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!

 

[jpiniero]

5456Mhz is ST clock gain, not MT . Everything looks ok score wise, all core boost is great, around 5.3Ghz.

It is going to lose to the locked 6+8 i5s with the E Core Spam. Maybe not by much (@ 110 W) but probally 15-16k. This based upon NBC's ADL mobile results.

NBC even has a 75/90 W 6900HS laptop which does about 14k. Granted that is 8 cores versus 6.

 

[qmech]

we know Intel 4 will be garbage. Why else would they need to follow it up/replace it with Intel 3 immediately after?

That is a misinterpretation of what Intel 4 is. Rather, the need for a quick successor in Intel 3 is a function of eschewing features not strictly needed for compute dies. This is a risk reduction and faster time-to-market effort and says nothing about the performance characteristics of Intel 4.

From a technical perspective, Intel 4 will only feature a single standard library (high performance), as opposed to the 3+ (high density, ultra-high performance) of previous nodes. This limits the target usage of the node, basically mandating the use of a tile based approach, since only the compute part is practical without a fuller set of features.

Additionally, the use of EUV is limited to the layers that absolutely need them, resulting in some layers (even the M0 layer) still being done with DUV and requiring SAQP.

For Intel 3 a full set of standard libraries will be available and EUV use will expand to reduce cost and increase throughput - and enabling a tightening of some pitches and design rules.

The physical characteristics of Intel 4 are well-known by Intel at this time. They have full Meteor Lake compute tiles in hand, so expecting "complete garbage" seems unwarranted.

The only remaining question is what kind of yields are achieved (and how quickly) when ramping outside the Hillsboro development fab. Even here, extremely low yields would seem unlikely given the focus on minimizing the issues that plagued 14nm/10nm development (no more Co metal layers!).

 

[nicalandia]

It is going to lose to the locked 6+8 i5s with the E Core Spam. Maybe not by much (@ 110 W) but probally 15-16k. This based upon NBC's ADL mobile results.

NBC even has a 75/90 W 6900HS laptop which does about 14k. Granted that is 8 cores versus 6.

Not in ST.. Let's face it if you get a 7600X it will not be for MT workloads, but for Gaming. Pair that with a High end GPU and you are set for the next few years(as you would if you have selected any late 6C/12T CPU built in the last 2 years)

 

[Abwx]

I hope this wasn't posted yet.
AMD Ryzen 5 7600X tested in Cinebench R23 with AMD Core Performance Boost disabled

	R5 5600x(Link)	R5 7600x (CPB Off)	R5 7600x (CPB On)
CB R23	11225	13003	14767
Clockspeed	~ 4550 MHz	4705 MHz	5456 MHz
Power consumption	~76 W ?	60 W	110 W
Performance gain	100 %	116 %	132 %
Power draw	100 %	79 %	145 %
Power efficiency	100 %	147 %	91 %

If this will hold true even for retail version, then as I thought, If you want max performance you will have to sacrifice some efficiency.
But 110W in CB R23 is better than I expected based on the rated 105W TDP(142W PPT), I expected 120-125W.
If you don't need max performance, then efficiency is much much better than Zen3.

P.S. If someone has 5600x, then please post his package power during CB R23. I am not sure, If 76W is accurate.

At same frequency than a 5600X it should use less than half of the power given the more efficient IOD, so there s no way that it require 60W at 4705MHz, that seems to be a badly set manual setting voltage wise.

 

[PJVol]

P.S. If someone has 5600x, then please post his package power during CB R23. I am not sure, If 76W is accurate

Yep, 76W is stock ppt limit. As for the 7600x results, the 2nd score seems temp. limited (modest cooling), so I would guess the clocks hardly exceeded 5.0ghz in the MT test.
CBP realy shines if adequate cooling applied together with the reducing ATE and BTC voltage guardbands as much as possible.
As a side note, running ram oc'ed makes such a comparision less useful.

 

[Hans Gruber]

At same frequency than a 5600X it should use less than half of the power given the more efficient IOD, so there s no way that it require 60W at 4705MHz, that seems to be a badly set manual setting voltage wise.

My 5600 is manually set (1.35v) voltage. I run it at 4.7ghz locked in 24x7 and it uses a maximum 75watts running Cinebench or gaming. In Ryzen master the PPT is either 18-19% of 395w.

I cannot see how the 7600x goes above 80-85w power consumption. It's only a 6 core 12 thread CPU. It does not take as much juice as the 8 core 16 thread parts.

 

[Abwx]

There s not only the power that do not make sense, the temps are not logical at allegedly 110W.

If we look at the first pic at 60W temp is 56 °C.
Assuming 26 °C ambiant temp then the cooling thermal resistance is about 0.5 °C/Watt
At 110W the temp over ambiant should be 110 x 0.5 = 55 °C more than 26 °C, that is 81 °C for the CPU.

 

[PJVol]

If we look at the first pic at 60W temp is 56 °C.

My 5600x with the CPB turned off and ambient 23° running at ~40° consuming ~ 54W. When auto and max tuned Tctl is around 73°c and 105W PPT.
Doesn't your math looks a bit off here?

