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!

 

[gdansk]

Some napkin math from me

So if we assume that Cinebench is among the worst case (low end of the scale) in terms of "IPC" increase from Zen3 to Zen4, ref this picture: (Cinebench r20 ST is + 13% while geomean is 19%)
View attachment 62850

Zen3 fmax 5050mhz VS Zen4 fmax 5850mhz = ~15% higher peak clockspeeds (sustain clocks probably see an even bigger increase)

ST Performance base (Zen3) x 1.10 IPC x increased clock speed by 15% = ~26% higher ST performance in "worst case apps" like Cinebnench R23

In games, we can probably/maybe quickly get "+20% IPC" or more considering doubling the L2 cache that Cinebench does not care about.

Gaming performance (Zen3) x 1.2 IPC x increased clock speed by 15% = ~38% higher performance in "games". (or more thanks to better sustained clocks)

On top of this we get the benefit from going from DDR4 to DDR5 with hopefully much higher infinity fabric clockspeeds, which has limited Zen3 performance in games (latency) from the get-go.

All in all, not bad i would say

Here is my napkin maths. For pessimists such as myself.
Clock rate is closer to 5550 than 5850.
The 8% exists in the "8-10% IPC" increase for a reason.
5520/5050 * 1.08 = closer to 118% ST in CB23, methinks.

 

[Markfw]

In fact that's what people have been expecting for MT all along due to the higher TDP. You have to remember it's two different charts too. It may be 25% more performance per watt but that >35% is when not at the same power level. They're throwing another ~90W at it to get that last >10%. And unfortunately that will be putting its power consumption much closer to ADL than the 5950X was, out of the box. Of course it is promising for people who set their own PPT.

Where do you get that 90w figure ? NO wattage numbers except the SOCKET have been released. And getting 25% more performance/watt ? It must be using LESS power, not more. But lets wait until the reviews come out on that.

 

[Det0x]

Here is my napkin maths. For pessimists such as myself.
Clock rate is closer to 5550 than 5850.
The 8% exists in the "8-10% IPC" increase for a reason.
5520/5050 * 1.08 = closer to 118% ST in CB23, methinks.

Why are you comparing fmax for zen3 (5050mhz) with sustained clockspeeds for Zen4 in a game ? (5520mhz)

We can just as well say that Zen3 sustain ~4.8ghz in games while Zen4 sustain ~5.5ghz in games --> we still get the same +15% increase in clockspeed
While both you and me know that a stock Zen3 dont stay at 4.8ghz while gaming, even if it happened to be a low threaded game that only use 4-6 threads...

 

[exquisitechar]

ST Performance base (Zen3) x 1.10 IPC x increased clock speed by 15% = ~26% higher ST performance in "worst case apps" like Cinebnench R23

They showed 15% in Cinebench R23, right? I don't know what clock speed that was run at (assuming close to final), but then the IPC increase there is probably significantly less than 10%. I expect it to be a bad case, indeed. The doubled L2 cache will not bring a big improvement there, neither will DDR5, AVX-512 support, etc.

 

[Tuna-Fish]

Where do you get that 90w figure ? NO wattage numbers except the SOCKET have been released. And getting 25% more performance/watt ? It must be using LESS power, not more. But lets wait until the reviews come out on that.

35% more performance, +25% perf/watt implies increase in power use right there. Not a devastatingly high increase, but it's clearly there.

 

[inf64]

They showed 15% in Cinebench R23, right? I don't know what clock speed that was run at (assuming close to final), but then the IPC increase there is probably significantly less than 10%. I expect it to be a bad case, indeed. The doubled L2 cache will not bring a big improvement there, neither will DDR5, AVX-512 support, etc.

They showed >15%. The greater than sign is key. The footnote in the slide even mentions that R23 1T was included in the ST IPC calculation, along with Geekbench 1T and specint/specfp2017. R23 and Geekbench were on the lower end of the scale in Zen3's case.

Zen4 is most likely ~8% faster than Zen3 (IPC) and runs at >5.5Ghz in ST R23 test . Hence the >15% performance. If Zen4 reached 5.8Ghz for 1T, it could be actually scoring higher than 12900KS in R23 , getting around ~2100 pts (see here for 12900KS/5950X scores).

