AMD Sampling Next-Gen Ryzen Desktop "Medusa Ridge," Sees Incremental IPC Upgrade, New cIOD

[john_]

I was getting enthusiastic too, but I'm not sure AMD is willing to give regular desktop users more than 32 threads and erode their Threadripper family.
I hope for the 12-core single CCX CPU though, to replace my aging 8 core 3700X.

They will go 24 cores / 48 threads. They probably will not go the road Intel took with all those E cores, I doubt they want to make their design more complicated than what it is now and have to build something like Intel's Thread Director and have to put more software headaches on their heads to make sure that OS uses the correct cores for the correct tasks etc. What they have to do, they can't avoid it, no matter if this threatens their "low" core Threadripper models - it probably wouldn't anyway, TRs offer more than just more cores - is to at least match thread count to Intel's core count. So If Intel is going to push CPUs with 52 cores (16+32+4), then AMD will push 24 cores / 48 threads models to counter that.

 

[_roman_]

I need a processor where all Nodes are able for the same cpu instruction set, sometimes called cpu flags.

I doubt there will be changes which are am5 socket related.
DRAM standard is set in stone. I did not claim that. AM4 = DDR4 / AM5 = DDR5

I doubt they increase the CPU cache size. I always wanted that ages ago. Higher cache existed with special cpu cases. Intel had a few (intel 5775c or what they called it), amd had a few processors (x3d cpus).

I would accept zen 4 + zen4c (whatever they called the same cores with less cpu cache but same structure) or similar amd processors.

Note: 8 core - 5000 or 8000 series ryzen processors are kinda cheap

 

[Broken Processor]

I'll wait for real world benchmarks when the time comes to much is unknown currently.

 

[Grebber01]

Fu**Y** amd.... when will they finally replace the old 8700G??

 

[DemonicRyzen666]

latency hit, depending on the design, and for once AMD, make 8000MT's speed not feel like worse using 6000MT's..lol..

would be nice if the CCD's had link to each other in stead of relying on going through to the I/O die & back. Two I/O dies could use the same link, but this also extra cost.

 

[TheGeekn°72]

new dual memory controller architecture

I SEE IT ! FINALLY, AFTER ALL THESE YEARS ! THE PROMISED LANDS OF CONSUMER QUAD CHANNEL !
I know it's still dual channel here for Zen6 but COME ON AMD, YOU CAN DO IT, MAKE IT THICKER AND GIVE US QUAD CHANNEL ON ZEN7

Looking forward to it, would be nice if the consumers get a bump in core count again, we have had 16 core 32 thread parts since Ryzen 3000.

I mean, does anyone *really* need more ? the problem right now as it is the undoubtely *STUPID* amount of software that runs single-core/single-thread in this day and age
the still-too-few softwares that do run multi-core run fairly well and if you absolutely need more, Threadripper is there (though I'll be the first to admit it's definitely not a wallet friendly option)
though moving from 8 to 12 core CCDs is definitely nice for the mid-high end, as long as the price impact isn't too high...

Quad channel will required at-least 40% more pads and bigger CPU socket, traces issue can be solved by placing 2 CAMM like slot on back side of PCB where CPU socket is placed, to have shortest route from CPU.

always wondered why there wasn't more things going on backside, mobo mfers experimented putting M.2s backsides a few times but it's not a regular thing (yet ?) and zero cable projects put all the power connectors on the backside of the mobo (obviously), CAMM2 ought to go on the back of mobos, right ?

The big difference won't be just in the memory controller but also the interconnect between the IOD and CCDs. We're going to get the same style interconnect as Strix Halo but clocked MUCH faster.

after a putting the cache below the CCD, better IMC, better interconnect, higher clocks across the board, AMD is cooking holy shit

No rdna 3.5 igpu upgrade on that iod nor an npu?
I have a hard time believing this will be all there is to zen 6

well it's not all there will be to Zen6, I mentioned more of the Zen6 additions/changes above but tbh, I do hope the iGPU gets an upgrade along the IMC rebuild and they manage to sneak in an XDNA2 NPU (or maybe even XDNA3 ? who knows, I haven't heard much about XDNA development...)

dozens of cores on the desktop mean that the half a dozen core CPU would be cheaper (although this is amd we are talking about, they might as well keep charging you 400$ for 8 cores).

it used to be an Intel tactic, I hope AMD won't adopt it...