 

[IntelUser2000]

As you see, when enabling CPB it has ~16% clock uplift(5456/4705) but only has 13.56%(14767/13003) R23 score gain. I would rather wait for the BIOS being more stable.

Is the R23 relation to clockspeed linear or something? I know up to R20 the scaling is 85-90%. 13.56/16 equals 84.75. Seems perfectly fine to me.

Remember, linear scaling is an indicator of a terrible benchmark.

 

[IntelUser2000]

Why else would they need to follow it up/replace it with Intel 3 immediately after?

The analysis is based on IEDM and Hot Chips presentations and few reputable sites already commented on that.

*Note: This is at a process level only.

I also think some people need to look up when was the last time a Intel node was used in a low power or power efficient design (or worse, in a successful/innovative one)... I can wait. Even Samsung probably beats them here.

Again, their design side also has issues. Why do you think Pentium M did far better than Pentium 4? Cause they used a different process?

Last but not least, they seem more expensive than TSMC too. Even though capacity may not be a problem for them. Intel Alchemist anyone?

Alchemist doesn't even use Intel node. It's on TSMC N6.

Intel 4 is having a short lifespan so they can accelerate process node introductions and catch up and eventually exceed the competition. Dismissing it when as you already noted, when it doesn't even exist is weird.

 

[PJVol]

My 5600 is manually set (1.35v) voltage. I run it at 4.7ghz locked in 24x7 and it uses a maximum 75watts running Cinebench or gaming.

Are you sure about 75W? For the six-core Zen3 it's just not enough to run at 4700@1.35V. Though power could vary depending on temps and load-line slope set, it should be closer to 100W.
In CB R23 MT PPT was ~ 99W with my 5600X running 4700@1.35V and Vdroop 45-50mv.

Is the R23 relation to clockspeed linear or something?

Close to linear I think, so the MT clock had to be around 5300mhz @110W. Impressive, to say the least.

 

[deasd]

It is going to lose to the locked 6+8 i5s with the E Core Spam. Maybe not by much (@ 110 W) but probally 15-16k. This based upon NBC's ADL mobile results.

E core SPAMMING would shine in some situation but others do not, especially if you consider the complexity of scheduler.

And locked i5 doesn't have enough clock to compete against 7600x. Only if you care about the well-thread benches which don't make heavy use of AVX256/512.

I have some horrible examples showing how bad the big-LITTLE perform in realworld but I wouldn't post them in AMD thread.

Not in ST.. Let's face it if you get a 7600X it will not be for MT workloads, but for Gaming. Pair that with a High end GPU and you are set for the next few years(as you would if you have selected any late 6C/12T CPU built in the last 2 years)

'MT workload' is too ambiguous. It all depends on how you use your PC......

Remember, linear scaling is an indicator of a terrible benchmark.

Close to linear I think, so the MT clock had to be around 5300mhz @110W. Impressive, to say the least.

Cinebench scales pretty well with clock. But the 7600x the VCZ benched lacked some details like clock and voltage behavior. Especially I just wonder the CPB and PBO differences inside Zen4, is CPB behavior closed to PBO limit?

 

[PJVol]

But the 7600x the VCZ benched lacked some details like clock and voltage behavior. Especially I just wonder the CPB and PBO differences inside Zen4, is CPB behavior closed to PBO limit?

I don't think boosting algorithm has changed much (if at all). Not sure I understand what you meant "CPB close to PBO limit". PBO/PBO2 limits are just CPB(aka Precision Boost) overridables for the x86 space or BIOS.

 

[igor_kavinski]

Remember, linear scaling is an indicator of a terrible benchmark.

Sorry to bother you but I have to ask:

Would you be kind enough to provide some examples of linear scaling benchmarks and some non-linear scaling (hence good) benchmarks?

 

[igor_kavinski]

Dismissing it when as you already noted, when it doesn't even exist is weird.

Intel Meteor Lake silicon is up and running in Intel labs. It exists!

 

[deasd]

I don't think boosting algorithm has changed much (if at all). Not sure I understand what you meant "CPB close to PBO limit". PBO/PBO2 limits are just CPB(aka Precision Boost) overridables for the x86 space or BIOS.

I mean if default CPB is that high(>5.3Ghz all core for 7600x) and temperature is more than 90C, would there any headroom left for PBO? It would be a shame if PBO just give 100Mhz to us.

 

[Hans Gruber]

Are you sure about 75W? For the six-core Zen3 it's just not enough to run at 4700@1.35V. Though power could vary depending on temps and load-line slope set, it should be closer to 100W.
In CB R23 MT PPT was ~ 99W with my 5600X running 4700@1.35V and Vdroop 45-50mv.

Close to linear I think, so the MT clock had to be around 5300mhz @110W. Impressive, to say the least.

I did Cinebench again. 102.7-106.65w all core under full load. In bios my vcore is set to 1.35v. In real world testing the voltage went up to 1.38-1.39v according to CPU-Z. Temps on my 240mm AIO were peak 82-84C according to Zen Master. All numbers except vcore are Zen Master numbers. I was giving single core PPT watts in my earlier post.

