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!

 

[Bigos]

Cinebench only tests rendering performance and only using one software stack. I myself don't care much about 128-bit vector performance with small working set.

 

[Hitman928]

If your goal is only to see the ideal maximum performance, you have Cinebench for that. But that is further from real world expectation than GB.

Depends on your use case. I don't do rendering but I bought a 12c/24t CPU for a reason and for me, Cinebench is a much better representation of performance than GB6.

 

[moinmoin]

If your goal is only to see the ideal maximum performance, you have Cinebench for that. But that is further from real world expectation than GB.

Completely disagree. Cinebench is a bad test for benchmarking MT since while it scales embarrassingly it also doesn't depend on the memory subsystem, core-to-core latency etc. so reflects too much of an ideal case hiding shortcomings of many chips with huge amount of cores. GB6 on the other hand is a bad suite for benchmarking MT since it tests the MT performance of limited isolated workloads, not what a chip is capable of MT performance when running multiple workloads in parallel or an embarrassingly parallel workload. So in GB6's case "MT" is a complete misnomer.

 

[Heartbreaker]

Completely disagree. Cinebench is a bad test for benchmarking MT since while it scales embarrassingly it also doesn't depend on the memory subsystem, core-to-core latency etc. so reflects too much of an ideal case hiding shortcomings of many chips with huge amount of cores. GB6 on the other hand is a bad suite for benchmarking MT since it tests the MT performance of limited isolated workloads, not what a chip is capable of MT performance when running multiple workloads in parallel or an embarrassingly parallel workload. So in GB6's case "MT" is a complete misnomer.

No one MT benchmark will be applicable for everyone.

You should look for benchmarks on your specific applications if you want some kind of result that applies specifically to you.

 

[Hail The Brain Slug]

I can only hardly imagine this to be an ASUS feature - I'd eat a broom stick if it was.

Will you record video and post it on the forums? Wood, metal or plastic?

 

[BorisTheBlade82]

Will you record video and post it on the forums? Wood, metal or plastic?

The Haribo style ones 😁

 

[moinmoin]

No one MT benchmark will be applicable for everyone.

That's not the point, and you know it.

GB6's "MT" benchmark is perfectly suited to people who buy processors to then only run one isolated workload at a time. How much do you think this applies to people buying processors with many cores?

 

[Heartbreaker]

That's not the point, and you know it.

GB6's "MT" benchmark is perfectly suited to people who buy processors to then only run one isolated workload at a time. How much do you think this applies to people buying processors with many cores?

That doesn't even make sense. No benchmark covers people running multiple different applications at the same time. If it did it would be even more niche to that specific multiple application workload.

 

[moinmoin]

That doesn't even make sense. No benchmark covers people running multiple different applications at the same time. If it did it would be even more niche to that specific multiple application workload.

Nearly nobody even cares about the specifics, about the make up of benchmarks, whatever useless niche they are (Cinebench is a workload that'd be better suited to GPU compute after all). It all in the score they produce. And GB6's "MT" score has little to nothing to do with a processor's multi-threading capability, only when the lack of cores is a serious bottleneck the score becomes telling.

 

[Heartbreaker]

And GB6's "MT" score has little to nothing to do with a processor's multi-threading capability, only when the lack of cores is a serious bottleneck the score becomes telling.

Again, that makes no sense. The more cores you have the higher the GB MT score. It's a de Facto MT benchmark.

Just because not every sub-component is embarrassingly parallel, doesn't change that.

 

[Markfw]

Again, that makes no sense. The more cores you have the higher the GB MT score. It's a de Facto MT benchmark.

Just because not every sub-component is embarrassingly parallel, doesn't change that.

For a 256 thread dual EPYC 7763 Milan to score less on an MT score than a 14c/20t 13600k, it means the MT benchmark is crap, and worthless.

 

[Heartbreaker]

For a 256 thread dual EPYC 7763 Milan to score less on an MT score than a 14c/20t 13600k, it means the MT benchmark is crap, and worthless.

That does sound terrible. Though when I searched the results for GB6 the only Epyc 7763, were in virtual machines and were only allowing 4 cores...

 

[Hans Gruber]

I think AMD needs to speed up their CPU release cadence. Zen 4 is kind of a bust. AMD processors are subject to needing fast memory whereas Intel is less dependent on memory timings and frequency for performance. AMD has a serious problem when they say that Zen 4 (comments before Zen 4 as released) is a memory OC'ers dream. They can't even run 6400mhz stable. 8000mhz already exists in the wild. Intel Raptor Lake systems run DDR5 memory at 7000mhz+ frequencies without issues.

Zen 5 cannot come soon enough. The other problem is motherboard pricing. Demand is in the gutter and the pricing of AMD motherboards is not following the principles of supply and demand.

 

[Mopetar]

Zen 4 is fine, but Zen 3 was an incredible swan song for the AM4 platform and a lot of AMD customers are still upgrading to Zen 3 because AMD gave them an upgrade path. If I had bought in at Zen 2, I'd probably snag a 5800X3D now that it's had a price cut and run with that for several years while both AM5 and DDR5 mature and come down in price.

