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!

 

[nicalandia]

Most real world workstation uses have a substantial single or lightly threaded component, where Golden Cove's superior single thread performance pulls it ahead.

This poster would like to have a word with you about "Real World" Applications.

Page 213 - Discussion - Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Page 213 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

LibreOffice (Make), 5950X 35% faster
Unreal Engine (VS2019-MSBuild), 5950X 28% faster
Chromium (Ninja), 5950X 14% faster
Blender (CMake), 5950X 14% faster

 

[Markfw]

It's pretty ironic to complain about people ignoring facts when you first claimed that no such review exists, and then proceeded to not even read it when I gave you an example. Turns out not everything is Blender.

I never said a review did not exist. I said SHOW me. And its an outlier of course.

 

[Exist50]

This poster would like to have a word with you about "Real World" Applications.

Page 213 - Discussion - Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Page 213 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

LibreOffice (Make), 5950X 35% faster
Unreal Engine (VS2019-MSBuild), 5950X 28% faster
Chromium (Ninja), 5950X 14% faster
Blender (CMake), 5950X 14% faster

Those are compilation benchmarks, and the 5950X is faster in those, yes. But you're not seriously going to tell me that you don't think the Adobe suite, Davinci Resolve, etc. are not "real world applications", are you? Certainly they're more representative of typical workstation use than compiling a scattering of open source apps.

I never said a review did not exist. I said SHOW me. And its an outlier of course.

You didn't even read it, and yet immediately call it an outlier? How are Photoshop, Lightroom, etc. worse benchmarks than e.g. Blender, which isn't even run on the CPU in the real world? But we've been over this before. Your sole definition for non-gaming workloads is apparently embarrassingly parallel scientific compute.

 

[AtenRa]

No, it definitely cannot. Write the sentence again using 50% less time as basis of calculation.

If you take the context with time, then less time = faster and more time = slower

In that context of time, you can say that finishing the benchmark at 31% less time equals 31% faster that the competition.
After all, you need 31% less time to finish the same job and that equals for that CPU to be 31% faster as of using 31% less time.

Same applies for 50% less time, you just finished the same job at half the time or 50% faster in the context of time.

 

[Markfw]

Those are compilation benchmarks, and the 5950X is faster in those, yes. But you're not seriously going to tell me that you don't think the Adobe suite, Davinci Resolve, etc. are not "real world applications", are you? Certainly they're more representative of typical workstation use than compiling a scattering of open source apps.


You didn't even read it, and yet immediately call it an outlier? How are Photoshop, Lightroom, etc. worse benchmarks than e.g. Blender, which isn't even run on the CPU in the real world? But we've been over this before. Your sole definition for non-gaming workloads is apparently embarrassingly parallel scientific compute.

All of this is off-topic. I replied to one user who thinks Alder Lake rules number one in everything, and Zen 4 is coming late. Now you are arguing about Alder lake vs Zen3. DROP IT !

 

[nicalandia]

Those are compilation benchmarks, and the 5950X is faster in those, yes.

Please stay on Topic, I would not like to report you for derailing this thread.

 

[AtenRa]

No.

UK GCSE revision topic

In the UK going from 297s to 204s would always be described as a 31% reduction in time (or other synonyms like less time). If you were to use the term faster you either need to state it as Zen 4 is 93s faster than 12900K or you need to calculate the speed and compare on that basis which makes it 45% faster.

If I was to answer an exam question using the working of some people here, on reddit, over at TPU etc I would have failed GCSE maths and physics.

Did you not noticed I have changed the context to "less time = faster" and "more time = slower" ???

 

[Exist50]

All of this is off-topic. I replied to one user who thinks Alder Lake rules number one in everything, and Zen 4 is coming late. Now you are arguing about Alder lake vs Zen3. DROP IT !

I responded to the claims you made, and clearly thought were on topic. And it's going to translate to discussions about Zen 4 as well. Clearly we're seeing a substantial boost to multithreaded boost clocks in particular, but that will translate differently to different benchmarks. Quite notably, a long Blender run will see a greater gain than most actual workstation uses, so anyone trying to extrapolate those gains from the test AMD showed will be poorly served. But thanks to these two numbers (ST and long term MT), we can get a sense of best and worst case results.

 

[coercitiv]

In that context of time, you can say that finishing the benchmark at 31% less time equals 31% faster that the competition.
After all, you need 31% less time to finish the same job and that equals for that CPU to be 31% faster as of using 31% less time.