TRs offer more than just more cores - is to at least match thread count to Intel's core count.

I mean, 128 lanes of PCIe5 and quad/hexa/octachannel (I forgot how many are on current gen TRs) ECC memory is also a reason why you'd want such a platform to which Intel has no answer for besides Xeon that is massively expensive since server oriented instead of workstation

I doubt they increase the CPU cache size. I always wanted that ages ago. Higher cache existed with special cpu cases. Intel had a few (intel 5775c or what they called it), amd had a few processors (x3d cpus).

I mean, they might not increase the cache *per core* but they will likely increase the overall cache along the core count

Note: 8 core - 5000 or 8000 series ryzen processors are kinda cheap

5000 is on Zen3, it's a 2019 design, 6 years later, it's not that relevant anymore, the gap in IPC with Zen5 and soon-to-be Zen6 is huge already
also, 8000 series is a mobile design put in a desktop socket, it's not that surprising that it's cheap since not made to be really all that powerful

Fu**Y** amd.... when will they finally replace the old 8700G??

there's a lot of noise (because of a bunch of BIOS update patch notes) about 9000G coming to desktop apparently, maybe that's it ? always wanted to see at least Strix Point come to desktop and bring a thick iGPU (and honestly, a good NPU wouldn't be too bad either)

would be nice if the CCD's had link to each other in stead of relying on going through to the I/O die & back. Two I/O dies could use the same link, but this also extra cost.

you'd need an IO controller on the CCD itself, kinda redundant tbh, not sure the improvement would be that significant, more interested in the better die interconnect that's also being added to Zen6

 

[Nhonho]

Example of how desktop Ryzens could be.

All components that are on the "main die" in this image must be made on a single die to ensure the fastest possible communication with each other. If any of these "main die" components were placed on a separate die, latencies go to increase significantly and overall performance go to drop significantly, similar to current Intel CPUs.

 

[Denver]

IF and IMC are bottlenecks limiting Zen5's real performance today. So, this is going to be revolutionary at least in terms of bandwidth and latency, solid gains in games and MT tasks.

Games today tend to be increasingly aggressive in multithreading, including shader compilation and initial installation processes, which are also highly multithreaded.

 

[Tek-Check]

I'm looking forward to a single 12 core CCD, also would be interesting if AMD can double the Infinity Fabric speed with dual memory controllers.

New Infinity Fabric should double the speed per single link to 64 Gbps, as EPYC Venice will support PCIe 6.0. IF speed is usually in line or close to currently supported PCIe standard.

 

[Daven]

Fu**Y** amd.... when will they finally replace the old 8700G??

The 8700G came out a year and 3 months later than the 7700X. The 9700X came out in August 2024 so I expect the 10700G (12 cores, 16 CUs) to come out in November 2025.

 

[Hecate91]

New Infinity Fabric should double the speed per single link to 64 Gbps, as EPYC Venice will support PCIe 6.0. IF speed is usually in line or close to currently supported PCIe standard.

More I/O would be nice as well, PCI-e 6.0 means we would get an X970E chipset, though I'd expect PCI-e 6.0 to not be on desktop until Zen 7.

 

[Tek-Check]

They will go 24 cores / 48 threads. They probably will not go the road Intel took with all those E cores, I doubt they want to make their design more complicated than what it is now and have to build something like Intel's Thread Director and have to put more software headaches on their heads to make sure that OS uses the correct cores for the correct tasks etc. What they have to do, they can't avoid it, no matter if this threatens their "low" core Threadripper models - it probably wouldn't anyway, TRs offer more than just more cores - is to at least match thread count to Intel's core count. So If Intel is going to push CPUs with 52 cores (16+32+4), then AMD will push 24 cores / 48 threads models to counter that.

There are probably another 2 LP cores on IOD, so 52 threads in total on the top SKU.

 

[RinSenpai]

Here we go again...

Tell us what desktop applications need dozens of cores other than video encoding by the CPU?

