Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[biostud]

Who said the current IOD was unable to work with more than two chiplets ?
If AMD planned for more than two chiplets, they probably built it already.

If they didn't, it for good reasons, meaning they either don't need more than two chiplets or the chiplet will change anyway with zen5 or something along those lines...

I think the market for personal computers with more than 16 "Performance" cores and a limitation af 128GB memory is quite small. If you have work that can use more cores, it can probably also use more memory and more memory bandwidth etc. -> Threadripper

For that little niche of people who would be in the market for the less expensive Threadrippers, I can understand the frustration that they are going the way of the dodo, but from a business standpoint I think it makes perfectly sense for AMD.

 

[Thibsie]

I think the market for personal computers with more than 16 "Performance" cores and a limitation af 128GB memory is quite small. If you have work that can use more cores, it can probably also use more memory and more memory bandwidth etc. -> Threadripper

For that little niche of people who would be in the market for the less expensive Threadrippers, I can understand the frustration that they are going the way of the dodo, but from a business standpoint I think it makes perfectly sense for AMD.

Oh I completely agree, I was merely considering the technical point of view.

They may feel they to compete with Intel with number of cores though. I don't care a iota but marketing has its weird ways...

 

[Vope45]

A little birdie told me lower skus will have a hard time competing with raptor, higher end skus when pushed will win and trade blows in some cpu intensive games. Average single threaded performance uplift is 25%, don't know about multi. Apparently amd worked hard to scale the core to achieve ultra high clock. Heat will be a problem when really pushed. Raptor will get a beat down from v-cache version, only 8 cores and 12 cores will receive it (??). Little birdie = co worker in a firm which used to work with nvidia.

 

[biostud]

A little birdie told me lower skus will have a hard time competing with raptor, higher end skus when pushed will win and trade blows in some cpu intensive games. Average single threaded performance uplift is 25%, don't know about multi. Apparently amd worked hard to scale the core to achieve ultra high clock. Heat will be a problem when really pushed. Raptor will get a beat down from v-cache version, only 8 cores and 12 cores will receive it (??). Little birdie = co worker who worked me in a firm who used to work with nvidia.

I think Intel will get quite some wins on MT because of their e-cores, but the question is how important that is to the buyers. Gamers will look at 3D cache models vs RPL, content creators 7900X and 7950X vs 13900K and the majority of those buying the lower SKUs will probably never know the difference of what CPU they have in their computer.

 

[Gideon]

this time the chiplets are literally side by side.

My little pet-hope is this is for huge V-cache enabled chips, with the top die spanning over both CCDs combining the entire L3 of the two (and adding at least 128MB)

One can dream

 

[DrMrLordX]

I think Intel will get quite some wins on MT because of their e-cores

Intel's problem is they're having to add more cores. AMD is staying at 16c/32t and gaining massively from IPC and clockspeed increase. All software that ran well on a 5950X will run better on a 7950X. That won't necessarily be the case between the 12900k and 13900k.

 

[Joe NYC]

My little pet-hope is this is for huge V-cache enabled chips, with the top die spanning over both CCDs combining the entire L3 of the two (and adding at least 128MB)

One can dream

I had a similar dream, but then I came across some Geona SKU leak, and it appears that +64MB per chiplet is the max, at this time.

 

[nicalandia]

It's almost here....

 

[Zepp]

I think the market for personal computers with more than 16 "Performance" cores and a limitation af 128GB memory is quite small.

I maybe out of touch here but I would have said the market for personal computers with more than 8 cores and 64GB memory limit was relatively small as well.
And likely the majority that bought 10-12-16 core CPU's did so more for bragging rights so to speak than an actual need.

 

[CakeMonster]

During the ~2 year duration of the upcoming generation I think we'll see those people more annoyed with PCIE lanes and versions than lack of cores. You might have to choose between PCIE5 capability on your GPU and an NVME drive, and toward the end of that period I predict most new drives and GPU's will support it. People will also increasingly use more than 1 NVME drive.

 

[nicalandia]

Look what we have here. Most likely the one used for the Geekbench 5 Benchmark




As you have already been told before, you must provide your own commentary when posting a picture.

Daveybrat
AT Moderator

 

[biostud]

During the ~2 year duration of the upcoming generation I think we'll see those people more annoyed with PCIE lanes and versions than lack of cores. You might have to choose between PCIE5 capability on your GPU and an NVME drive, and toward the end of that period I predict most new drives and GPU's will support it. People will also increasingly use more than 1 NVME drive.

Obviously, but content creators, semi proffs, streamer, small business etc. is still a market.