 

[nicalandia]

40-45% gain over previous gen 16-core part is simply unachievable, even with increase in clock speeds. Go look at each 6/8/12/16 Core SKU from Zen 2 -> Zen 3.

It will be close to 40% due to being advertised as >35 so there...

 

[inf64]

It will be close to 40% due to being advertised as >35 so there...

Pretty much as expected : 1.1 (IPC) x 5.2GHz (all core boost for Zen4) / 4.2Ghz (all core boost for 5950X) ~= 1.36 or 36%

If all core boost is >5.2Ghz, the MT performance could be even higher.

edit: there is a slide in PC desktop presentation where they list ~8% higher IPC in "desktop" specific workloads and they cite R23 1T and Geekbench being in the mix. This was basically the worst case scenario for Zen3, but let's see how other desktop workloads perform when Zen4 launches.

They also listed >5.5Ghz ST clock, but no end target. So it's possible they can get 5.6-5.8GHz and that's why the ambiguous >15% ST uplift which basically references the stupid R23 1T workload.

Linky:

https://d1io3yog0oux5.cloudfront.net/_19fa4c9a17d360ec4e0900539d09c524/amd/db/812/6837/presentation/FAD+2022_Saeid+Moshkelani_Final+Post.pdf

 

[gdansk]

Where do you get that 90w figure ? NO wattage numbers except the SOCKET have been released. And getting 25% more performance/watt ? It must be using LESS power, not more. But lets wait until the reviews come out on that.

TDP goes from 105W to 170W, that's acknowledged by AMD. The PL2 is a linear estimation (142/105 * 170 = 230W). I would love to see the math on using less power than more... how do you reckon that

Why are you comparing fmax for zen3 (5050mhz) with sustained clockspeeds for Zen4 in a game ? (5520mhz)

We can just as well say that Zen3 sustain ~4.8ghz in games while Zen4 sustain ~5.5ghz in games --> we still get the same +15% increase in clockspeed
While both you and me know that a stock Zen3 dont stay at 4.8ghz while gaming, even if it happened to be a low threaded game that only use 4-6 threads...

Because that's the highest AMD showed. Everything else is speculative optimism. I find everyone saying AMD is sandbagging 10% to be exhausting and not really based on anything except "leaks".

 

[Hans Gruber]

View attachment 62851

35% more performance, +25% perf/watt implies increase in power use right there. Not a devastatingly high increase, but it's clearly there.

That is wrong. Clock for clock on the same number of cores 5950x vs 7950x would yield a 25% improvement in power consumption for Zen 4 in theory. People get confused by increasing the TDP in Zen 4. The reason is to better handle the 24 core or 32core Zen 4 CPU's yet to be released. Since the clocks for Zen4 will be 5.5ghz or more that power efficiency could be lost on paper because power consumption would be greater for higher clock speed but should be less than the 5950x by 5-10%. I like to ballpark my estimates.

 

[gdansk]

That is wrong. Clock for clock on the same number of cores 5950x vs 7950x would yield a 25% improvement in power consumption for Zen 4 in theory. People get confused by increasing the TDP in Zen 4. The reason is to better handle the 24 core or 32core Zen 4 CPU's yet to be released. Since the clocks for Zen4 will be 5.5ghz or more that power efficiency could be lost on paper because power consumption would be greater for higher clock speed but should be less than the 5950x by 5-10%. I like to ballpark my estimates.

You are implying they aren't even hitting the TDP when in an all core workload?

 

[trexfromouterspace]

Remember when people were circlejerking about how Zen 4 would provide >20% IPC uplift over Zen 3?

Pepperidge Farm remembers.

 

[Hans Gruber]

You are implying they aren't even hitting the TDP when in an all core workload?

They wouldn't increase it if they didn't hit it.

The 170w TDP has been leaked by sources. They also said AMD would release a 24core and 32core Zen 4 chip. I am implying that the 170w TDP is for a 24core or 32core Zen 4 CPU. I say this because 105w processor. That is on 7nm and Zen 4 is on 5nm with improved efficiency.