Just play 'Cities: Skylines 2' and you'll see Intel outperforming their respective AMD counterparts despite not having HyperThreading on its P-cores for Core Ultra.
Lots of games don't care about what are E-cores/Zen 5c cores, they just try to use them and that's it, and it works most of the time.
And if we're going to debate about little cores being useless (E-cores AND Zen 5c cores, because they technically all run at way lower frequencies than main cores), then why is AMD selling us 3 and 4-cores CPUs in €1500 laptops in 2025?

 

[Tek-Check]

Example of how desktop Ryzens could be.

Zen6 package will look similar to Strix Halo, with a smaller IOD. That's all to it.

 

[dyonoctis]

Here we go again...

Tell us what desktop applications need dozens of cores other than video encoding by the CPU?

Code compilation, shaders compilation in games engine, after effects, a few 3D render engines who have additional functionality in CPU mode vs GPU mode, other like Renderman who's using both the GPU and CPU to full throttle. The CPU is also used to decode some high quality/raw video formats that aren't accelerated (Prores/red/Blackmagic video files)

Other thing, haven't you noticed that the highest core count CPUs on mainstream platforms also have the best ST performance? AMD and Intel aren't dumb, they know that mixed workload is important for the mainstream platform. They won't sacrifice ST perf for MT. This isn't the Core 2 Quad era. Even in MT, the 9950x boost higher compared to the 9700x.

Abother thing: even if you personally won't have any use for anything above 8 cores, increasing the core count will make your 8 cores CPU cheaper, so you still win.

Last thing, there's also a strong chance that many apps don't scale that well past 8 cores not because they can't, but because 6-8 cores was probably the most common core configuration for a while. A decades ago many apps also didn't scale that well past 4 cores. Zen 1 was better for productivity only in a few select application. Since that era apps have improved in MT.

Example of how desktop Ryzens could be.

All components that are on the "main die" in this image must be made on a single die to ensure the fastest possible communication with each other. If any of these "main die" components were placed on a separate die, latencies go to increase significantly and overall performance go to drop significantly, similar to current Intel CPUs.

Ah yes, and getting rid of the thing that made Ryzen so flexible in the first place : being able to use the same compute die and wafers used for Epyc and threadripper, and having to dedicate a whole wafer allocation just for gamers.

 

[john_]

although this is amd we are talking about, they might as well keep charging you 400$ for 8 cores

The tray version of 8400F(6 cores/12threads), sells for around 90 euros in Greece. The tray version of 8500F(8 cores/16threads) sells for around 150 euros. Full box 6 cores start at around 130 euros and 8 cores at around 250 euros.

 

[_roman_]

It's all fine when you run unoptimised, low end code which relies on many cores. One of many flaws of windows binary software. Legacy code will not bother much if it's a P or E-core as the advanced instructions are never used in the first place. Windows requiring a very old instruction set suddenly for w11 shows how unoptimised the binary code of windows is.
It'S time AVX512 will be a hard requirement for the next windows release with all the other newer cpu instructions.

When I want many cores without any cpu instruction sets I can use a arm platform also.

 

[rv8000]

Quad channel will required at-least 40% more pads and bigger CPU socket, traces issue can be solved by placing 2 CAMM like slot on back side of PCB where CPU socket is placed, to have shortest route from CPU.

This would literally cook DRAM ICs. CAMM doesn’t seem likely to be a thing in the desktop space outside of a (pointless) show piece for news articles.

I’m excited to see a potentially faster fabric and higher memory frequencies + cudimm support. I want to chase some higher frequencies for fun, my IMC doesn’t like 8400 unfortunately

 

[JustBenching]

The tray version of 8400F(6 cores/12threads), sells for around 90 euros in Greece. The tray version of 8500F(8 cores/16threads) sells for around 150 euros. Full box 6 cores start at around 130 euros and 8 cores at around 250 euros.

Those are old cutdown parts man and the prices you are quoting are used, not new. The average price of 8core chips from amd on time of release is 400+$. (5800x, 5800x 3d, 7700x, 9800x, 9800x 3d). Lets not spend any more time arguing about the obvious here.

