AMD Unveils 5 nm Ryzen 7000 "Zen 4" Desktop Processors & AM5 DDR5 Platform

[R0H1T]

IPC numbers can be different depending on applications, Intel & AMD generally use 10+ varying workloads to show avg IPC gains.

If so, which is is the right one since they can't be two ways and each one gives a different result.

Both strictly speaking, since we're not solely depending on just one benchmark.

Maybe different application gives different impression of the CPU's speed.

Yes & even (different) RAM configurations will give different "IPC" numbers.

maybe we should wait for the general performance metric instead

I'll wait for that fully loaded zen4 (4d?) die!

 

[DeathtoGnomes]

Well, I guess the potential negative reviews will push AMD to decrease the prices of the platform.
Because otherwise, AMD will not sell well. The prices are too high as is.

Price creep cant be helped.

 

[Valantar]

My favorite part about NFTs is if someone drops an image in your wallet, like Cheese Pizza, you cant delete it, you can only hide it or send it somewhere. Super secure!

I thought that asrock's repeated failure of low end motherboard designs and them going after reviewers would do it for you.

I'm really confused with some comments here. 15% ST gain is quite a bit in my opinion. Then you have the IPC gain which some people misinterpret. AL has the same IPC as 5000 Series AMD CPUs. According to GURU 3d.
View attachment 248539

So 15% increase is not nothing I would think. Also you have the frequency boost. I'm puzzled how AMD measures the IPC to be honest.
Consider this. The 5800x3d has the same IPC as 5800x and 12900k according to guru3d and yet it is way faster in games due to 3dvcache but lacks in other apps like MT apps in comparison to 5800x due to lower frequency. So IPC is one thing, ST performance is another and general performance is totally different thing. I haven't watched the presentation yet but I'm really going to refrain from speculations and guess what it will be like. Especially, if this is supposed to be something totally different than 5000 series CPUs.
What I'm trying to say is, the IPC and frequency etc. is misleading in any way. You have to look at the bigger picture here.

That ... is a very poor way of measuring IPC. Fixed frequency is ... fine, though not ideal (as IPC is highly dependent on caches, interconnects and RAM, locking clocks can present an unrealistic image of actual real-world IPC as you're changing the relative speeds of those separate clock domains), but a single application is not an IPC benchmark. If you're going to talk about real-world IPC (and not architectural-level execution ports etc.), you need a broad range of applications to give any kind of representative overview. A single application just isn't enough.

 

[HisDivineOrder]

I was expecting more. That being said, having been there for the early days on AM4, I'll let others beta test AMD's new platform. Hopefully, they'll have ironed it out by Zen5.

 

[dont whant to set it"']

*"pre-production model" its in the video , as words by company's CEO.

 

[Valantar]

Come on it's obvious they have something up their sleeve, these are either zen4c cores or they have x3d designs lined up for later!

Isn't Zen 4c supposed to be a lower area, lower clocked core for higher density implementations? That definitely makes zero sense for an MSDT platform topping out at the same core count as its predecessor. It's definitely not unlikely that there will be X3D SKUs later on, but who knows when/if those will arrive? At least there's no mention of them for now.

 

[HD64G]

Nice thing for AMD and us customers that they seem to sandbag with Zen4 performance. Intel won't know what will be competing with and will be cautious in pricing. Many people will stop going for DDR5 platforms for months and prices will drop. If AMD isn't snadbagging then they won't have crazy pricing and Intel won't be able to increase their pricing after launching their 13th gen CPUs. If Zen4 is faster than it seems so, Intel will cut pricing as they did vs Zen3 when they lost the performance crown and AMD will follow suit. Competition helps us more than the duopole they retain in the CPU market for decades now.

 

[Steevo]

I really am doubting any double digit actual IPC increase at this point in X86-64 architecture **unless** we all agree that they should sell us CPUs that have KNOWN speculation and branch vulnerabilities since they don’t need to use hardware to check threads.

I would be OK with 5.5Ghz and a few vulnerabilities for another 10% IPC.

 

[Crackong]

204s is 31% faster than 297s; 297s is 45% slower than 204s.

