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!

 

[Makaveli]

Great, that means the 40% MT Boost is correct because ALL Cores will be boosting 5Ghz+

Yup not a problem at all.

 

[tamz_msc]

Great, that means the 40% MT Boost is correct because ALL Cores will be boosting 5Ghz+

Downside being that efficiency goes down the drain.

 

[nicalandia]

Downside being that efficiency goes down the drain.

It will still be better than ADL/RTL

 

[tamz_msc]

It will still be better than ADL/RTL

Barely.

 

[Timmah!]

IIRC there's no rendering project(Blender, vray, corona) support AVX512, since this instruction draw too much power/heat, and maybe it's the reason why disable it in ADL which is power limited. Very interesting to see how will AMD do with AVX512 with superior process.

Yup , i wonder whether AMD will have to downclock as hard, or at all, as Intel CPUs (like Skylake-X) had to, when running AVX-512 code.
Its rather amusing how things moved from Intel pushing AVX-512 on their HEDT line-up, while AMD being no-show in that regard, to situation, when the only consumer AVX-512 capable CPUs will be AMD ones.

Anyway, i am unaware of any mainstream application that would be using it.

 

[nicalandia]

Barely.

A Win is a Win, by an Inch or a mile(in this case by Watts)

 

[Anhiel]

You're also using wrong logic on this, as you ignore the difference between time and rate.

To check your formula assume Zen4 finishes a benchmark in 150s and 12900K does the same bench in 300s. According to you the calculation should be 150/300=0.5 => 1-0.5=50% However, you can clearly see from the numbers that a CPU that finishes a workload in half the time is 100% faster, not 50% faster.

The correct way is to translate the time numbers into rate or speed.

• In one second Zen 4 completes R1=1/204 of the bench.
• In the same one second the 12900K completes R2=1/297 of the bench.

If we want to see how much faster Zen4 is, we can now compare the rates R1/R2=(1/204)/(1/297)=297/204=1.45

Nice, but your rate evaluation is pointless as you can see in your equation yourself they the rate mathematically elimiate each other out. You got the same results. lmao

 

[tamz_msc]

Anyway, i am unaware of any mainstream application that would be using it.

Anything that incorporates MKL, from Anaconda distribution of Python, to R, and mathematical software like Matlab and Mathematica, will be able to use it. All of these are mainstream applications.

 

[Henry swagger]

Great, that means the 40% MT Boost is correct because ALL Cores will be boosting 5Ghz+

Lets wait for benchmarks and amd to present it before assuming.. remember 37% single core performance ?

 

[Ranulf]

Duh?

It's obviously half as faster!

But does it have what plants crave?

Here is my shot at Intel. they dropped hyperthreading in the 9700k not because they didn't think they needed it. They dropped it because of the spectre issue. People asked where the hyperthreading went. Intel said, you don't need hyperthreading not answering the reason why it disappeared for a CPU generation. Then they dropped AVX-512 from Alder Lake because the hybrid structure would not work with it.

This is getting off-topic IMO but I would enjoy a seperate thread to hear more about the theory.

 

[deasd]

Yup , i wonder whether AMD will have to downclock as hard, or at all, as Intel CPUs (like Skylake-X) had to, when running AVX-512 code.
Its rather amusing how things moved from Intel pushing AVX-512 on their HEDT line-up, while AMD being no-show in that regard, to situation, when the only consumer AVX-512 capable CPUs will be AMD ones.

Anyway, i am unaware of any mainstream application that would be using it.

Geekbench has AVX512 support and being a composition of GB score.
PlayStation 3 Emulator, RPCS3 utilize AVX512 would provide 15-20%+ FPS depends on which game.
Corel Painter utilize AVX512 for photo editing.
That's what I remember.

 

[nicalandia]

Lets wait for benchmarks and amd to present it before assuming.. remember 37% single core performance ?

But it was already confirmed that the reason for Higher TDP is to boost MT Performance and that the 5950X was held back by that power constrain

 

[Anhiel]

I see many cannot do maths and do not understand speed = distance / time.

How much faster is an 8.83ms frame time vs a 16.67ms frame time?

I presume you meant it like this
8.83/16.67=0.5 => 1/0.5=2x faster
This is correct in terms of x-times faster but not in terms of percent as being discussed.

 

[Makaveli]

I presume you meant it like this
8.83/16.67=0.5 => 1/0.5=2x faster
This is correct in terms of x-times faster but not in terms of percent as being discussed.

We need to stay on topic please stop responding to post about math.

 

[biostud]

Uh oh....

https://www.tomshardware.com/news/amd-corrects-socket-am5-for-ryzen-7000-power-specs-230w-peak-power-170w-tdp

Or simply future proofing for 16c+ CPUs

 

[nicalandia]

Robert Hallock (AMD Director of Technical Marketing) on Why the bumped the TDP Rating

"We did this to mainly improve multi-thread performance as many of the core count chips were actually held back in overall compute performance by relatively modest socket power"

 

[leoneazzurro]

Barely.

Barely? Even if the 12900K and the hypothetical 7950X had the same power consumtpion (but 12900K has a PPT of 240W and the peak power consumption of the AM5 socket is 230W) the CPU in AMD's test was 45% faster in Blender. So that would translate (if power consumtion was the same) a 45% perf/W advantage in that particular test. Granted, different tests may have a lower ratio and RPL may change things (at the same process node it's unlikely to dramatically alter the situation, though), but "barely" compared to ADL it's not an understatement. It's completely wrong.