 

[gdansk]

The 170w TDP has been leaked by sources. They also said AMD would release a 24core and 32core Zen 4 chip. I am implying that the 170w TDP is for a 24core or 32core Zen 4 CPU. I say this because 105w processor. That is on 7nm and Zen 4 is on 5nm with improved efficiency.

Hmm, we'll see. Why not use the new power limit right out of the gate? Intel is already over it, why artificially limit yourself.

 

[Ajay]

Zen has been server first from the very beginning. The very first Zen design, the Zeppelin die, was designed to work in a group of four, for servers. The consumer oriented designs are the mobile APUs. Arguably the first actual desktop specific design is Raphael's IOD including a minimalist iGPU, though that may or may not be part of mobile Dragon Range as well.

Point taken, from that perspective - but AMD marketed heavily in the desktop space, specifically DIY. They needed early adopters to start proving out their technology. If the performance gains weren't their for desktop (over the construction cores) it would have been game over . Server users had become very skeptical of anything not Intel.

 

[Markfw]

Hmm, we'll see. Why not use the new power limit right out of the gate? Intel is already over it, why artificially limit yourself.

Even if it was 170 for a 7950x, thats 28 watts more, not 90. Again, NO chip full load wattage has been announced, but >35% total performance and >25% efficiency. I will stick with that until reviews come out.

Edit: and Intel was desperate to beat the 5950x in ANYTHING, just to say they were in the rear view mirror. Thats why they jacked the power to the limits. AMD is not so stupid (we will see, but again, I will stick with this for now.)

 

[moinmoin]

Point taken, from that perspective - but AMD marketed heavily in the desktop space, specifically DIY. They needed early adopters to start proving out their technology. If the performance gains weren't their for desktop (over the construction cores) it would have been game over . Server users had become very skeptical of anything not Intel.

AMD likes targeting DIY and gamers for mindshare and easy publicity, see also X3D having been pushed as game cache 2.0 initially while clearly being workspace/server tech first and foremost. For the first market share growth that audience was very important as well since mobility is the highest there. But to solidify market share OEM and business sales are more important which is why direct to consumer sales are affected of scarcity.

Back to the design point, I agree desktop influence is the biggest in marketing, and marketing is pushing the visible competition aspect against Intel, so matching/passing core count (Zen), performance at better efficiency (Zen 2), ST (Zen 3 until ADL), and frequency (Zen 4). The latter two arguably are pretty much focused on desktop indeed, though AMD did find out that there's demand for higher frequencies in server chips as well (Epyc 7xF3 series), and the effort in Zen 4 should help smooth out the F/V curve which will help in DC as well.

 

[inf64]

At the end of the day, it's the net performance and perf./watt that matters.

From Zen2 (3950X)->Zen3 (5950X) we had 24% ST uplift, 15% MT uplift, and 32% gaming uplift.
Reference: https://www.computerbase.de/thema/prozessor/rangliste/#abschnitt_die_computerbaseperformanceratings

From Zen3->Zen4, we potentially have ~28% maximum ST uplift (~5.8Ghz, +10% IPC), >36% MT uplift (~>5.2Ghz? all core turbo, +10% IPC), and potentially ~30% gaming uplift (~5.35Ghz gaming all core boost?, +10% IPC).

So, Zen4 could have higher 2 out of 3 performance jumps than what Zen3 brought, where MT could be drastically better than what we got with Zen2->Zen3, and ST could be noticeably better. Gaming could be a "wash", ~32% vs potential ~30%. This doesn't count in the Zen4Vcache that should add another 10-15% on top of Zen4 numbers.

Zen4 could pretty much be noticeably higher performance jump than Zen3, even if it sounds crazy. Ask yourself, would you rather have Zen4 with ~19% IPC that boosts to ~5.2Ghz ST and ~4.6GHz all threads, OR a Zen4 with 8-10% higher IPC that boosts to 5.8GHz ST and ~5.2-5.3Ghz all threads? The later choice has higher ST and MT performance.

 

[Hans Gruber]

Hmm, we'll see. Why not use the new power limit right out of the gate? Intel is already over it, why artificially limit yourself.

