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!

 

[pakotlar]

I guess I need to spell it out and say 'Whoosh"....

We're looking at an 7xxx maxed out marchitecture with:
PCIe 5
DDR 5
AVX512 (fwiw)
+/- 5.5Ghz
Built-in RDNA 2 Gfx
>15% gain in single-threaded work
>35% overall performance gain (multi-threaded workloads)
>25% performance-per-watt gains

And this will likely be at a price Intel was only 5 years ago charging for its upper i7 systems....

IIRC I was doing the pencil trick on my Celeron 300a in an Abit BP6.
And yes, this moaning and groaning that AMD is somehow slacking off and acting like Intel of yore vis-a-vis milking it, is both maddening and laughable.

At least Timorous got it.

So the answer to my question is yes. I want more for consumers, you want less, and apparently 7 others are avidly in support of companies giving us ever smaller gen on gen gains. Seems dumb, but ok.

 

[Timmah!]

That's a mess of different drives. It looks like you are currently using 2 X M.2 slots and 5 SATA ports. Not knowing what your budget is, for my personal usage, I would probably save my pennies until I could afford the following:

Buy 4 of a decently rated 6TB Hard Drive. In the end, reliability over performance, avoid SMR.
Buy a single 12TB+ Hard Drive.

Keep the 950 Pro for boot.
Use Windows Storage Spaces (via powershell) to create a tiered storage array with the 4 X 6TB drives in a RAID 0 (no redundancy, stripped) as a HDD tier and the 960 EVO as the SSD Tier.
Configure the 12TB drive to automatically backup the storage array nightly.

That covers 2 M.2 NVME ports and 5 SATA ports. If you have a 6th SATA port, you can retain the 850 512GB for games and non-essentials.

The above configuration will give you effectively NVME speeds for any of your active tasks while keeping 24GB of storage available to you. You keep all of your programs and data on it, save your games if you have the SATA port for it. You have near-line disaster recovery with the backup drive. You should already have an offline backup, so I did not consider that.

I know the above works because I have almost that exact configuration in my son's computer and have set similar configurations for multiple others.

The rules, make sure that the SSD tier NVME drive is healthy. Always, ALWAYS, check that your backup drive is making good backups. Not having that drive is not an option for a work computer. Once you build the configuration, it will work almost seamlessly.

There are a lot of people that are not fans of MS Storage Spaces. You need to do your research on how to make the configuration as you shouldn't use that if you don't know how to build it. I have had excellent success with it and haven't yet had an array fail for anything other than a mechanical drive failure. That's what the backup is for.

Sorry not to respond sooner. Thanks for your suggestion, its indeed interesting proposition, one that i did not even think of.

However, the 6TB pricing you mention in your other post, 70 bucks, is not reality where i live, at least i dont know where, i only found drives of that size in the 170~200 EUROs range. So quite a lot of cost for 5 HDDs, while i am looking to build truly expensive rig, my funds are not unlimited.

I do agree with you my current drive setup is rather messy and i am looking to simplify it and reduce number of drives. This is not exactly that So i guess i just stick to my initial plan of getting couple of new M2 drives, store my actual work files in there and set-up some backing up system to one of the HDDs. Or maybe i will get some NAS for that purpose.

I may still however reconsider, nothing is decided at this point.

 

[pakotlar]

Did ya not read the sarcasm in that post? That post was dripping with it.

Ah yes, “sarcasm”.

 

[pakotlar]

What's funny is someone found a mistake in PPT slides that AMD officially described "Cinebench R23 1T 35% improvement" in the foot note:

https://twitter.com/x/status/1547434913862803461

,,,,,,,if it turns out to be real. or it would have lawsuit lol.

NT refers to MT, not 1T lol. Wishful thinking. They provided 15% figure for 1T, that was based on R23 1T, which is likely what the endnote refers to.

 

[gdansk]

So the answer to my question is yes. I want more for consumers, you want less, and apparently 7 others are avidly in support of companies giving us ever smaller gen on gen gains. Seems dumb, but ok.

Yes we all want 50% more IPC and 32 cores. But be reasonable. They can only diverge so much from existing designs without delaying their launch or taking a big risk. If things slip they are eaten by competitors who did make incremental improvements.

 

[eek2121]

NT refers to MT, not 1T lol. Wishful thinking. They provided 15% figure for 1T, that was based on R23 1T, which is likely what the endnote refers to.