 

[TheGeekn°72]

All components that are on the "main die" in this image must be made on a single die to ensure the fastest possible communication with each other. If any of these "main die" components were placed on a separate die, latencies go to increase significantly and overall performance go to drop significantly, similar to current Intel CPUs.

I find this design quite interesting tbh

IF and IMC are bottlenecks limiting Zen5's real performance today. So, this is going to be revolutionary at least in terms of bandwidth and latency, solid gains in games and MT tasks.

Games today tend to be increasingly aggressive in multithreading, including shader compilation and initial installation processes, which are also highly multithreaded.

nothing like a STEAM DOWNLOAD to benchmark my CPU and see it revv up to max boost on all core to handle the files lmfao (multigig/s download bandwidth required)

New Infinity Fabric should double the speed per single link to 64 Gbps, as EPYC Venice will support PCIe 6.0. IF speed is usually in line or close to currently supported PCIe standard.

does that mean IF is essentially an interCCD PCIe ?

More I/O would be nice as well, PCI-e 6.0 means we would get an X970E chipset, though I'd expect PCI-e 6.0 to not be on desktop until Zen 7.

yeah no, we "just" got PCIe5, I don't think we'll get consumer grade boards with gen6 soon, datacenters get all the brand new shiny toys after all !

Just play 'Cities: Skylines 2' and you'll see Intel outperforming their respective AMD counterparts despite not having HyperThreading on its P-cores for Core Ultra.

I don't know if that's the best example to take considering the unoptimized shitstorm that it is... I mean, I guess you can bruteforce through sheer compute

Lots of games don't care about what are E-cores/Zen 5c cores, they just try to use them and that's it, and it works most of the time.
And if we're going to debate about little cores being useless (E-cores AND Zen 5c cores, because they technically all run at way lower frequencies than main cores), then why is AMD selling us 3 and 4-cores CPUs in €1500 laptops in 2025?

...what are you talking about ? which laptops ? nevermind the price you're claiming, what mobile CPU did AMD launch in 2025 that's only 3 cores ?

Ah yes, and getting rid of the thing that made Ryzen so flexible in the first place : being able to use the same compute die and wafers used for Epyc and threadripper, and having to dedicate a whole wafer allocation just for gamers.

I mean, it would kinda make sense : the bottom IO die doesn't have to be sourced from TSMC if it has to be cheap, it just has to work and not be power hungry, surely another foundry can provide that, the main die, if on N2X can be made super small/very efficient and the CCDs on the left could be Epyc/TR CCDs, combining production costs while making a super fast central unit and beefy extension CCDs

This would literally cook DRAM ICs.

why ? CAMM draws less power and since it'd be exposed on the backside, just put a passive heatsink

CAMM doesn’t seem likely to be a thing in the desktop space outside of a (pointless) show piece for news articles.

unfortunate, we'd stand to gain from that, I wonder what's "wrong" with desktop CAMM design that they won't commercialize it...

I’m excited to see a potentially faster fabric and higher memory frequencies + cudimm support. I want to chase some higher frequencies for fun, my IMC doesn’t like 8400 unfortunately

Intel CPUs always enjoyed improved performance from high freq memory, I wonder if that will also be true of Ryzen once they can clock just as high and if it'll matter with X3D CPUs (it doesn't right now at 6000-6200MTs)

 

[Onasi]

I mean, it would kinda make sense : the bottom IO die doesn't have to be sourced from TSMC if it has to be cheap, it just has to work and not be power hungry, surely another foundry can provide that, the main die, if on N2X can be made super small/very efficient and the CCDs on the left could be Epyc/TR CCDs, combining production costs while making a super fast central unit and beefy extension CCDs

I mean, if AMD wanted to burn money for no particular reason that WOULD be one way to do it, yes.

 

[john_]

Those are old cutdown parts man and the prices you are quoting are used, not new. The average price of 8core chips from amd on time of release is 400+$. (5800x, 5800x 3d, 7700x, 9800x, 9800x 3d). Lets not spend any more time arguing about the obvious here.