 

[uzzi38]

Uh oh....

https://www.tomshardware.com/news/amd-corrects-socket-am5-for-ryzen-7000-power-specs-230w-peak-power-170w-tdp

Waiting on a response if that applies to the demo itself. The clarification is worded specifically to talk about the capabilities of the AM5 socket, not the demo, whereas Robert Hallock/AMD marketing have said 170W PPT multiple times for the demo already.

 

[Anhiel]

Also, I remember someone commenting a day or two ago (which is probably like 300 messages back by now...) that Blender takes advantage of AVX-512, not sure if that point was discussed, but it doesn't really seem to be the case:

View attachment 62130

It came up because of the discrepancy between ST and MT.
I've heard AVX-512 is only used in one specific test, though.
I personally don't think this is the case because the discrepancy is too large (15% -> 31%) even with SMT.
AMD's SMT gain is bigger than Intel's arch design but still no more than 35%. So none of the numbers seem to fit any verifiable metrics.

Based on Ryzen 5 5600X 6c 3.7/4.6GHz 65W TDP and people's stock testing 4.65/4.85GHz
and presuming this engineering sample to be something along the lines of 7950X and 125W TDP to hit 170W PPT;
and considering node gains 15% clock speed or 30% power reduction (ratio 1:2)
and running at relative power limits
then back calculated for max turbo boost for 1-2 cores 5.52GHz I get 5.29GHz for all core boost. This seems to fit perfectly with the demonstration. That said it's also obvious the balance is nearly all going into speed here.

 

[uzzi38]

It came up because of the discrepancy between ST and MT.
I've heard AVX-512 is only used in one specific test, though.
I personally don't think this is the case because the discrepancy is too large (15% -> 31%) even with SMT.
AMD's SMT gain is bigger than Intel's arch design but still no more than 35%. So none of the numbers seem to fit any verifiable metrics.

Based on Ryzen 5 5600X 6c 3.7/4.6GHz 65W TDP and people's stock testing 4.65/4.85GHz
and presuming this engineering sample to be something along the lines of 7950X and 125W TDP to hit 170W PPT;
and considering node gains 15% clock speed or 30% power reduction (ratio 1:2)
and running at relative power limits
then back calculated for max turbo boost for 1-2 cores 5.52GHz I get 5.29GHz for all core boost. This seems to fit perfectly with the demonstration. That said it's also obvious the balance is nearly all going into speed here.

You're not accounting for the fact that the 5950X doesn't even get remotely close to it's boost clock in MT loads. In R20/R23, a stock 5950X sits closer to ~3.8GHz.

There is a LOT more headroom available in MT loads than ST loads.

 

[jpiniero]

Or simply future proofing for 16c+ CPUs

I'd say the bigger issue is that boards are going to be pricey like Intel.

 

[Hans Gruber]

But does it have what plants crave?



This is getting off-topic IMO but I would enjoy a seperate thread to hear more about the theory.

It's not a theory. It's a fact. Disabling hyperthreading on coffee lake processors prevented spectre from being a problem. When hyperthreading was enabled, spectre was a problem. The 9900K wasn't affected by Spectre. I can't remember why but as another poster here said. The 9700k doesn't need hyperthreading against the 2700x because it can compete just fine without it. That was an accurate quote but the reason why hyperthreading was disabled for the 9700K had to do with the spectre security issue.

I could point out the delusional belief that the B350/X370 boards could not run memory at 3733mhz or 3800mhz. For a long time they couldn't and it was a bios firmware issue that AMD fixed 4 or 5 years later. At least AMD fixed the memory issues on the original Ryzen AM4 boards. Intel's way of doing business, oh well wait until the next generation motherboards and CPU's and buy a new computer.

I still firmly believe that AMD still owes us long term customers. More than 10 years of darkness and misery with worthless AMD CPU's. I personally have a desk drawer full of my very old AMD CPU's from the golden era of AMD. Starting with the Athlon 1800+ to the dual Core AMD 64's.

I should point out that I currently have AMD Ryzen 5600. That processor was released in April 2022 and I am getting grief from some AMD fanboys in this thread. It should have been released in early 2021.

 

[Hans Gruber]

Robert Hallock (AMD Director of Technical Marketing) on Why the bumped the TDP Rating

"We did this to mainly improve multi-thread performance as many of the core count chips were actually held back in overall compute performance by relatively modest socket power"

The reason AMD would increase the TDP on the AM5 socket. To better handle 24core 48thread and 32core 64thread CPU's that AMD will inevitably release during the Zen 4 era at some point.

 

[uzzi38]

Waiting on a response if that applies to the demo itself. The clarification is worded specifically to talk about the capabilities of the AM5 socket, not the demo, whereas Robert Hallock/AMD marketing have said 170W PPT multiple times for the demo already.

Just FYI as to why we have to ask, there is a 125W processor group (w/170W PPT) as well. So uh... it's really not very clear ATM.

 

[Anhiel]

You're not accounting for the fact that the 5950X doesn't even get remotely close to it's boost clock in MT loads. In R20/R23, a stock 5950X sits closer to ~3.8GHz.

There is a LOT more headroom available in MT loads than ST loads.

My point was using the power limits as a proven reference for projection without adding any speculation. The 5600X was best simply fit on the metrics. Your 5950X is power limited by the 142W PPT that's why it can't get faster.