600 reliably. The Gene is pretty much your only option due to AIBs refusal to make 2dimm boards available (no tachyon release, no apex, no unify).

There is a B650 Tachyon.

 

[rv8000]

I dont need USB 4 for the hubs with video functionality, just need it for CFx-B card reader so I can do without passing the video through controller.


There is a B650 Tachyon.

You cannot purchase this motherboard.

 

[pzqking]

Standards of AMD YES this time:
1.16C with 64m L3 xpu to be a 6GHZ monster,with PPT lower than 250 W,no more than 300W,wind cooling is essential for users require stability for 5 year to submit a official consume in China.
2.APUs are not urge for too much cores,but integrated graphics must be strong enough,even a 4C8T APU is acceptable.A 4C8T 5GHz all zen5 APU with 12CU,which starts with lower than 149USD seems perfect.Cheap APUs absolutely will be able to gain some favor,most of CPUs today are too much for common gamers.
3.X3D version needs to be lower to 6C12T.Most important is the price.We need G3D,at least something like 5600x3d.
4.Upgrade that d**n fclk.APUs needs a better memory frequency provide,even 6400mhz cannot fulfill 7840HS,it needs more.APUs with integrated high speed memory or seperated Gmem for integrated graphics seems nice,but it shouldn't too expensive.
I'm counting on it to retire my R5 5600 for already 4-year-used.Wish this time it can get something really inspring.

 

[dyonoctis]

Well, your wish will be fulfilled, I just hope you know what you're wishing for.
It will still be the same as integrated Thunderbolt on motherboards now, i.e. that you need cables from the graphics card to the I/O shield inputs to run graphics over the USB4 interface.

For what its worth, Win 11 support dGPU pass-through through the iGPU even on the desktop. I couldn't detect any latency worth mentioning, but I'm not a pro gamer.

 

[AnotherReader]

Standards of AMD YES this time:
1.16C with 64m L3 xpu to be a 6GHZ monster,with PPT lower than 250 W,no more than 300W,wind cooling is essential for users require stability for 5 year to submit a official consume in China.
2.APUs are not urge for too much cores,but integrated graphics must be strong enough,even a 4C8T APU is acceptable.A 4C8T 5GHz all zen5 APU with 12CU,which starts with lower than 149USD seems perfect.Cheap APUs absolutely will be able to gain some favor,most of CPUs today are too much for common gamers.
3.X3D version needs to be lower to 6C12T.Most important is the price.We need G3D,at least something like 5600x3d.
4.Upgrade that d**n fclk.APUs needs a better memory frequency provide,even 6400mhz cannot fulfill 7840HS,it needs more.APUs with integrated high speed memory or seperated Gmem for integrated graphics seems nice,but it shouldn't too expensive.
I'm counting on it to retire my R5 5600 for already 4-year-used.Wish this time it can get something really inspring.

APUs can use memory as fast as LPDDR5 7500. The link between the memory controller and the CPU isn't fast enough to utilize all of that, but the link to the GPU is fast enough to utilize an unrealistic DDR5 11200.

The GPU has four 32B/cycle ports to fabric, letting it get enough memory bandwidth even at low fabric clock.

 

[Chaitanya]

You cannot purchase this motherboard.

I think thats a niche SKU GB created specifically for either in house or pro overclockers. Not sure why they havent made it available for sale so far.

 

[efikkan]

Let's hope for a smooth update on existing motherboards then.

How about a more affordable (but decent) motherboard selection? Things still look a bit pricy to me, compared to how AMD platforms have been in the past (and compared to LGA1700).

DDR 6400? Is this a joke?

If you need more bandwidth for your workload, then there is always 4/8C Threadrippers and Xeon-Ws.

The design is already cache-starved, and it's well on its way to be memory bandwidth starved too. Dual channel is becoming a joke on AMD. I guess this is AMD dropping their pants thinking they will be superior to Intel, I'm not so confident.

Do you mean L2 or L3?
More L2 could be useful, but more L3 is mostly useful for poorly optimized code.
But cache efficiency always comes down the specific characteristics of the architecture, and unless you have in-depth knowledge of the Zen 5 design and performance, you couldn't make much of a qualified assesment. Cache has increased and decreased between generations before, and one cache configuration may favor latency, while others favor hitrate or bandwidth. If for instance a new architecture have a very different design, a differently configured cache might be beneficial, even if it's smaller than strongly opinioned forum warriors might want.

 

[TheLostSwede]

I think thats a niche SKU GB created specifically for either in house or pro overclockers. Not sure why they havent made it available for sale so far.

I know why, but I can't share it publicly.

 

[phubar]

dual CCD chips are a different story when it comes to any software that can leverage the bandwidth increase.