 

[Mopetar]

But how many of those would want a TR instead if it were available? The streamers and small businesses are the only ones who don't really benefit from more cores, but they also don't really benefit from more than 16 as well. Even 16 may be overkill for many.

 

[coercitiv]

You might have to choose between PCIE5 capability on your GPU and an NVME drive, and toward the end of that period I predict most new drives and GPU's will support it. People will also increasingly use more than 1 NVME drive.

My understanding is Zen4 offers enough PCIe 5.0 lanes so that won't be an issue:

• 16 lanes for the GPU slot
• 4 lanes for an M2 drive
• 4 lanes for another M2 drive or Thunderbolt 4 / USB4
• 4 lanes for chipset (which will function at 4.0 speeds due to chipset limitations)

What we might see happen on some boards such as this will be exposing the second set of 4x PCIe 5.0 lanes from the CPU through an expansion slot, giving the user another degree of freedom towards I/O use. Other boards will offer 2 PCIe 5.0 M2 slots with independent access to CPU and call it a day.

 

[inf64]

Hassan from Wccftech is hinting on twitter that he's heard that Raphael announcement might be delayed for another 2 weeks. So if this is true, we will see it in October

 

[Vope45]

No need to be sarcastic, I like amd a lot, just more critical of them than the average forum poster.



It matters when Amd hit the ipc plateau which will happen eventually or when tsmc lose their edge against intel foundry. It seems to me intel is trying to go lean bulk which takes a long time. To compete in the next 5+ years I think amd need their own fab for sure.

Just a slight off topic. I followed @SkyJuice60 aka Angstronomics after he leaked the correct zen 4 ipc uplift, awhile back. Same guy just leaked https://www.angstronomics.com/p/amds-rdna-3-graphics. RDNA 3 allegedly is all about area. That means amd is wafer constraint, meaning zen 4 is a bigger compromise than the previous zen iterations aka smaller ipc increase when die size plays a larger role.

 

[inf64]

Just a slight off topic. I followed @SkyJuice60 aka Angstronomics after he leaked the correct zen 4 ipc uplift, awhile back. Same guy just leaked https://www.angstronomics.com/p/amds-rdna-3-graphics. RDNA 3 allegedly is all about area. That means amd is wafer constraint, meaning zen 4 is a bigger compromise than the previous zen iterations aka smaller ipc increase when die size plays a larger role.

It's all relative. Zen 4 should bring larger performance uplift in both ST and MT workloads than Zen 3 brought versus Zen 2. If AMD did it with less die area then kudos to them. ~10% IPC combined with 20%+ clock uplift is massive generational uplift (cumulative).

edit:

Here's a quote from SkyJuice60 about Zen 4 IPC target (full article here):

"As for the claimed Single-Thread Uplift of ‘greater than 15% expected‘, Angstronomics can confirm this is a conservative value, done at below final frequencies and using Maxon’s Cinebench R23 Single Thread Benchmark. We can independently confirm that the Performance Per Clock (PPC) targets for the Zen4 core are targeted at +7% Single-Thread PPC, +10% Multi-Thread PPC over their Zen3 core, with significantly higher PPC for memory sensitive workloads thanks to DDR5 while core execution bound workloads like Cinema4D have a lower PPC improvement. "

So in summary, the IPC should be much greater than 10% for memory sensitive workloads (such as games for example). R23 and other compute bound workloads are bringing the average IPC figure down by quite a bit, which is interesting.

 

[Vope45]

It's all relative. Zen 4 should bring larger performance uplift in both ST and MT workloads than Zen 3 brought versus Zen 2. If AMD did it with less die area then kudos to them. ~10% IPC combined with 20%+ clock uplift is massive generational uplift (cumulative).

Amd apparently spent a lot of transistors for scaling so they ran out of space for higher ipc uplift. Also perf increase due to uplift isn't linear, I was told to expect 25% on average in the best case scenario - what ever that means. They could have done more to be honest but it was time for zen to clock past 5ghz. If zen 4 gaming performance is spectacular due to a high sustained clock, it'd be worth it. If not zen 5 on 4nm could not be as competitive as the previous zen. Note this what I'm reading between the lines with common sense.

 

[eek2121]

I think the market for personal computers with more than 16 "Performance" cores and a limitation af 128GB memory is quite small. If you have work that can use more cores, it can probably also use more memory and more memory bandwidth etc. -> Threadripper

For that little niche of people who would be in the market for the less expensive Threadrippers, I can understand the frustration that they are going the way of the dodo, but from a business standpoint I think it makes perfectly sense for AMD.