No
It is the other way around

If you treat the whole job as 1
Ryzen 7000 needs 204s to finish it make it 1/204 = 0.00490 jobs per second
12900k needs 297s to finish it = 1/297 = 0.00337 jobs per second
Take the zeros off
(490-337)/490 = 31% slower
(490-337)/337 = 45% Faster

So
12900k is 31% slower than the 16 core Ryzen 7000
Ryzen 7000 is 45% faster than the 12900k

 

[bobsled]

No
It is the other way around

If you treat the whole job as 1
Ryzen 7000 needs 204s to finish it make it 1/204 = 0.00490 jobs per second
12900k needs 297s to finish it = 1/297 = 0.00337 jobs per second
Take the zeros off
(490-337)/490 = 31% slower
(490-337)/337 = 45% Faster

So
12900k is 31% slower than the 16 core Ryzen 7000
Ryzen 7000 is 45% faster than the 12900k

 

[Crackong]

"Who is faster?" is a "Higher is better" scale.
It might sounds a little confused for the question when the data is presented with time and time is "Lower is better"
Therefore we need to convert them to"Higher is better"
For example
David needs 10s to run a 30m distance
Paul needs 20s

To find out "Who is faster" we need to divide and find the speed
Therefore it is 30/10 vs 30/20
So 3m/s vs 1.5m/s

So David runs 3m/s and he is 100% faster than Paul's 1.5m/s

Same situation applies to the test today.

 

[TheLostSwede]

Not sure if this is a 100% accurate, there are definitely some discrepancies compared to the information that has come out today, but it seems pretty close.

Site Launch Exclusive: All the Juicy Details on AMD's Quirky Chipset Solutions for AM5!

Eastbridge? Westbridge?

angstronomics.substack.com

 

[Deleted member 24505]

Yeah, if the Ryzen 9 5950X is already 20% faster than Core i9-12900K, while the newest Ryzen 7000 CPU is only 30% faster than Core i9-12900K, then Raptor Lake will have the door very wide open to beat it and to reign supreme.

Which means that AMD is in a trouble and in the winter we may see for the first time since 2017, that Intel overtakes AMD in the DIY sales ala German MindFactory...

Was considering jumping to AM5 but might wait and see what raptor is like as it will fit in my 690 board.

 

[oxrufiioxo]

Nope.
View attachment 248535

I should've been more specific the 5950X can be as much as 20% faster in Blender depending on what workload is chosen especially for the longer ones.

Unfortunately unless comparing identical workloads you can't compare at the same time both Intel/AMD choose what makes them looks best also them not talking about increases over the previous flagship is odd. Hopefully the comparison was made on a workload that the 5950X finishes in about the same time as the 12900k.

 

[FeelinFroggy]

Does it use the 3d V-Cache? I assume that it does, but I did not see it mentioned anywhere and since the 5800x3d is a locked chip I did not know if it could support 5.5ghz clock speed.

 

[oxrufiioxo]

Was considering jumping to AM5 but might wait and see what raptor is like as it will fit in my 690 board.

Guessing they will be showing off Raptor Lake soon I think it will be similar enough that buying an AM5 motherboard won't be worth it. Kinda like reverse from already owning a decent AM4 motherboard and switching to Z690. The X670E boards the really interesting ones IO wise will probably be stupidly expensive as well.

Does it use the 3d V-Cache? I assume that it does, but I did not see it mentioned anywhere and since the 5800x3d is a locked chip I did not know if it could support 5.5ghz clock speed.

I think if it did they would have mentioned it. I do find it odd that there was only mention of L2 cache increasing. Wouldn't be shocked if they just make a cpu focused on gaming that uses it again. The gains in application performance seem to be nonexistent or lower on the 5800X3D vs the 5800X as it is.

 

[Valantar]

No
It is the other way around

If you treat the whole job as 1
Ryzen 7000 needs 204s to finish it make it 1/204 = 0.00490 jobs per second
12900k needs 297s to finish it = 1/297 = 0.00337 jobs per second
Take the zeros off
(490-337)/490 = 31% slower
(490-337)/337 = 45% Faster

So
12900k is 31% slower than the 16 core Ryzen 7000
Ryzen 7000 is 45% faster than the 12900k

That makes no sense. We already have a unit: seconds (per job), not jobs per second. There was a single job run. 297/204=1.45, i.e. 297s is 1.45x 204s, or 297s is 45% slower than 204s. 204/297=0.68686868=~0.69, i.e. 204s is 0.69x 297s, or 204s is 31% faster than 297s. No amount of juggling numbers and changing units will change those basic relations.

