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!

 

[naad]

People forget not all benchmarks are either one core or "embarrassingly parallel" loads, Raptor lake has a whole 8 cores more than Raphael, any workload(plenty of those in PTS) not putting those 8 extra core to use will make the 13900k look like last gen perf compared to 7950x

tldr for workloads:

1-8 core = tie with 2-3% differences here and there
8-16 core = Raphael clearly stomps
16-24 core = mostly tie, around 2-5% difference again


IMO from the plethora of benchmarks I've seen around most of them usually stop scaling at around 10-12 cores

 

[Abwx]

People forget not all benchmarks are either one core or "embarrassingly parallel" loads, Raptor lake has a whole 8 cores more than Raphael, any workload(plenty of those in PTS) not putting those 8 extra core to use will make the 13900k look like last gen perf compared to 7950x

tldr for workloads:


16-24 core = mostly tie, around 2-5% difference again

Above 8T the 2-5% difference is only at 32T.

ST Zen 4 or RPL = 100.

8T will yield 800 for both RPL and Zen 4.

ST perf of small core is 60.

9 threads yield 900 for Zen 4 and 860 for RPL
10 threads is 1000 for Zen and 920 for RPL.

16T is 1600 for Zen and 1280 for RPL.

Above 16T the Zen 4 advantage will gradually decrease and tie is at 32T, the greater advantage is hence at 16T with 25% better throughput.

 

[inf64]

Also we have to take into account that small E cores have half the AVX2 width so anything that is using AVX and is multithreaded up to 32 threads is going to be hampered versus 32T Raphael that has 16 "fat" cores (32T) with full width AVX(2). Also, take into account the fact that E cores are going to be running at much lower MT boost clocks VS Zen 4 cores - all in all it's no surprise why 7950X will be overall superior product.

 

[Abwx]

I took account of the e cores lower frequencies but not of AVX2 throughput, all in all and on average it should be close to those numbers.

 

[Det0x]

AMD Zen 4 Memory Latency / Bandwidth

Seems to be diminishing returns above 2000 FCLK with static 1:1 UCLK 3000mhz and MCLK 6000MT/s *edit* see my new post below

Now we only need to find out how high we can run the UCLK and MCLK 1:1 with locked FCLK at 2000mhz

https://twitter.com/x/status/1571326977881313282

 

[naad]

The capped memory bandwidth kinda worries me for EPYC, but I guess anyone needing max bandwidth will be using NPS4 anyway, plus 12 channels is a lot more than the 8 the competition has

 

[inf64]

The capped memory bandwidth kinda worries me for EPYC, but I guess anyone needing max bandwidth will be using NPS4 anyway, plus 12 channels is a lot more than the 8 the competition has

Well, memory BW is capped because the chart shows the same DDR5 6000 numbers, other variables are not important for it.

 

[PJVol]

What is this "the competition" language, anything like this ?

Spoiler

"Imagine, the task force guy says, telling a passenger on arrival that a dildo
kept her baggage on the East Coast. Sometimes it's even a man. It's airline
policy not to imply ownership in the event of a dildo. Use the indefinite
article.
A dildo.
Never your dildo.
Never, ever say the dildo accidentally turned itself on.
A dildo activated itself and created an emergency situation that required
evacuating your baggage"
(c) 1996 Chuck Palahniuk

 

[Kaluan]

The capped memory bandwidth kinda worries me for EPYC, but I guess anyone needing max bandwidth will be using NPS4 anyway, plus 12 channels is a lot more than the 8 the competition has

Capped? What do you mean? That's not really a memory scaling chart, it's a (memory in relation to) FCLK scaling one.

If anything it just means FCLK = 2000 is enough for optimum use of 6000MTs DDR5... as we've already kinda known.

6400MTs (or more) + going beyond 2000MHz FCLK might be a different story. ("golden sample" talk aside).

Also Raphael IOD/IMC =/= Genoa IOD/IMC, I wouldn't bother speculating too much based on one for the other.

As some of the leakers and speculations said, people are overestimating how memory sensitive Zen4 is.

 

[naad]

As expected, Raptor Lake gaming nor MT results look like much of a threat to Raphael, that 700MHz ST gap between i9 and i5 will not do them any favors in the eye of gamers (along with no pcie5 M2 and dead end socket)

 

[naad]

Capped? What do you mean? That's not really a memory scaling chart, it's a (memory in relation to) FCLK scaling one.

If anything it just means FCLK = 2000 is enough for optimum use of 6000MTs DDR5... as we've already kinda known.

6400MTs (or more) + going beyond 2000MHz FCLK might be a different story. ("golden sample" talk aside).

Also Raphael IOD/IMC =/= Genoa IOD/IMC, I wouldn't bother speculating too much based on one for the other.

As some of the leakers and speculations said, people are overestimating how memory sensitive Zen4 is.