AMD's design means that memory is a bit wonky due to the way it has to work with the infinity fabric. There's an area of memory speed where it's not going to gain results because it doesn't line up with the fabric speed. Theoretically if the memory were even faster, it would align well enough again where you could see better results.

Releasing Zen 5 before it's ready is pointless and AMD would be fools to try rushing a release. There's nothing wrong with Zen 4 and for many workloads or use cases it's the best product on the market. Even in gaming it's comparable with Intel now and with the v-cache parts releasing soon it will likely be the top choice in that market segment as well.

They should take time to work the kinks out and ensure that when Zen 5 does ship that it's a smooth rollout with minimal headaches for early adopters.

 

[Dave3000]

I think AMD needs to speed up their CPU release cadence. Zen 4 is kind of a bust. AMD processors are subject to needing fast memory whereas Intel is less dependent on memory timings and frequency for performance. AMD has a serious problem when they say that Zen 4 (comments before Zen 4 as released) is a memory OC'ers dream. They can't even run 6400mhz stable. 8000mhz already exists in the wild. Intel Raptor Lake systems run DDR5 memory at 7000mhz+ frequencies without issues.

Zen 5 cannot come soon enough. The other problem is motherboard pricing. Demand is in the gutter and the pricing of AMD motherboards is not following the principles of supply and demand.

If Ryzen CPU's benefit greatly from faster memory, wouldn't a 7950X and 7900X be much better in some games where memory read performance is critical than a 7700X and 7600X since they have much better memory read bandwidth than the 7700X and 7600X as shown in AIDA64's memory benchmark? Also if there is a situation where DDR5-6000 is giving me better performance than DDR5-5200 on a Ryzen 7700X would that mean that that application would benefit from the higher memory read bandwidth of the 7950X and 7900X?

 

[Hans Gruber]

If Ryzen CPU's benefit greatly from faster memory, wouldn't a 7950X and 7900X be much better in some games where memory read performance is critical than a 7700X and 7600X since they have much better memory read bandwidth than the 7700X and 7600X as shown in AIDA64's memory benchmark? Also if there is a situation where DDR5-6000 is giving me better performance than DDR5-5200 on a Ryzen 7700X would that mean that that application would benefit from the higher memory read bandwidth of the 7950X and 7900X?

They said Zen 4 was good up to DDR5 6400mhz but may are struggling to get above DDR5 6000mhz. Then they argue lower latency is more important than higher memory frequency. Yes, I agree but when you get memory frequency up near 8000mhz, bandwidth wins over tight timings.

Maybe AMD is intentionally crippling the bios with regards to memory frequency. The other possibility is they need to fix the bios to support memory speeds beyond 6400mhz. It makes no sense to me that they would not support fast memory speeds considering Zen 4 benefits from it. Zen 4 benchmarks with slow memory timings and low frequency really take a big hit in performance vs. Intel.

Let the Zen 4 experts here explain it. I just cannot understand how 6400mhz would be the top and few can achieve stable 6400mhz systems right now. B-Die DDR4 kits smoke what Zen 4 DDR5 kits are able to do right now.

 

[In2Photos]

I think AMD needs to speed up their CPU release cadence. Zen 4 is kind of a bust. AMD processors are subject to needing fast memory whereas Intel is less dependent on memory timings and frequency for performance. AMD has a serious problem when they say that Zen 4 (comments before Zen 4 as released) is a memory OC'ers dream. They can't even run 6400mhz stable. 8000mhz already exists in the wild. Intel Raptor Lake systems run DDR5 memory at 7000mhz+ frequencies without issues.

Zen 5 cannot come soon enough. The other problem is motherboard pricing. Demand is in the gutter and the pricing of AMD motherboards is not following the principles of supply and demand.

Let's try this again. Zen4 gains more performance going from 4800MHz base timings to tuned 6000MHz than Intel does going from 4800MHz to 7200MHz. Where's the problem?

 

[Dave3000]

They said Zen 4 was good up to DDR5 6400mhz but may are struggling to get above DDR5 6000mhz. Then they argue lower latency is more important than higher memory frequency. Yes, I agree but when you get memory frequency up near 8000mhz, bandwidth wins over tight timings.

Maybe AMD is intentionally crippling the bios with regards to memory frequency. The other possibility is they need to fix the bios to support memory speeds beyond 6400mhz. It makes no sense to me that they would not support fast memory speeds considering Zen 4 benefits from it. Zen 4 benchmarks with slow memory timings and low frequency really take a big hit in performance vs. Intel.

Let the Zen 4 experts here explain it. I just cannot understand how 6400mhz would be the top and few can achieve stable 6400mhz systems right now. B-Die DDR4 kits smoke what Zen 4 DDR5 kits are able to do right now.

