AMD to unveil Zen 4 CPUs at CES 2022

g67575 · 24 May 2022 at 22:22

I'm curious about the bit confirming the mem controller running at 1:1, with DDR5. Can someone give the timestamp where they confirmed that please?

Based on Robert Hallock's comments, it definitely sounds like V-cache will be a big thing for Zen 4, presumably without the clock speed/power limits that Zen 3 had.

uscool · 24 May 2022 at 23:12

Just what I wanted to hear , board seems over engineered because they want to support the socket for long term like AM4 also took shot at intel changing platform 1 - 2 years

g67575 · 24 May 2022 at 23:20

Thinking about recommending a PCIE 5 compatible NVME drive for a new AM5 PC build for my dad, as it sounds like the technology has become quite mature by this point. I wondering if cooling is likely to be an issue, just as was on some PCIE4 NVME drives?

timorous · 24 May 2022 at 23:20

humbug said:
I'm listening to the MLID podcast and to be fair to him he has spotted something interesting.

The strange way AMD got the "31% faster than Alder Lake"

The 12900K rendered 69% vs 100% of the image for the 7950X, so "31% faster" right? not quite, its 31 percentage points faster, not 31% faster, its actually 45% faster, in other words the 12900K would have to gain 45% performance to match the 7950X.

69 / 100 = 1.449 (44.9%)
To proof it
69 + 44.9% = 99.981 (100)

The 7950X is 45% faster than the 12900K.

I said this earlier, as did others.

AMD goofed there because 31% faster is just wrong. The 12900K was 31% slower.

Troezar · 25 May 2022 at 06:05

g67575 said:
I'm curious about the bit confirming the mem controller running at 1:1, with DDR5. Can someone give the timestamp where they confirmed that please?

Based on Robert Hallock's comments, it definitely sounds like V-cache will be a big thing for Zen 4, presumably without the clock speed/power limits that Zen 3 had.

Just listened to it again 1:07:35. He is asked about memory speeds running 1:1. It is not directly related to the demo. He confirms DDR5 will run 1:1 but not what speeds. That means either it was running at 1:1 in the demo or the did the demo at a less optimal ratio. Either way it's interesting but does not confirm they ran it at 3000MHz.
There is quite a bit of information from that timestamp on until the end of the video if anyone wants to skip the bulk of it.

g67575 · 25 May 2022 at 07:00

Thanks for checking that. I'm gonna guess it will work at 1:1 with DDR5 running at 5200mhz, as that is the officially supported frequency.

So, it sounds like AM5 will have an advantage over LGA1700 boards, for the memory controller.

Troezar · 25 May 2022 at 07:29

g67575 said:
Thanks for checking that. I'm gonna guess it will work at 1:1 with DDR5 running at 5200mhz, as that is the officially supported frequency.

So, it sounds like AM5 will have an advantage over LGA1700 boards, for the memory controller.

Yes I misunderstood the context first time. The interesting thing is then that the demo may have been a little faster if run 1:1 rather than at 6000MHz? Also heard some thoughts that 16C may not be the top SKU. Different package to allow 3x8 or maybe a chip without an iGPU. Lots of possibilities still at this point.

Dave2150 · 25 May 2022 at 07:39

Still very disappointed with only 15% ST performance. Just how pathetic is the IPC gain in the actual architecture, given the 'free' gains from DDR5, better process, much higher clocks etc?

Not even matching 12900k in ST performance, that launched in November, is really not what I was expecting.

Once Intel release the 13900k (rumoured 10%+ performance increase, 24 cores), AMD will be in for a rough time. What's more, 13900k may launch before Zen4 for all we know...

Armageus · 25 May 2022 at 07:40

Troezar said:
Also heard some thoughts that 16C may not be the top SKU. Different package to allow 3x8 or maybe a chip without an iGPU. Lots of possibilities still at this point.

The number of pads and increased TDP limit certainly points to there being other options over the lifetime of the socket, if not this generation

Robert896r1 · 25 May 2022 at 08:15

To run mem controller at 3000mhz at 1:1 you’re looking at Aida latency in the mid 20’s and that’s not going to happen. If you have taken some time to learn how memory works, you’ll understand why it’s laughable.

I think this is the same guy who made up a bunch of nonsense in the past when he was drawing things up on a clear whiteboard. If it is him, I’d temper my expectations severely.

alec · 25 May 2022 at 08:18

RavenXXX2 said:

weeding through his self back patting, the reasonable argument that Zen 4 is more or less node shrink of Zen 3 + more cache + avx512.