My mistake, didn't look properly, didn't see it was fclock scaling, sorry @inf64 as well

 

[deasd]

Now looks like the temperature is better than before.
7950X ~1.252v ~81.6C
7900X ~1.368v ~86.5C

https://twitter.com/x/status/1570944405800099840

https://twitter.com/x/status/1571424943077982209

 

[Det0x]

AMD Zen 4 Memory Latency / Bandwidth

Seems to be diminishing returns above 2000 FCLK with static 1:1 UCLK 3000mhz and MCLK 6000MT/s

Now we only need to find out how high we can run the UCLK and MCLK 1:1 with locked FCLK at 2000mhz
View attachment 67751

https://twitter.com/x/status/1571326977881313282

It's actually not diminishing returns, the graph is simply not using a consistent scale for the X-axis

Here's a corrected graph using the data points included. Note the gaping hole at 1900 MHz.
Looks close to linear for latency, while read and copy might just be logarithmic

 

[Hans de Vries]

Blender Open Data 3.3.0 (CPU-Windows)
https://opendata.blender.org/benchmarks/query/?compute_type=CPU&operating_system=Windows&blender_version=3.3.0&group_by=device_name (Thanks to HXL @9550pro)

 

[PJVol]

Looks like Zen3/4 six-core didn't fit )
Just curious of 7600X uplift, my 5600x scored 178 and 190, stock and tuned in 3.3, though still impressive for the 8-core parts and even more so for 16.

 

[Antey]

Remember that time when this guy richie rich was all over here with his IPC/PPC comparisons using just geekbench and such and you all hated him? well, i think it's interesting how different the situation is nowadays in that matter.

(1) Info - TOP 20 of the World's Most Powerful CPU Cores - IPC/PPC comparison | AnandTech Forums: Technology, Hardware, Software, and Deals

He estimated that the A14 would have 558 PPC and it tourned out to be close to 530, and it's funny because he estimated that the A15 would have 640 PPC! and it was once again 530. And now, the A16, even though this CPU isn't part of his estimations, but it's again, 540-545!

Four iterations (A13, A14, A15 and A16) and the apple big cores have increased their (geekbench) ''IPC'' just 8%.

And in the x86 world the Zen 4 now has 395-400 PPC! That's just 36-38% more ''IPC'' for the ARM world, and what about the X3? wasn't it 11% higher IPC than the previous core? what is it going to be? 460 PPC? a mere 15% better than their x86 counterparts?

at this rate...

Just noting that

 

[DrMrLordX]

Remember that time when this guy richie rich was all over here with his IPC/PPC comparisons using just geekbench and such and you all hated him? well, i think it's interesting how different the situation is nowadays in that matter.

It's really not worth stirring up that nonsense again. Everyone can see now what Apple has in mind for desktop/workstation, despite the fact that the performance of their cores on actual applications (rather than GB5) is somewhat muddied by the extensive use of fixed-function hardware. In any case, AMD is not worried about any such PPC comparisons, for good reason.

 

[Abwx]

Blender Open Data 3.3.0 (CPU-Windows)
https://opendata.blender.org/benchmarks/query/?compute_type=CPU&operating_system=Windows&blender_version=3.3.0&group_by=device_name (Thanks to HXL @9550pro)
View attachment 67763

By the look of the scores the 7950X/7700X/7600X have the same all core frequency.

Scores being proportional to TDP and assuming a worst case figure of 230W for the 7900X then the 7700X use about 115W and the 7600X is around 90W.

 

[Asterox]

As expected considering the global situation,..................................huh

AMD Ryzen 7000 series now listed in Spain, 7950X for 852 EUR - VideoCardz.com

AMD Ryzen 9 7950X for 852 EUR The prices of Ryzen 7000 for Spain are now known, reports Professional Review. Spanish prices for Ryzen 7000 are unfortunately high. AMD’s first AM5 processors will retail at 371 to 852 EUR depending on the model. While those prices are considerably higher than...

videocardz.com

 

[scineram]

I almost forgot about it. Who else nailed the poll of this topic? Or was it just me being brilliant?

 

[Kaluan]

Looks like Zen3/4 six-core didn't fit )
Just curious of 7600X uplift, my 5600x scored 178 and 190, stock and tuned in 3.3, though still impressive for the 8-core parts and even more so for 16.

It did ~229 (so should've been at the bottom), but apparently not in that version.
I wouldn't put much weight in comparison there, particularly across the different versions of Blender. I think they had runs in one of the versions were 4700G seemed faster than 5800X.
Anyway the general idea seems to be that 7600X is slightly faster than 5800X and 12600K in productivity (which also means > efficiency than both, but unclear if it can muster better efficiency than 5600X's 76W @ typical load).

My mistake, didn't look properly, didn't see it was fclock scaling, sorry @inf64 as well

Hehe, no worries mate. I was pretty confused of what the chart was actually showing at first too.

 

[nicalandia]

AMD Ryzen 9 7950X 16-Core Retail CPU Already Being Sold In China For $850 US

We are a week away from the launch of the AMD Ryzen 7000 CPUs but the flagship Ryzen 9 7950X is already being sold in China.

wccftech.com

Let see how many of those make their way to USA before the NDA is lifted

 

[Det0x]

AMD Ryzen 9 7950X 16-Core Retail CPU Already Being Sold In China For $850 US

We are a week away from the launch of the AMD Ryzen 7000 CPUs but the flagship Ryzen 9 7950X is already being sold in China.

wccftech.com

Let see how many of those make their way to USA before the NDA is lifted

Sadly its a papir weight without supporting motherboard.
I think and hope there will be good amounts of stock this Ryzen release..

 

[Markfw]

Sadly its a papir weight without supporting motherboard.
I think and hope there will be good amounts of stock this Ryzen release..

Lets hope so... I have memory, case, PSU HSF(AIO) and video card all ready to go. I just need CPU, motherboard and NVME. Waiting on that hoping for PCIE 5.0 compatible. If not PCIE 4.0 will do.

 

[igor_kavinski]

I just need CPU

Keep an eye on AMD shop like a hawk and swoop in on your prey the moment it rears its lovely head