Actually, nT means anywhere from 1 to infinite number of threads. For example, you can run the multithreaded bench on a single thread rather than running the single threaded test itself.

 

[eek2121]

4.4GHz base despite AVX512 ?

Unless those instructions are not enabled for some SKUs that would be boldly technically speaking.

Bold of you to assume that AMD has to down clock for AVX-512 instructions. Remember this is a completely different implementation from Intel.

Are there any technical reason having high base clock? 4.4Ghz is the highest base clock i ever seen. And it seems this is not AVX512 clocks though. If this base clock is all due to process enhancement, all core turbo might be easily surpass 5Ghz even for 6C12T.......

Base clocks are typically the guaranteed minimum all core clocks for the heaviest workloads. That means when running something like blender, the chip will always run at (4.4ghz in this example) or higher than the base clock. Note this has nothing to do with idle clocks.

If the base clock on this chip is 4.4ghz or higher, Zen 4 is going to be something special indeed. That is 700mhz higher than the 5600x.

 

[MadRat]

Wouldn't AVX be independent of the CPU clock and operate in lockstep of some ratio off fsb or memory speed?

 

[jpiniero]

Wouldn't AVX be independent of the CPU clock and operate in lockstep of some ratio off fsb or memory speed?

No. The issue with Intel is that they have to raise the voltage when using AVX-512 at the same clock for it to remain stable, which obviously increases power.

 

[pakotlar]

Yes we all want 50% more IPC and 32 cores. But be reasonable. They can only diverge so much from existing designs without delaying their launch or taking a big risk. If things slip they are eaten by competitors who did make incremental improvements.

Then we’re in agreement. I would have been thrilled with the 8-10% IPC and 24 cores.

They’re delivering a rumoured 35% at +65W (+60%) when using 6000 MT DDR5 vs 3600 MT DDR4 on Zen 3, on a new full node (5nm). This is much worse than Zen 3 delivered without a new node or new memory generation (+19% IPC on the same 7nm). Yeah, MT is better than Zen 3, but it’s essentially an overclocked part + expensive DDR5 that gets us there.

Am I unhappy with this? Not really, it’s probably still better than Raptor Lake, and is pushing Intel to do better. Is it the weakest generational improvement since Bulldozer? Yep.

The exciting changes for me are AM5 (DDR5, PCIE5) and AVX512 support. The actual core is ok, but not nearly as exciting as exciting as Genoa and Threadripper Pro. It’s clear that they’re focusing their energy on server and workstation markets, hence all the core increases, 4+ channel memory is restricted to parts that costs thousands. Slowly leaving behind regular consumers. No one should be thrilled or defensive about that.

 

[pakotlar]

Actually, nT means anywhere from 1 to infinite number of threads. For example, you can run the multithreaded bench on a single thread rather than running the single threaded test itself.

Cinebench nT commonly refers to >1T; it’s “n” because the number of threads isn’t fixed, but variable with the number of threads offered on the CPU. The 1T tests are referred to as such. I agree it’s ambiguous.

edit: see an example below, hope that helps

https://www.tomshardware.com/news/intel-13th-gen-raptor-lake-release-date-specifications-pricing-benchmarks-all-we-know-specs

 

[RnR_au]

Additionally, they clearly showed they can be every-bit as bad as Intel, in areas where they have currently no competition - look no further than the whole Threadripper situation.

Couldn't that be explained by competing demands from various units inside AMD for precious silicon?

 

[carrotmania]

They’re delivering a rumoured 35% at +65W (+60%) when using 6000 MT DDR5 vs 3600 MT DDR4 on Zen 3, on a new full node (5nm). This is much worse than Zen 3 delivered...

Why do you insist on misquoting AMD? Whats your agenda here?

AMD states ">15%" but you determine that means "15%".
AMD state that AM5 socket can do 170W, but have in NO way released info on what individual CPUs will use.

7950X could still be 105W, but you continually push to the worst case scenario. Calm it.

 

[DrMrLordX]

Wouldn't AVX be independent of the CPU clock and operate in lockstep of some ratio off fsb or memory speed?

AMD doesn't have separate logic for their temp/clockspeed/voltage curve when running AVX code. At least as of Zen2, the standard boost algo constrains voltage based on current and total power draw. AVX2 (and presumably AVX512 workloads on Zen4) tends to pull a lot of current, which hampers the boost algo's ability to feed higher voltage to the cores (since it still has to obey PPT limits). That has the consequence of modifying the temp/clockspeed/voltage curve so that clocks have to come down based on voltage constraints. It's no coincidence that overall power draw tends to be high in these circumstances, along with temps (which is no surprise given the current draw).

