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!

 

[reb0rn]

While your logic does have some merit, There are 1,000 variables out there, like core to core and core to small core latency with the interconnects, memory controllers, etc...

And of course how fast will Zen 4 really go ? and how long will Raptor lake be able to sustain full speed due to power and heat ???

Just for power alone (from what I read) I will take a Zen 4, unless things do not turn out anything like we have been reading.

The next month or so is going to be epic (or should I say EPYC)

And you know 7950x will also use almost same power to run 5.5Ghz as Raptor lake, then valid question is also same for both???

 

[nicalandia]

That's about right. Someone on these forums did a head to head comparison between the 12900K and the 5950X as you ramp up the thread count, and it showed that at certain thread counts, having half as many big cores was better than more E cores. Note, it was just based on approximations and not based on a benchmark.

Cinebench is one those apps that love e core SPAM..!



Extrapolating this to Raptor Lake(speed of P cores and e cores and IPC performance) that 2P + 32E score would be about 38-40,000 points matching/exceeding the actual 13900K wich is 8P + 16E

 

[nicalandia]

And you know 7950x will also use almost same power to run 5.5Ghz as Raptor lake, then valid question is also same for both???

Not at all. 7950X will be much more energy efficient than Raptor Lake, do doubt about it.

 

[Abwx]

And you know 7950x will also use almost same power to run 5.5Ghz as Raptor lake, then valid question is also same for both???

13900K will use 250W to match the 7950X@154W.

Edit : 13900K will be competitive in perf/watt against the 5950X, same way as a moderatly clocked 12900K can be competitive up to a 5900X when it comes to perf/Watt.

 

[Markfw]

16 e cores are more or less equal to 8c/16T P cores but that will depend on apps, but most of the time they are somewhat less

Somebody calculated this. Not sure if it was in AMD or Intel thread, but it was ONE P-core=1.8 E-core., so yes, thats what I think based on what I have seen.

 

[Markfw]

And you know 7950x will also use almost same power to run 5.5Ghz as Raptor lake, then valid question is also same for both???

Ahhhh, NO. 170 watt <> 350 watt. That is what I have seen.

 

[Abwx]

Somebody calculated this. Not sure if it was in AMD or Intel thread, but it was ONE P-core=1.8 E-core., so yes, thats what I think based on what I have seen.

That s at same frequency, if the P core run at 5Ghz with the e core at 4GHz then the ratio is 2.16 e cores for 1 P core.

 

[Markfw]

That s at same frequency, if the P core run at 5Ghz with the e core at 4GHz then the ratio is 2.16 e cores for 1 P core.

Well, I guess we will see.

 

[reb0rn]

Ahhhh, NO. 170 watt <> 350 watt. That is what I have seen.

So you say 5.5Ghz 16/32 full hard load core AMD will use no more then 170W where intel use ~300W (real use) I do not care for TDP rating and rest...

Someone memorise this for first user bench review, as I for long trust none of paid shills most review sites are

 

[deasd]

7600X Geekbench5 ST score around 2032 according to hint below

https://twitter.com/x/status/1563367468944330753

 

[Det0x]

7600X Geekbench5 ST score around 2032 according to hint below

https://twitter.com/x/status/1563367468944330753

Nice find

1613 * 1.26 = 2032
Lay of the land:

 

[uzzi38]

Worth noting that AVX512 is used for Geekbench as a part of the crypto score. It only accounts for (I think?) 5% of the final score, but still worth noting that it is there and will have an impact.

 

[cellarnoise]

I've been searching a bit and can't seem to find how often programs can run freely as a single core. Can someone point out how valuable single core bench is compared to two core or more in most real world apps including games?

Doesn't O.S., other background, and fetch from NVM memory /storage consistently suck up 20 to 50+% of a thread worth of cpu? Seems performance and frequency of 2 threads on two separate cores would be more valuable than an artificial ST benchmark anymore? But might not be as energy efficient?

 

[tamz_msc]

Doesn't O.S., other background, and fetch from NVM memory /storage consistently suck up 20 to 50+% of a thread worth of cpu?

Fire up task manager and see for yourself how much a hardware thread is used at idle.

 

[DrMrLordX]

By being born in the next door Slovakia.

If you married one of your national neighbors, would that be Czechmate?

(sorry, had to do a tasteless Czech joke)

On Geekbench Raptor Lake is in for a Beating. The e cores latencies and lack of AVX512 will be against them.

Geekbench kinda sucks though, so aside from obsessing over leaks, who really cares?

Intel is increasing prices up to 20%. That's kind of vague and surely will vary depending on SKU.

My suspicion is that the actual Raptor Lake SKUs (13600k and up) will see the biggest price increases. The rehashed Alder Lake products will maybe see a smaller bump versus previous gen. Despite schlepping an older uarch, the lower-end 13000-series chips are still selling more die area to the customer at a given product point.

AMD is likely pricing against the "real" Raptor Lake since they aren't really trying to compete on the low end anyway.

 

[Gideon]

7600X Geekbench5 ST score around 2032

Hmm Looks like my estimate of 2100 points for 7950X was pretty accurate:

Page 48 - Question - Raptor Lake - Official Thread

Page 48 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[Kaluan]

Hmm Looks like my estimate of 2100 points for 7950X was pretty accurate:

Page 48 - Question - Raptor Lake - Official Thread

Page 48 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

Given what we know about 7600X and 7950X clocks (boost/fmax), 7950X might score 2.200+, based on this 7600X leak.

But I suppose conservative estimates should the go-to with these things. You get the possibility of being pleasantly surprised later on too 😂

 

[PJVol]

Given what we know about 7600X and 7950X clocks (boost/fmax), 7950X might score 2.200+, based on this 7600X leak.