I am starting to agree that single digit IPC increase is likely.
Zen4 achieved 45% win over 12900K thanks to
- higher (sustained all-core) clocks thanks to better node
- much higher DDR5 bandwidth.
- Blender supports AVX512!
- well, IPC helped too

KompuKare · 25 May 2022 at 08:24

A

Troezar said:
Different package to allow 3x8 or maybe a chip without an iGPU. Lots of possibilities still at this point.

The whole idea about chiplets is re-usability, so unless some other product also has a use for a smaller I/O die without a iGPU, I think it's unlikely that for this gen there's going to be any AM5 with more than 2x8.

welshrat · 25 May 2022 at 08:25

Dave2150 said:
Still very disappointed with only 15% ST performance. Just how pathetic is the IPC gain in the actual architecture, given the 'free' gains from DDR5, better process, much higher clocks etc?

Not even matching 12900k in ST performance, that launched in November, is really not what I was expecting.

Once Intel release the 13900k (rumoured 10%+ performance increase, 24 cores), AMD will be in for a rough time. What's more, 13900k may launch before Zen4 for all we know...

How disappointed are you though Dave. I reckon on a scale of 1 - 10 its zero.

uscool · 25 May 2022 at 09:11

Dave2150 said:
Still very disappointed with only 15% ST performance. Just how pathetic is the IPC gain in the actual architecture, given the 'free' gains from DDR5, better process, much higher clocks etc?

Not even matching 12900k in ST performance, that launched in November, is really not what I was expecting.

Once Intel release the 13900k (rumoured 10%+ performance increase, 24 cores), AMD will be in for a rough time. What's more, 13900k may launch before Zen4 for all we know...

Raptor lake will be on a dead platform, AM5 is engineered to have lasting longevity drop in CPU upgrades . I know which I would prefer

G J · 25 May 2022 at 10:33

Well the single digit increase rumors get worse the lower down the CPU stack you go for the 7600/7800 etc as they have lower clocks.

These could overclock well though.

Troezar · 25 May 2022 at 13:29

Robert896r1 said:
To run mem controller at 3000mhz at 1:1 you’re looking at Aida latency in the mid 20’s and that’s not going to happen. If you have taken some time to learn how memory works, you’ll understand why it’s laughable.

I think this is the same guy who made up a bunch of nonsense in the past when he was drawing things up on a clear whiteboard. If it is him, I’d temper my expectations severely.

Exactly which then means they ran the demo at something other than 1:1 which I presume is less than optimal?

LePhuronn · 25 May 2022 at 13:49

Robert896r1 said:
To run mem controller at 3000mhz at 1:1...

Yeah, that's not what was said. Clarification with timestamps are literally in this post thread.

cliffy · 25 May 2022 at 16:31

Dave2150 said:
Still very disappointed with only 15% ST performance. Just how pathetic is the IPC gain in the actual architecture, given the 'free' gains from DDR5, better process, much higher clocks etc?

Not even matching 12900k in ST performance, that launched in November, is really not what I was expecting.

Once Intel release the 13900k (rumoured 10%+ performance increase, 24 cores), AMD will be in for a rough time. What's more, 13900k may launch before Zen4 for all we know...

Still 15% is better than the confirmed <10% of RPL. With 7950X beating the 12900K in blender by approx 46% i would guess RPL even with its 8 extra e cores may get its pants pulled down in multithreaded workloads. I also have a sneaky feeling that the gaming benchmarks are going to be pretty good, after all no one said it di not match ADL in single threaded performance just that in R23 it was about 15% fatser then teh 5950x, this is only one test

g67575 · 25 May 2022 at 16:45

I think Zen 4's cache increases could swing it towards AMD, in games. If not, there's always the upcoming vcache versions.

g67575 · 25 May 2022 at 23:11

Dave2150 said:
Once Intel release the 13900k (rumoured 10%+ performance increase, 24 cores), AMD will be in for a rough time. What's more, 13900k may launch before Zen4 for all we know...

I'm doubtful there will be any IPC improvement with Raptor Lake, especially since they only said there would only be an improvement to multithreaded performance.

Both Intel and AMD are pushing for the highest frequency clocks now. AMD can apparently reach 5.2-5.5ghz with a maximum CPU power consumption of just 170w.