"Who is faster?" is a "Higher is better" scale.
It might sounds a little confused for the question when the data is presented with time and time is "Lower is better"
Therefore we need to convert them to"Higher is better"
For example
David needs 10s to run a 30m distance
Paul needs 20s

To find out "Who is faster" we need to divide and find the speed
Therefore it is 30/10 vs 30/20
So 3m/s vs 1.5m/s

So David runs 3m/s and he is 100% faster than Paul's 1.5m/s

Same situation applies to the test today.

This ... is some really creative mathematical nonsense. "Who is faster" is only a "higher is better" scale if you translate it from lower time=better to something inversely proportional to that. The base assumption of asking "who is faster" for a predefined task is "who performs that task the quickest", not "who performs the most iterations of that task in a given amount of time". That is an entirely different question. Unless it is defined beforehand that what you're looking for is rate (repetitions/time) and not speed (time/repetition), then the base assumption is that anyone asking what is fastest is that they're asking about speed.

 

[R0H1T]

Isn't Zen 4c supposed to be a lower area, lower clocked core for higher density implementations?

Why would they clock lower the (lower cache) variants? You think that makes sense. There's a ton of rumors around this but bigger cache variants are almost certainly "more dense" & IMO lower clocked as well.

 

[Valantar]

Why would they clock lower the (lower cache) variants? You think that makes sense. There's a ton of rumors around this but bigger cache variants are almost certainly "more dense" & IMO lower clocked as well.

You clock them lower because that allows a higher number of cores within a given power limit - which is also why you want a smaller core to begin with. The architectural differences mean these clock drops aren't linear compared to clock scaling for the regular core, but they will most definitely be there, unless 4c manages to trade off its simplicity with astoundingly good clock scaling.

Which of course also means that Zen4c makes no sense in a low core count implementation, as the sacrifices made to afford higher core densities in HPC/server applications don't make any sense in that scenario, even if you would be able to clock them higher than in a dense server die.

 

[Crackong]

That makes no sense. We already have a unit: seconds (per job), not jobs per second. There was a single job run. 297/204=1.45, i.e. 297s is 1.45x 204s, or 297s is 45% slower than 204s. 204/297=0.68686868=~0.69, i.e. 204s is 0.69x 297s, or 204s is 31% faster than 297s. No amount of juggling numbers and changing units will change those basic relations.

However your explanation doesn't make sence mathematically.
Because 0.69 only means 0.69 or 69% , it doesn't mean 31%

Your equation doesn't work bothways since you need to decide when to put the 1 in front / after the answer (ie 1-0.69 or 1.45-1)

My equation is consistent
I suggest you check #87 to see how to calculate "Who is faster" which is a "bigger is better" scale.
The thing is
You need to calculate the "Speed" of the process.
Dividing both time doesn't give you the speed
"Distance / time" gives you the speed.

It is simple maths.

 

[oxrufiioxo]

I did find her wording on AM4 during the presentation interesting.

AMD confirms its AM4 platform "will continue for many years to come" - VideoCardz.com

AMD AM4 socket is not going anywhere yet AMD next-gen AM5 platform might have just been unveiled at Computex, but AMD is not saying goodbye to AM4 just yet. The AM4 socket has been a great success for AMD and consumers, especially those on the first gen 300-series motherboards, which have just...

videocardz.com

 

[Valantar]

However your explanation doesn't make sence mathematically.
Because 0.69 only means 0.69 or 69% , it doesn't mean 31%

Uh ... percentages are relative. Relative to something else. If you're talking time to finish a job, and your baseline is 100%, then a result of 69% is indeed 31% faster. If, on the other hand, you redefine the baseline to be your new result, then that old 100% result becomes 145% of that, making it 45% slower.

What you're doing here is attempting to redefine the base variable from "time elapsed" to "jobs performed". This is an explicit reversal of what was presented.

I suggest you check #87 to see how to calculate "Who is faster" which is a "bigger is better" scale.

....did you miss the part where I quoted that post directly? Also, that post entirely fails to explain this supposed point.

The thing is
You need to calculate the "Speed" of the process.