Nah its not true even then:

KLEVV CRAS V RGB DDR5-6400 CL32 2x 16 GB Review

KLEVV is at it again and brings us another CRAS product. What we have here today is the CRAS V RGB DDR5-6400 memory kit marketed towards PC enthusiasts and gamers alike. Follow along as we test this memory kit and see how it stacks up to the competition!

www.techpowerup.com

You'll see a nice boost in synthetic benches but in real world work loads there is hardly any difference going over 6000 for Zen4 dual CCD or 1 CCD.

I think you're mistaking the increase from uncore/IF bus or latency improvements from clocking the RAM high with tight (c34-36 + tight secondary/tertiaries at 7800 is FAIRLY tight for AMD right now, Intel too) timings for bandwidth gains.

The caches on these chips are all quite large, low latency, and fast so its not surprising that they're not all that limited by bandwidth.

 

[tugrul_SIMD]

Huuuuh?

DDR5 is already quad channel (4x 32bit instead of 2x 64bit) and its already been shown multiple times by different people that going over 6000 gets you almost nothing for Zen4.

Zen4 is much more sensitive to latency than bandwidth. That is why once people get to 6000 they focus on getting the timings as tight as possible. Going any faster usually gets you almost nothing and getting tighter timings is incredibly difficult so its not worth it.

Going big on cache also usually doesn't get you huge gains across the board (as has been shown with 7800x3D) so I'm not sure where you're getting the idea its 'cache starved' from. Cache does help with some games a fair degree but even with gaming its not providing the gigantic 100%+ gains across the board to justify 'cache starved'.


The x3D chips hardly get any benefit in most games from going over 6000.

Some people have gotten 7800x3D's to 7800 or 8000 DDR5 and benched it and gains are super minimal. The giant cache makes memory overclocking mostly moot.

This is a good thing!

You can buy the cheap RAM, or use what you already got, instead. OC'ing RAM is pretty pricey and even on Intel getting things stable at high clocks with low timings is still incredibly difficult. Many people have had to lower their standards for what they consider 'stable' to get DDR5 8000 working on either vendor.

When data is copied to graphics card, it is mapped from RAM to VRAM. Then it uses RAM banwdith for reads. In the same time, CPU can write to RAM from its cache. Then RAM will require a higher bandwidth to not lag much. High RAM bandwidth helps when doing physics simulations on CPU. Many games do physics sim on CPU. Because either developers can't spend time on doing it on GPU or they optimize more for low-end GPU players.

If data does not directly go from RAM to VRAM, but through CPU infinity fabric, then its not a good design as infinity fabric is very limited (32GB/s write, 64GB/s read, per ccd).

 

[phubar]

When data is copied to graphics card, it is mapped from RAM to VRAM. Then it uses RAM banwdith for reads.

OK sure this is broadly true BUT if Zen4 was bandwidth limited, in any CCD config, then we should be seeing LARGE gains with increases in bandwidth in games or other apps that use lots of bandwidth.

But we're not.

Gains are extremely minimal past 6000. Even all the way to 8000.

Unless of course you can also minimize the timings and crank the uncore/IF bus clocks....but that primarily effects latency NOT bandwidth.

BZ has a vid on this, for both AMD and Intel, showing virtually 0 gains in performance with overclocking DDR5 if you don't also tighten timings for instance for this reason.

The extra bandwidth can certainly help the iGPU on the Zen4/5 APU's of course but iGPU's aren't Zen4/5 CPU cores!

If data does not directly go from RAM to VRAM, but through CPU infinity fabric, then its not a good design as infinity fabric is very limited (32GB/s write, 64GB/s read, per ccd).

Even top end video cards are barely effected by PCIe 3 vs PCIe 4 (which is double the bandwidth) or PCIe 5 (double 4) so your example here for 'IF bus/bandwidth is limited' isn't making any sense.

If anything games are typically more limited by the latency of the PCIe bus (which is high) not the IF bus/system RAM bandwidth!

The stuff that will be limited by the IF bus or bandwidth on AM5 is the stuff that is bandwidth limited pretty much all the time on all platforms so its irrelevant to talk about it in a practical sense here. Especially if you want to talk about common desktop apps like gaming.

 

[Nordic]

AMD MO is to squeeze the early adopters and then drop prices. The X3D and non X products are so much better than the early releases.

What hardware company doesn't have this MO? Corporations will extract as much value as they can.

I know why, but I can't share it publicly.

Then why say anything at all? All you are doing is teasing us

 

[TheLostSwede]

Then why say anything at all? All you are doing is teasing us

Sorry, let's just call it business related for the company in question and maybe someone here can figure it out.

 

[rv8000]

Nah its not true even then:

KLEVV CRAS V RGB DDR5-6400 CL32 2x 16 GB Review

KLEVV is at it again and brings us another CRAS product. What we have here today is the CRAS V RGB DDR5-6400 memory kit marketed towards PC enthusiasts and gamers alike. Follow along as we test this memory kit and see how it stacks up to the competition!

www.techpowerup.com

You'll see a nice boost in synthetic benches but in real world work loads there is hardly any difference going over 6000 for Zen4 dual CCD or 1 CCD.