Ryzen 7000 series only officially supports up to DDR5-5200. There is no guarantee that past DDR5-5200 will work or be stable on a random Ryzen 7000 series CPU and that's down to the silicon lottery. On 13th gen Intel CPUs it DDR5-5600 for DDR5 and DDR4-3200 for DDR4 and likewise no guarantee that it will work or be stable on a random 13th gen CPU with faster memory than that. Personally when I buy memory for my system, I only buy the official max supported speed that the CPU supports, and in my case I purchased DDR-3200 for my Ryzen 5800X even though my motherboard supports faster speed memory and even though there are people that said that they have no issues with DDR4-3600 on this CPU.

 

[Hans Gruber]

Let's try this again. Zen4 gains more performance going from 4800MHz base timings to tuned 6000MHz than Intel does going from 4800MHz to 7200MHz. Where's the problem?

I want to climb Everest 8000mhz but K2 is 7200mhz which is where Intel is at. I guess we are stuck at Mt Fuji with 6000mhz. So in a Zen 4 world. 6000mhz is still the most beautiful mountain in the world at 6000mhz but nowhere close to Everest in elevation.

Agreed AMD gains much more with higher frequency and lower timings than Intel. Which begs the question of why Zen 4 cannot run DDR5 at 7200mhz currently. I simply want to know why the AMD bios will not support memory beyond 6000-6400mhz.

 

[DrMrLordX]

I think AMD needs to speed up their CPU release cadence. Zen 4 is kind of a bust.

Genoa disagrees with you. Desktop is an afterthought.

 

[Dave3000]

I want to climb Everest 8000mhz but K2 is 7200mhz which is where Intel is at. I guess we are stuck at Mt Fuji with 6000mhz. So in a Zen 4 world. 6000mhz is still the most beautiful mountain in the world at 6000mhz but nowhere close to Everest in elevation.

Agreed AMD gains much more with higher frequency and lower timings than Intel. Which begs the question of why Zen 4 cannot run DDR5 at 7200mhz currently. I simply want to know why the AMD bios will not support memory beyond 6000-6400mhz.

DDR5-7200 is way faster than the official 5200 speed supported by Zen 4, so I'm not surprised. Personally, if I upgrade to a Ryzen 7950X system today, I'd purchased DDR5-5200 JEDEC memory or if I upgraded to a 13900K system, I'd purchase DDR5-5600 JEDEC memory. I prefer a reliable system even if it means losing 5-10% performance in gaming due to slower memory (not going past the official maximum supported memory speed of the CPU). I'm not saying that going faster than the maximum official supported memory speed will make my system unreliable, it's that I'm risking having reliability issues as far as the memory subsystem is concerned and prefer not to take that risk for the 5-10% better gaming performance.

 

[Markfw]

I want to climb Everest 8000mhz but K2 is 7200mhz which is where Intel is at. I guess we are stuck at Mt Fuji with 6000mhz. So in a Zen 4 world. 6000mhz is still the most beautiful mountain in the world at 6000mhz but nowhere close to Everest in elevation.

Agreed AMD gains much more with higher frequency and lower timings than Intel. Which begs the question of why Zen 4 cannot run DDR5 at 7200mhz currently. I simply want to know why the AMD bios will not support memory beyond 6000-6400mhz.

Its not the bios, its the memory controller and the IF in the CPU that are stopping it from going higher. If you look at the post below, you will see that with the right timings and memory Zen 4 can equal 13900k in gaming, which is currently the only weak spot for Zen 4.

Page 483 - Discussion - Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Page 483 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

 

[itsmydamnation]

They said Zen 4 was good up to DDR5 6400mhz but may are struggling to get above DDR5 6000mhz. Then they argue lower latency is more important than higher memory frequency. Yes, I agree but when you get memory frequency up near 8000mhz, bandwidth wins over tight timings.

This makes Zero sense ......

what is going on in this thread........

 

[Hans Gruber]

This makes Zero sense ......

what is going on in this thread........

Clock for clock. Cas 14 latency DDR4 @ 3800mhz would be equal to Cas 28 DDR5 7600mhz if you double the speed. Obviously DDR5 has some raw performance gains in architecture. Even if you have E die 16Cas @ 3800DDR4 would be Cas 32 DDR5 7600mhz.

I am equaling out the performance to give you an idea of the latencies of top tier DDR4 vs. DDR5.

My point is that AMD needs much higher frequency support for DDR5.

 

[Markfw]

Clock for clock. Cas 14 latency DDR4 @ 3800mhz would be equal to Cas 28 DDR5 7600mhz if you double the speed. Obviously DDR5 has some raw performance gains in architecture. Even if you have E die 16Cas @ 3800DDR4 would be Cas 32 DDR5 7600mhz.

I am equaling out the performance to give you an idea of the latencies of top tier DDR4 vs. DDR5.

My point is that AMD needs much higher frequency support for DDR5.

You can't just double all the numbers, it does not work like that. Your posts are approaching trolling.