I think the 7600x is going to push higher temps (90C) 122-130w more than Zen 3 if the Zen4 (7600x) clocks are 5.3-5.5ghz all cores on Cinebench.

I think people may want to consider getting 280-360mm AIO's for Zen4. Or they will have to crank up their AIO fan speeds.

 

[igor_kavinski]

How many people here are planning fish tank builds for Zen 4 and Raptor Lake? They would offer the best performance, using the non-conductive 3M coolant.

PC Submerged In Cooling Liquid Built by Der 8auer. : pcmasterrace (reddit.com)

Behind GIGABYTE’s Fish Tank Oil System: A New Take on Liquid Cooling (anandtech.com)

 

[PJVol]

In bios my vcore is set to 1.35v. In real world testing the voltage went up to 1.38-1.39v according to CPU-Z. Temps on my 240mm AIO were peak 82-84C according to Zen Master

Out of curiosity, what's your LLC CPU settings (if other than Auto), and what "CPU Core Voltage (SVI2)" and "CPU Core VID (effective)" reported in HWInfo64 during CB R23 Multi?
The voltage reported by CPU-z isn't accurate as its source is SuperIO. Either it's reporting nonsense or something is off with the board's VRM settings.

 

[PJVol]

I mean if default CPB is that high(>5.3Ghz all core for 7600x) and temperature is more than 90C, would there any headroom left for PBO? It would be a shame if PBO just give 100Mhz to us

I see... you meant Boost Override. But does it really matter how Fmax is derived? I'd rather like to know how conservatively (or not) AMD chose Fmax for each model this time around

 

[Hans Gruber]

Out of curiosity, what's your LLC CPU settings (if other than Auto), and what "CPU Core Voltage (SVI2)" and "CPU Core VID (effective)" reported in HWInfo64 during CB R23 Multi?
The voltage reported by CPU-z isn't accurate as its source is SuperIO. Either it's reporting nonsense or something is off with the board's VRM settings.

The CPU-Z was a screenshot after running Cinebench. It normally says 1.34-1.35v. Mine LLC is 3rd highest setting. My bios setting for vcore is 1.35v. I have my multiplier locked in at 47. The voltage on CPU-Z works and goes above 1.35v with the all core Cinebenc

Out of curiosity, what's your LLC CPU settings (if other than Auto), and what "CPU Core Voltage (SVI2)" and "CPU Core VID (effective)" reported in HWInfo64 during CB R23 Multi?
The voltage reported by CPU-z isn't accurate as its source is SuperIO. Either it's reporting nonsense or something is off with the board's VRM settings.

The screenshot was idle on CPU-Z. Under load it went up to 1.38-1.39v. Normally it shows 1.34v-1.35v when idle.

 

[PJVol]

The screenshot was idle on CPU-Z. Under load it went up to 1.38-1.39v. Normally it shows 1.34v-1.35v when idle.

That's what I'm talking about, for the AMD CPUs it should be the opposite, i.e. 1.30-1.33 under load.
It seems like asus initial load line setting is too agressive, so your temps and power look a bit high for the 5600.
But I'd first doublecheck the voltages from the VRM telemetry (marked green on the HWInfo screenshot)

 

[IntelUser2000]

Sorry to bother you but I have to ask:

Would you be kind enough to provide some examples of linear scaling benchmarks and some non-linear scaling (hence good) benchmarks?

Dhrystone, Geekbench are popular examples of linear scaling. Dhrystone for example scales linearly because it fits entirely in the L1 cache. Great for testing out the capabilities of the execution units but what's the point when you need the things *around* the execution units to make it work? It's not like you buy just the ALU!

Cinebench is pretty decent. Yea every other generation it favors one architecture over the other but you can't avoid that.

There's also the case of scaling too low in regards to clock speeds. Games are an example. We all know performance in games can be very different from applications. HPC workloads are also another example. It requires too much memory bandwidth so scaling is often 50-60% and again that's not good.

SPEC is decent single benchmark but you need to take into account the limitations like compilers used and how different the subtests are compared to your workload. Like most agree .gcc portion is extremely difficult to cheat or optimize abnormally for.

You also need to take into account what SPEC you are running. SpecRate is garbage for client and probably for lots of enterprise. Rate scales nearly linearly with cores and processors. SpecFP is not at all relevant to client but seems pretty decent for HPC. When I say decent I am talking about SpecInt portion running single thread. It has a scaling rate similar to Cinebench at about 85%.

Again, no single benchmark will work well. You need extensive testing for a reason, so it can expose weaknesses of a particular architecture and also it's strengths.

 

[nicalandia]

Dhrystone, Geekbench are popular examples of linear scaling. Dhrystone for example scales linearly because it fits entirely in the L1 cache.

But Geekbench really poor MT scaling after 32C/64T.

 

[nicalandia]

Also..

32 Core ThreadRipper 5975WX vs Fastest 32 Core Genoa so far

PEGATRON Lightning vs LENOVO 30E0CTO1WW - Geekbench

browser.geekbench.com