What does it mean "cut down" parts? Is an 9900X irrelevant because it is a "cut down" version of a 9950X? And no, those prices are for new parts, directly from shops with warranty.
If you don't want to spend more time arguing, just don't ignore reality when it doesn't suit your narrative and don't call my post misleading by saying that I am quoting second hand prices. I am quoting real pricing here for new parts. You are quoting MSRP prices of products that are released in a market where other products exist. Obviously when AMD, Intel, Nvidia, EVERYBODY has flooded the market with products, they will not price their new line of products in such a way that will make the older products impossible to sell. Are we going to throw logic in the trash just to create a point?

 

[TheGeekn°72]

I mean, if AMD wanted to burn money for no particular reason that WOULD be one way to do it, yes.

I genuinely don't see what part of this would be burning money : if the main die is made small and self contained, they could put so many on a single wafer and if the CCD orders are grouped with TRs/Epycs as multipurpose CCDs, it would enjoy the economy of scale that comes with it instead of ordering specifically desktop-oriented CCDs and datacenter CCDs separately
and I doubt TSMC would make the IO die for cheap considering they're very much focused on the bleeding edge nodes and lower value lithography are almost dime a dozen in so many other types of tech, AMD could review a few foundries and pick a few one to make a cheap, effective IO die for the applications described in the picture

 

[rv8000]

I find this design quite interesting tbh

nothing like a STEAM DOWNLOAD to benchmark my CPU and see it revv up to max boost on all core to handle the files lmfao (multigig/s download bandwidth required)

does that mean IF is essentially an interCCD PCIe ?

yeah no, we "just" got PCIe5, I don't think we'll get consumer grade boards with gen6 soon, datacenters get all the brand new shiny toys after all !

I don't know if that's the best example to take considering the unoptimized shitstorm that it is... I mean, I guess you can bruteforce through sheer compute

...what are you talking about ? which laptops ? nevermind the price you're claiming, what mobile CPU did AMD launch in 2025 that's only 3 cores ?

I mean, it would kinda make sense : the bottom IO die doesn't have to be sourced from TSMC if it has to be cheap, it just has to work and not be power hungry, surely another foundry can provide that, the main die, if on N2X can be made super small/very efficient and the CCDs on the left could be Epyc/TR CCDs, combining production costs while making a super fast central unit and beefy extension CCDs

why ? CAMM draws less power and since it'd be exposed on the backside, just put a passive heatsink

unfortunate, we'd stand to gain from that, I wonder what's "wrong" with desktop CAMM design that they won't commercialize it...

Intel CPUs always enjoyed improved performance from high freq memory, I wonder if that will also be true of Ryzen once they can clock just as high and if it'll matter with X3D CPUs (it doesn't right now at 6000-6200MTs)

You can’t place the camm socket directly behind the CPU as they suggested, CPUs dump a considerable amount of heat behind the socket. That plus no airflow would cook the ICs which reduces frequency potential, and ultimately would provide no benefit if I had to guess.

There is ZERO benefit to CAMM on desktops, CU-DIMMs match and exceed frequencies of camm modules without changing pcb/motherboard design. They offer nothing.

 

[TheGeekn°72]

You can’t place the camm socket directly behind the CPU as they suggested, CPUs dump a considerable amount of heat behind the socket.

How much of the powerdraw turned heat makes it to the backside exactly ?

That plus no airflow would cook the ICs which reduces frequency potential, and ultimately would provide no benefit if I had to guess.

current RAM are still shipped with either no heat spreader or just fancy looking ones, I think a small slab of aluminium would be more than enough to help heatsink a few degrees away from the ICs just through sheer radiation... plus, backside mobos don't have *no airflow*, it has *a little* which should still be enough to make a passive cooler of sort work well enough I reckon
an NVMe consumes max between 7W (gen4) to 12W (gen5) and you probably can figure out the size of the coolers we can find for those, just make it the same for CAMM2, just less tall and more spread out, even hanging off the module's board, that should be fine

There is ZERO benefit to CAMM on desktops, CU-DIMMs match and exceed frequencies of camm modules without changing pcb/motherboard design. They offer nothing.

I mean, low profile memory displaced to the back that sips power compared to DIMM, I'd buy
plus it would mean zero cooler incompatibilities, more space for bigger air coolers or a bulky water reservoir