Extrapolating 5950X score from the linked browser.geekbench page, 7950X is more likely to score ~2150.

Doesn't O.S., other background, and fetch from NVM memory /storage consistently suck up 20 to 50+% of a thread worth of cpu? Seems performance and frequency of 2 threads on two separate cores would be more valuable than an artificial ST benchmark anymore?

No and no.
For the real world scenarios this view is overly simplistic. Instead what really matters is the cores' CC0 residency (active state) and there's no other way to know dNcores-loaded/dt
You can watch the "Average Active Core Count" metric in HWInfo64 panel for that purpose.

 

[nicalandia]

Worth noting that AVX512 is used for Geekbench as a part of the crypto score. It only accounts for (I think?) 5% of the final score, but still worth noting that it is there and will have an impact.

It's more than 5% specially on MT score. That is why Ice Lake show very strong MT when compared to AVX512 Gimped Alder Lake.


12900K 8P Only on Linux vs Ice Lake Xeon(8 Cores) AVX512 Performance

 

[Mopetar]

Ahhhh, NO. 170 watt <> 350 watt. That is what I have seen.

I know what you meant, but I think it's strictly true that 170 W < 350W. No need for the extra > in there.

 

[Markfw]

I know what you meant, but I think it's strictly true that 170 W < 350W. No need for the extra > in there.

Well, he said they would be equal, so this is an exact refutal as NO they will not be equal

 

[eek2121]

And you know 7950x will also use almost same power to run 5.5Ghz as Raptor lake, then valid question is also same for both???

So you say 5.5Ghz 16/32 full hard load core AMD will use no more then 170W where intel use ~300W (real use) I do not care for TDP rating and rest...

Someone memorise this for first user bench review, as I for long trust none of paid shills most review sites are

Yes, Intel chips will use significantly more power than AMD chips for the same performance.

They do not have the same amounts of threads ^^

The comparisons should be between the following cpus eventho some of the guys from IDL is scared of the 7950x and only want to compare 13900vs7900 and whatnot since "name'ing":

• 13900k(8P+16E 32t) vs 7950x(16c 32t)
• 13700k(8P+8E 24t) vs 7900x(12c 24t)
• 13600k(6P+8E 20t) vs 7700x(8c16t)
• 13500(6P+8E 20t)
• 13400(6P+4E 16t) vs 7600x (6c12t)

This could possibly mean that its on the low-end intel will do the best vs Zen4 this generation, if you only play Cinebench 24/7
Pricing will decide the winner below the 13700k i guess

Do note 13600, 13500 and 13400 is based on (refreshed) Alder Lake silicon, not the new Raptor lake
View attachment 66646

It was a typo. I was dead tired when I wrote that. Should have been comparing the 13700k.

It is not that simple. First you have 8 "P" cores with hyperthreading for both. I suspect RL big core will be a bit faster than Zen 4, but for the sake of argument, lets call them equal.
So 8 big Zen cores vs 8 big RL cores, call it a wash.
So you have left:
8 big Zen cores vs 8 E cores -- obvious Zen wins
But you still have 8 real E cores vs 8 Zen threads from hyperthreading. Advantage should be to Intel.
So the final performance will depend on how 16 real E cores will perform against 8 hyperthreaded Zen cores.

Yes, except the AMD chips have a base clock of 4.5-4.7ghz according to leaks, which is much higher than the base clock of the P cores (3 GHz). This means that AMD will be able to sustain much higher clocks at maximum power. If AMD's cores are roughly equal to Intel's P cores in terms of performance, and AMD's cores clock much higher in a worst-case scenario (the base clock), that tells me AMD will ultimately win that fight. For the "E" cores, the clocks are rumored to be increasing over the 12900k. WCCFTech claims 4.3 all core/4.7 single core boost. However, Zen 4 again wins here. Not only will the core be over 2x as fast, but the base clock is again higher than the all-core boost of the "E" cores.

I will honestly be shocked if AMD is not competitive in benchmarks, and it wouldn't surprise me if they end up having the fastest chip. It also will not surprise me if they lose in the mid-range. Their mid-range offerings have always been a bit 'meh' in terms of performance.

Regardless, it will be an exciting time for hardware enthusiasts. We have more competition now than ever in both the CPU and GPU space. It will be great.

 

[Kaluan]

Extrapolating 5950X score from the linked browser.geekbench page, 7950X is more likely to score ~2150.

Yeah maybe, they're same-y enough for me to not really care that much.

But regardless, my speculation was based around averaging rumoured boost clocks (5,3 and 5,7... 7,4%) with deduced fmax clocks (5,35 and 5,85... 9,3%), which comes out as 8,4% higher clocks and multiplying that with the rumoured 2032 of 7600X, which comes up as ~2203 points but anyway.

 

[nicalandia]

7600X Geekbench5 ST score around 2032 according to hint below

https://twitter.com/x/status/1563367468944330753

I am more interested on MT Performance. The AVX512 performance will tip the balance to the 7950X part(as in doubling the AES-XTS Performance of the 13900K)

Because I fully expect the High Clock 32T Zen4 CPUs to beat/match Slow Clock 32T Ice Lake Xeons on AVX512 and the 13900K with it's 32T even at High Speeds can't hold a candle to those slow xeons at AVX512 performance

 

[Joe NYC]

After @Kepler_L2 first tweeted that RDNA 3 chiplets might have stacking of SRAM, and SkyJuice confirmed it, there was one tidbit in SkyJuice piece, saying that AMD could have 2 layers of stacking of SRAM.

I was wondering why no one caught up on this, since it could have wide implications.

And just today, this is what Graymon posted:

https://twitter.com/x/status/1563502540183875585

He says it can apply to both CPU and GPU.