The 5950x TDP is 105w. Going from 105w to 170w is a 62% increase in power on paper. I think the 170w TDP number originally came from Gigabyte docs stating what power requirements the socket would require on x690 motherboards.

It was speculated that AMD would release either a 24core or 32core Zen 4 CPU on the release of Zen 4. Since the plan changed, people are attributing a 170w TDP to the 16 core 7950x. Remember, the AM5 socket is new and is designed for future processors beyond Zen 4.

 

[scannall]

The 5950x TDP is 105w. Going from 105w to 170w is a 62% increase in power on paper. I think the 170w TDP number originally came from Gigabyte docs stating what power requirements the socket would require on x690 motherboards.

The 170 watt is the power requirement for the motherboard. The chips themselves may or may not pull that much at stock. Have to wait on reviews to know that for sure.

 

[uzzi38]

Edited the post

Right. So then, my response changes to: stop trying to project your wet dream fanfics on reality.

Oh wait, that's not all that different, is that?

 

[gdansk]

Even if it was 170 for a 7950x, thats 28 watts more, not 90. Again, NO chip full load wattage has been announced, but >35% total performance and >25% efficiency. I will stick with that until reviews come out.

AMD would like to issue a correction to the socket power and TDP limits of the upcoming AMD Socket AM5. AMD Socket AM5 supports up to a 170W TDP with a PPT up to 230W. TDP*1.35 is the standard calculation for TDP v. PPT for AMD sockets in the “Zen” era, and the new 170W TDP group is no exception (170*1.35=229.5).

This new TDP group will enable considerably more compute performance for high core count CPUs in heavy compute workloads, which will sit alongside the 65W and 105W TDP groups that Ryzen is known for today.

— AMD Representative to Tom’s Hardware

TDP: 105W to 170W (+65W)
PL2: 142W to 230W (+88W)
In Cinebench MT they're probably at PL2. Which is still less than Intel.

Hans thinks we are seeing a 16 core 105W part. I think we're seeing a 16 core 170W part for the >35% performance comparison and a 105W configuration for the 25% performance per watt comparison.

 

[Markfw]

TDP: 105W to 170W (+65W)
PL2: 142W to 230W (+88W)
In Cinebench MT they're probably at PL2. Which is still less than Intel.

Hans thinks we are seeing a 16 core 105W part. I think we're seeing a 16 core 170W part for the >35% performance comparison and a 105W configuration for the 25% performance per watt comparison.

There is NO WAY the 7950x is going to use 230 watt. That I will bet on. You really need to look at your math, its out of whack compared to what AMD has announced, and unlike Intel, they have been honest with Zen on release statistics, UNLIKE Intel

 

[Hans Gruber]

TDP: 105W to 170W (+65W)
PL2: 142W to 230W (+88W)
In Cinebench MT they're probably at PL2. Which is still less than Intel.

Hans thinks we are seeing a 16 core 105W part. I think we're seeing a 16 core 170W part for the >35% performance comparison and a 105W configuration for the 25% performance per watt comparison.

I know the power consumption is higher than 105w but what the CPU is rated at matters. Look at the 5800x rated @105w, that is a 5950x with 8 cores disabled but with the same power usage as the 5950x. People have to manually (5800x) detune the voltage. The 5700x does it for you out of the box and is rated at 65w. Again, I am totally aware that the power usage is more than the TDP out of the box.

 

[gdansk]

There is NO WAY the 7950x is going to use 230 watt. That I will bet on. You really need to look at your math, its out of whack compared to what AMD has announced, and unlike Intel, they have been honest with Zen on release statistics, UNLIKE Intel

They said it is 230W socket. Do you think they're going to leave available performance when they will be so close to Raptor Lake?
We'll see but it seems like the >35% performance must be at least 5GHz based on their IPC estimation. Maybe a bit more. How much power does say 5050MHz x 16 need? 230W/16 is only 14W per core. 1 Zen 3 core at 5050MHz draws over 20W.
It's still an improvement in efficiency if they can go over 5050MHz all core in a 230W power limit.