 

[inf64]

Why do you insist on misquoting AMD? Whats your agenda here?

AMD states ">15%" but you determine that means "15%".
AMD state that AM5 socket can do 170W, but have in NO way released info on what individual CPUs will use.

7950X could still be 105W, but you continually push to the worst case scenario. Calm it.

~15% ST improvement over Zen3 will likely show up in the lowest of Zen4 lineup (7600X?).
Simple math we've all been through countless times (and pakoltar is well aware of it): 10% IPC x ST clock increase = 1.1x5.7/4.9 ~= 28% in case some Zen4 models boost to 5.7Ghz.
5.7Ghz seems like a lot but they demoed 16C ES hitting 5.5Ghz in a gaming workload so pure 1T workloads likely can run on ~3.6% higher clock (5.7Ghz).

 

[yuri69]

Here we go again. Many people are upset since AMD presented their Zen architecture as an IPC-first approach. There were multiple IPC trend slides like:

Those were true in range Zen 1 - Zen 3. But Zen 4 comes nearly 2 years after Zen 3 with ~10% IPC.

 

[inf64]

Here we go again. Many people are upset since AMD presented their Zen architecture as an IPC-first approach. There were multiple IPC trend slides like:

Those were true in range Zen 1 - Zen 3. But Zen 4 comes nearly 2 years after Zen 3 with ~10% IPC.

Although it was a disappointing IPC uplift versus the previous Zen iterations, we will still get a larger generational uplift than what Zen2->Zen3 brought.
Zen4 will be a great server chip and unsurprisingly a great desktop chip. I expect it will have significantly better FP performance than Zen3, so this will factor in the average as well.

 

[szrpx]

https://twitter.com/x/status/1548717866198831104

Take with grain of salt, ENG sample, but still some interesting stuff.

 

[nicalandia]

Numbers are Up for Genoa, by Yuuki_AnS...

https://twitter.com/yuuki_ans

Sapphire Rapids QS vs AMD Genoa ES on CineBench.



For some reason Cinebench is not detecting the full 196 Cores of Genoa so it's only working with 128C/256T and even with that it's breaking the 100,000 points. It would be close to 150,000 with 192 cores.





Edit.

I am checking with YuuKi_AnS a few points.

He is listing the 8480/8480+ as 112C for 2S System, but those should be 120C/240T per 2S System. Perhaps he meant the 8470 which are 56 Cores.. English is not his first language so it's kinda hard to follow him.

 

[Saylick]

Jesus, that's an absolute MT beat down for Genoa over Sapphire Rapids. A straight up massacre.

Not sure why Genoa's 128c R20 MT score is lower than Milan's 128c score though, but then reverses for R23. Does R23 take advantage of AVX512?

 

[Exist50]

Not sure why Genoa's 128c R20 MT score is lower than Milan's 128c score though, but then reverses for R23

I think it's best just to throw out these results. None of the ES/QS chips' scores really make sense.

 

[Markfw]

I think it's best just to throw out these results. None of the ES/QS chips' scores really make sense.

Sure, just throw out anything that doesn't make Intel look good.

 

[JasonLD]

Sure, just throw out anything that doesn't make Intel look good.

Aren't you the one who keeps insisting that we should wait for the official reviews from reputable sites?

 

[Exist50]

Sure, just throw out anything that doesn't make Intel look good.

Golden Cove performing far worse than Zen 3 makes sense? Or how about Genoa showing next to no improvement (if not an outright regression) over Milan?

You truly have abandoned all reason.

 

[Markfw]

Aren't you the one who keeps insisting that we should wait for the official reviews from reputable sites?

Early data is not totally reputable, but his statement was to throw out the results. They make sense, in that in most cases genoa beats Milan. And as for the above comment. Golden cove <> sapphire rapids just the same as 5950x core <> Milan core.. Due to being a server, they certainly can be clocked lower. You Intel supporters should not comment on Zen 4 if all your feedback is negative. Wait for official reviews, and talk about them in Intel threads, otherwise it looks like you are troilling AMD threads.

And speaking of abandoning all reason, I say its you that has lost all reason.