I have to disagree here. There is definitely a market for 24/32 core on AM5 or something similar. AMD HEDT Is Threadripper Pro. The Threadripper Pro platform starts at between $4,000 and $8,000 for a 32c/64t system. There are plenty of small shops that can't afford spending dropping 15-20 grand on a new build for a couple of employees. I used to work at such a company. Most of the workload was related to rendering, graphics work, and video editing/encoding. Another example: I do content creation and software development for websites I own. I do a lot of similar stuff. It's not that I personally can't afford 4-8 grand for such a system, it's that I'd rather invest that money elsewhere. I don't need the memory bandwidth or PCIE lanes, I just need the cores. There are others that ONLY need more PCIE lanes, which is actually why I think AMD should push out third socket for client. Stick the platform halfway between Ryzen and Threadripper in terms of features and sell it for about halfway between in cost.

If AMD had a 24c part for $999-$1,199 and a 32c part for $1,299-$1,499 for AM-5, they'd sell.

 

[inf64]

Amd apparently spent a lot of transistors for scaling so they ran out of space for higher ipc uplift. Also perf increase due to uplift isn't linear, I was told to expect 25% on average in the best case scenario - what ever that means. They could have done more to be honest but it was time for zen to clock past 5ghz. If zen 4 gaming performance is spectacular due high an sustained clock, it'd be worth it. If not zen 5 on 4nm could not be as competitive as the previous zen. Note this what I'm reading between the lines with common sense.

Well we have the average IPC number and the max boost clock, so it's pretty straightforward to calculate the maximum ST uplift.

It's trickier for MT but AMD listed >35% in R23 example which is arguably the worst case scenario (due to being compute bound; see my edit above). If the 16C part can sustain 5.3Ghz boost on all threads (vs ~4.2Ghz for 5950X), coupled with average ~10% IPC you get ~38% better MT performance in things like R23 (matches well with >35% supplied by AMD). If it's memory bound, it can be much higher (think ~50%)

 

[Vope45]

Well we have the average IPC number and the max boost clock, so it's pretty straightforward to calculate the maximum ST uplift.

It's trickier for MT but AMD listed >35% in R23 example which is arguably the worst case scenario (due to being compute bound; see my edit above). If the 16C part can sustain 5.3Ghz boost on all threads (vs ~4.2Ghz for 5950X), coupled with average ~10% IPC you get ~38% better MT performance in things like R23 (matches well with >35% supplied by AMD). If it's memory bound, it can be much higher (think ~50%)

I don't recall amd saing zen 4 is 35% faster in R23?

 

[biostud]

I have to disagree here. There is definitely a market for 24/32 core on AM5 or something similar. AMD HEDT Is Threadripper Pro. The Threadripper Pro platform starts at between $4,000 and $8,000 for a 32c/64t system. There are plenty of small shops that can't afford spending dropping 15-20 grand on a new build for a couple of employees. I used to work at such a company. Most of the workload was related to rendering, graphics work, and video editing/encoding. Another example: I do content creation and software development for websites I own. I do a lot of similar stuff. It's not that I personally can't afford 4-8 grand for such a system, it's that I'd rather invest that money elsewhere. I don't need the memory bandwidth or PCIE lanes, I just need the cores. There are others that ONLY need more PCIE lanes, which is actually why I think AMD should push out third socket for client. Stick the platform halfway between Ryzen and Threadripper in terms of features and sell it for about halfway between in cost.

If AMD had a 24c part for $999-$1,199 and a 32c part for $1,299-$1,499 for AM-5, they'd sell.

I have no doubt there is a market, the question is whether it is profitable enough, or if it eats the margins of Threadripper Pro.

 

[inf64]

I don't recall amd saing zen 4 is 35% faster in R23?

They have, on Financial Analyst Day, June 9, 2022. Presentation link (slide #13):

 

[Hans Gruber]

I have no doubt there is a market, the question is whether it is profitable enough, or if it eats the margins of Threadripper Pro.

It's not about profitability at the top SKU in Zen 4 (consumer grade) CPU's. It's bragging rights. Intel has the 13900k. A 24core and 32core Zen 4 CPU would run the table against anything in Raptor Lake. The 3D stacked V-cache would take care of the rest.

I am assuming the voltage issue with Zen 3 3D-cache will not be an issue with Zen 4. So v-cache CPU's will clock just as high as non v-cache Zen 4 chips.

 

[dnavas]

I have no doubt there is a market, the question is whether it is profitable enough, or if it eats the margins of Threadripper Pro.

I'm not paying $100/core for their workstation part -- no margin loss here.
I'd buy a 24/32 core 7950ZX. Particularly if there's a motherboard out there that splits some of those pcie5 lanes into a crapton of more useful pcie4 lanes.... I'd pay decent money for more lanes and memory as well -- I just don't need workstation-level capability, and I'm not willing to overpay for stuff I'm not going to use....