No you cannot! Really, what the heck is going on?! You're bending Math as if it's an urban dictionary.

Something that takes 50% less time is 100% faster, with no exception. Either you understand this and reconsider your position or we have no common ground to communicate on.

 

[Exist50]

No you cannot! Really, what the heck is going on?! You're bending Math as if it's an urban dictionary.

Something that takes 50% less time is 100% faster, with no exception. Either you understand this and reconsider your position or we have no common ground to communicate on.

This is why speedup is the preferred metric. In this case, S = T_old / T_new. I honestly am kinda confused why this is still even a discussion.

 

[biostud]

So:
Heavily treaded work software: AMD winner
Lightly threaded work software: Intel Winner
Gaming: Slight advantage to AMD, but pretty close on average.
Performance/watt gaming and lightly threaded work: Tie
Performance/watt full load all cores: AMD

And the buying advice is: Look at what software you use, and buy accordingly.

 

[nicalandia]

No you cannot! Really, what the heck is going on?! You're bending Math as if it's an urban dictionary.

This is really getting out of Hand, AMD did messed that Up on the Slide, this is not the only forum where people are getting confused about how much faster is that pre-production 16 Core sample vs 12900K

Just use this calculator and make up your mind

Percentage Calculator - percentagecalculator.guru

Percentage calculator to find percentage of a number, calculate x as a percent of y, find a number given percent. How to work out percentage formulas. Add and subtract percentages.

percentagecalculator.guru

 

[Mopetar]

In context of time , 31% less time can be translated in to 31% faster and yes you need to be 45,6% faster in performance in order to finish the benchmark in 31% less time

That doesn't work. Otherwise something that takes 100% less time (in other words something that completes instantly) would only be considered 100% faster, or twice as fast.

It's easy to see why people can become confused, but that doesn't mean the confusion should be encouraged.

 

[dnavas]

Those are compilation benchmarks, and the 5950X is faster in those, yes. But you're not seriously going to tell me that you don't think the Adobe suite, Davinci Resolve, etc. are not "real world applications", are you?

For workstation use, it seems to me that Premiere and Resolve are, Photoshop isn't. YMMV.

Certainly they're more representative of typical workstation use than compiling a scattering of open source apps.

Compilation is what I principally use my workstation for, and I frankly would like about twice the cores than I have and a qsfp28 link to storage. For video work, I could definitely use the cores, but I don't need the network upgrade. The problem with "certainly more representative" is that I frankly don't think we know to whom "workstations" are sold. Puget, of course, sells to the video editing crowd, which their benchmarks should make clear.

For video editing, quicksync should make a pretty significant difference in Alderlake's favor when comparing cpu to cpu. I would hope that AMD including decode logic in CPUs could help, but they'll need to round up support (and invest in VCE, because :urk: ). Anyway, that's kind of an interesting difference in that world, but it isn't "typical." Similarly for AVX512 support. Worse, I expect workloads are going to become increasingly specialized, which is going to make this sort of arguing kind of pointless, and 10% one way or the other isn't really likely to make a difference. Compounded velocity will make a winner, or neck-n-neck will make it more about price.

 

[deasd]

Let's back to topic. I don't want to join Zen3 vs Golden, 5950x vs 12900k debate...

DDR5-only factor makes the speculation about Zen4 even more complicated.... these comments/questions below are what I'm wondering too....

https://twitter.com/x/status/1529220869976883210

https://twitter.com/x/status/1529229372971880448

 

[nicalandia]

Let's back to topic. I don't want to join Zen3 vs Golden, 5950x vs 12900k debate...

DDR5-only factor makes the speculation about Zen4 even more complicated.... these comments/questions below are what I'm wondering too....

https://twitter.com/x/status/1529220869976883210

https://twitter.com/x/status/1529229372971880448

It's clearly obvious that AMD is Sand Bagging.

If you look how the Zen3+ Rembrandt with DDR5 at ISO Power it's getting a 32% performance boost for Floating Point in MT and 12% on Single Threaded apps


From Anandtech Rembrandt Benchmarks

"But the fact of the matter is, CPU performance is more than just the microarchitecture and frequency. Beyond that, it’s the memory subsystem, which also contributes directly to IPC or performance per clock. This is why I’ve gone off iso-frequency testing for this sort of comparison, because each product is built with optimization points in mind, and moving simply the core frequency causes a different balance of resources compared to the ‘as built’ and ‘as sold’ metrics. This is why when we put Zen3+ up against Zen3 at a similar power level, we’re seeing a sizable uptick in performance"

 

[JoeRambo]

DDR5 and new IOD will help a lot in MT. Zen3 has very large miss handler queues, like 20-30 requests per core can be pending and more parallelism from DDR5 will help serve those requests faster. Probably impact on the level of having dual-rank memory versus single rank, without impact on clocks and timings.