I think you're mistaking the increase from uncore/IF bus or latency improvements from clocking the RAM high with tight (c34-36 + tight secondary/tertiaries at 7800 is FAIRLY tight for AMD right now, Intel too) timings for bandwidth gains.

The caches on these chips are all quite large, low latency, and fast so its not surprising that they're not all that limited by bandwidth.

It is though, single CCD Zen4 parts are hard limited on bandwidth. You could theoretically run DDR5 8000 c30 and still get latency improvements on a 7700X/7800X3D but no benefit to bandwidth which would be well in excess above 100gb/s, and youd be stuck around 70gb/s. Whereas as a dual CCD Zen4 part won’t have the same limitiation.

Synthetics aside, there is absolutely a benefit to increased bandwidth which would be application dependent.

I’ve spent hundreds of hours testing ddr5 6000-8000 on my 7900X3D. Dual CCD is a different ball game when it comes to memory overclocking benefits.

 

[phubar]

It is though,

I can litterally find no one with benches posted to show this though.

I can however easily find vids and benches of 7950's getting tested in games at 6400 or 7200 or 7800 or 8000 and usually getting hardly any benefit over 6000 at 1080p.*

I can also find benches of 7950's getting tested at 6400 here at this site, I linked you one, and showing little to no improvement over 6000 in nearly all real world apps.

If Zen4 was really all that bandwidth limited, in either CCD config, then benches showing LARGE gains (ie. 20, 50, 100%+) should be typical and easy to find. Instead all that is out there are typically very minor 1% here 2% there gains for 6400+ speeds, outside of a couple of interesting outliers, and of course some synth benches but they're synth benches so who cares.

*note that if they also minimize the timings then yes you start to see some gains BUUUT adjusting timings doesn't effect bandwidth, it effects latency

 

[AnotherReader]

It is though, single CCD Zen4 parts are hard limited on bandwidth. You could theoretically run DDR5 8000 c30 and still get latency improvements on a 7700X/7800X3D but no benefit to bandwidth which would be well in excess above 100gb/s, and youd be stuck around 70gb/s. Whereas as a dual CCD Zen4 part won’t have the same limitiation.

Synthetics aside, there is absolutely a benefit to increased bandwidth which would be application dependent.

I’ve spent hundreds of hours testing ddr5 6000-8000 on my 7900X3D. Dual CCD is a different ball game when it comes to memory overclocking benefits.

I don't think there's any RAM out there that can do DDR5 8000 CL30. TechPowerUp reviewed a DDR5 8000 kit that was rated at CL38. AMD also needs to improve the DDR5 controller to match the efficiency of their older DDR4 controller. The memory controller in the IO die can only utilize about 81% of the theoretical bandwidth of DDR5.

Contrast this with the memory controller in the IO die for Zen 2 which achieves over 92% of theoretical bandwidth.

Contrast this with the 7840HS whose memory controller can utilize 91% of theoretical bandwidth from a higher latency setup (DDR5 5600 CL46).

 

[R-T-B]

Do you know what else would be cool? If I had a unicorn.

No way, that would be too much glitter-poop to cleanup and you know it.

 

[rv8000]

I can litterally find no one with benches posted to show this though.

I can however easily find vids and benches of 7950's getting tested in games at 6400 or 7200 or 7800 or 8000 and usually getting hardly any benefit over 6000 at 1080p.*

I can also find benches of 7950's getting tested at 6400 here at this site, I linked you one, and showing little to no improvement over 6000 in nearly all real world apps.

If Zen4 was really all that bandwidth limited, in either CCD config, then benches showing LARGE gains (ie. 20, 50, 100%+) should be typical and easy to find. Instead all that is out there are typically very minor 1% here 2% there gains for 6400+ speeds, outside of a couple of interesting outliers, and of course some synth benches but they're synth benches so who cares.

*note that if they also minimize the timings then yes you start to see some gains BUUUT adjusting timings doesn't effect bandwidth, it effects latency

Go over to the AM5 memory overclocking thread on OCN, theres been plenty discussion and benchmark evidence outlining the CCD configuration limitations; tldr there is little to no point running ddr5 8000 on a 7800X3D, whereas there is/can be on a 7950X/X3D.

Games are probably the LEAST applicable piece of software here, and not really at all what I’m talking about (as I also exclusively stated “application dependent”).

 

[phints]

Very interested in doing a 9800X3D build timing will be good for me, but need to hear AMD has resolved the long boot time issue Ryzen 7000 has.

 

[Jism]

First gen Ryzens came with 2 cores per CCD - they are at 8 now.

 

[Launcestonian]

Oh goody, looking forward to new toys to play with come middle of this year & I don't need to upgrade any of my motherboards either!