You're confusing speed with rate. In this case, speed is time per job, rate is jobs per time.

Dividing both time doesn't give you the speed
"Distance / time" gives you the speed.

... there is no "distance" here, except metaphorically. But let's go with that metaphor: the "distance" is a single Blender render. That makes speed "how quickly do you finish one render?", not "how many renders/time are you capable of". The latter question asks for a rate, not a speed.

It is simple maths.

Except that your maths fail to understand the questions being asked, and are thus being misapplied.

I did find her wording on AM4 during the presentation interesting.

AMD confirms its AM4 platform "will continue for many years to come" - VideoCardz.com

AMD AM4 socket is not going anywhere yet AMD next-gen AM5 platform might have just been unveiled at Computex, but AMD is not saying goodbye to AM4 just yet. The AM4 socket has been a great success for AMD and consumers, especially those on the first gen 300-series motherboards, which have just...

videocardz.com

Hm, that's indeed interesting. Though most likely she's just referring to the fact that the platform will be supported for quite a while yet - i.e. CPUs aren't being discontinued immediately, nor will AM4 be aimed at a full-stack replacement any time soon. I could also see OEMs and business partners continue making AM4-based products for low cost markets, entry business PCs, etc. that don't need the fast I/O or extreme performance of AM5 - especially given how AMD doesn't have Intel's massive chipset tier list with delineations of PCIe generations, DDR support, etc. Still, there's always the potential of that meaning 6nm AM4 refreshes (even if only for OEM markets) down the line, as that should be pretty cheap and easy for them to make.

 

[Crackong]

Uh ... percentages are relative. Relative to something else. If you're talking time to finish a job, and your baseline is 100%, then a result of 69% is indeed 31% faster. If, on the other hand, you redefine the baseline to be your new result, then that old 100% result becomes 145% of that, making it 45% slower.

What you're doing here is attempting to redefine the base variable from "time elapsed" to "jobs performed". This is an explicit reversal of what was presented.

....did you miss the part where I quoted that post directly? Also, that post entirely fails to explain this supposed point.

You're confusing speed with rate. In this case, speed is time per job, rate is jobs per time.

... there is no "distance" here, except metaphorically. But let's go with that metaphor: the "distance" is a single Blender render. That makes speed "how quickly do you finish one render?", not "how many renders/time are you capable of". The latter question asks for a rate, not a speed.

Except that your maths fail to understand the questions being asked, and are thus being misapplied.

Here is a really really simple example
Person A uses 100s to finish a job
Person B uses 200s

With YOUR equation
200/100 so B is 100% slower than A
100/200 so A is 50% faster than B

Your equation is fundamentally flawed because in your equation, A will NEVER be 100% faster than B becasue it has to be finished with 0 second to do that, in YOUR equation.
Please, use your common sense.
If a person finish his job in 10s when the other guy needs 200s, he is 20x faster than the other guy, but in YOUR equation, he is just 95% faster.

Even if the first person only needs 1s to finish the job, in YOUR equation he is just 99.5% faster.
In reality he is 200x faster

Comon it is simple Maths

 

[ARF]

Was considering jumping to AM5 but might wait and see what raptor is like as it will fit in my 690 board.

Zen 4 on AM5 and Raptor Lake on Z790 will launch around the same time, plus minus a few weeks.

 

[Drash]

Here is a really really simple example
Person A uses 100s to finish a job
Person B uses 200s

With YOUR equation
200/100 so B is 100% slower than A
100/200 so A is 50% faster than B

Your equation is fundamentally flawed because in your equation, A will NEVER be 100% faster than B becasue it has to be finished with 0 second to do that, in YOUR equation.
Please, use your common sense.
If a person finish his job in 10s when the other guy needs 200s, he is 20x faster than the other guy, but in YOUR equation, he is just 95% faster.

Even if the first person only needs 1s to finish the job, in YOUR equation he is just 99.5% faster.
In reality he is 200x faster

Comon it is simple Maths

Perspective. change/original x 100 is the formula, where change = new - original (or the reverse, it's just a sign). Which is the original is the point. Faster implies slower is original, slower ...

eg I take a work cut - reduce to 4 days from 5 = 20% cut, but I get them back (4 days to 5) is a 25% increase!