Also new IOD is very important too. Zen3 IOD Infinity fabric is not really infinite and can become a bottleneck when a few cores have traffic to IMC. So called "loaded" memory latency is rising faster than on Intel's stuff due to IF limitations and queues.

There is no doubt Zen4 will be a beast in MT, but they have to be, as they are going to compete with 8+16 => 8+32 in near future.

 

[AtenRa]

No you cannot! Really, what the heck is going on?! You're bending Math as if it's an urban dictionary.

Something that takes 50% less time is 100% faster, with no exception. Either you understand this and reconsider your position or we have no common ground to communicate on.

Explain to me how half the time is 100% faster ???

In the context of "less time = faster" and "more time = slower" , saying something is 100% faster means that it needs ZERO (0) time.
You essentially saying that If CPU A needs 100 secs , then CPU B that is 100% faster vs CPU A it needs 0 secs.


You cannot say that something that takes 50% less time is also 100% faster , faster in what context ???
50% less time =measured in time
100% faster = in what ???


You can say that 2 seconds is half the time of 4 seconds

That means that 2s is 50% less time than 4s and that 2s is 50% faster in seconds (less time)
Also, 4 seconds is double the time of 2 seconds, that means that 4 seconds is 100% SLOWER vs 2 seconds.

You can also say that in order to finish the job in half the time or 50% less time, you need double the performance or 100% higher performance.
But you cannot say that half the time or 50% less time is 100% faster, this is incorrect.

Here, this will help
Hans says that in order for the ZEN 4 to finish the benchmark in 31% less time, it has 46% higher performance (instructions/s)
You cannot say 31% less time is 46% faster, the performance of the CPU is 46% higher and that translates in to finishing the benchmark at 31% less time of 31% faster.

https://twitter.com/x/status/1528719223786192896

 

[JoeRambo]

Explain to me how half the time is 100% faster ???

"Faster" means speed. Say you are driving 100km at 100km/h and 200km/h. 2nd option is half the time and 100% faster. 150km/h would be 50% faster and so on with lesser impact on time.

 

[desrever]

Explain to me how half the time is 100% faster ???

https://en.wikipedia.org/wiki/Relative_change_and_difference

 

[AtenRa]

"Faster" means speed. Say you are driving 100km at 100km/h and 200km/h. 2nd option is half the time and 100% faster. 150km/h would be 50% faster and so on with lesser impact on time.

In the benchmark "faster means less time", it has nothing to do with speed.


In the benchmark, "less time = faster" and "more time = slower"

2s = 50% faster vs 4s

4s = 100% slower vs 2s

we are only using time, and less time is faster , more time is slower

 

[AtenRa]

https://en.wikipedia.org/wiki/Relative_change_and_difference

This has nothing to do with less time being faster and more time being slower

 

[Timmah!]

Here, this will help
Hans says that in order for the ZEN 4 to finish the benchmark in 31% less time, it has 46% higher performance (instructions/s)
You cannot say 31% less time is 46% faster, the performance of the CPU is 46% higher and that translates in to finishing the benchmark at 31% less time of 31% faster.

https://twitter.com/x/status/1528719223786192896

I think you are arguing semantics. Technically you should not be even allowed to say "x percent less time is x percent faster", because what does that even mean? What exactly is faster? Its not the time to do the task, because time cant go "faster". Thus its the processor, which as you yourself agree, is 46 percent faster.
Thats the only conclusion to this and i dont see any point in further arguing about it. But whatever floats your boat.

 

[JoeRambo]

In the benchmark "faster means less time", it has nothing to do with speed.

At some point people will start to genuinely wonder if You are trolling.

Let's instead use render that takes 100 seconds @ 2ghz and scales 100% with clock speed, okay? 4Ghz CPU would complete that task in 50 seconds and would be 100% faster.

Or in computer games, how much cooldown reduction one needs to cut cooldown of some skill in half? 100% reduction, twice as fast in reducing cooldown as standard cooldown "reduction" rate is.

 

[desrever]

This has nothing to do with less time being faster and more time being slower

It shows how how to calculate percent faster. Tho it expect you to know how to do elementary level math.