Question Zen 6 Speculation Thread

[OneEng2]

Overpriced Epyc? They are selling in droves. Clearly people think they are worth the money.

Many server applications charge licensing on an annually on a "Per Core" basis. This cost so totally and completely eclipses the server hardware cost to the extent that it becomes flat out stupid not to buy the best server chip money can buy.

This is also why I feel that Intel's loss of SMT is a much bigger deal in DC than people understand. Sure, they are able to put lots of little cores on a die that do pretty well against AMD's Zen 5 in single thread, but once you throw MT at it, Zen 5 gets a 40% boost that Skymont doesn't.

If AMD continues to boost performance in DC applications (and Zen 6 appears ready to do exactly that), I think it is an excellent financially driven architectural decision.

The N2 versions of Zen 6 that will be used in EPYC will likely up the core count (double is my guess). The clock speeds may well be limited by the 1KW rumored per-socket power limit for the new EPYC's in 2026.

What might be interesting is Clearwater Forest with its Darkmont cores on 18A up against EPYC D Zen 6c on N2. I am guessing that software licensing isn't an issue in this market (or a less powerful "c" Zen vs full Zen wouldn't make sense).

 

[adroc_thurston]

Many server applications charge licensing on an annually on a "Per Core" basis

You cheese that with F SKUs.

 

[OneEng2]

You cheese that with F SKUs.

F SKUs?

 

[LightningZ71]

F SKUs are processors that focus on achieving the highest possible boost clocks and sustained clocks to maximize per core performance over maximum socket MT performance for applications that are licensed in a per-core manner. Essentially, they wrong as much performance as possible out of a few cores as opposed to having as many cores as possible active.

 

[adroc_thurston]

F SKUs?

Frequency SKUs.
Turin ones are particularly extra kosher.

F SKUs are processors that focus on achieving the highest possible boost clocks and sustained clocks to maximize per core performance over maximum socket MT performance for applications that are licensed in a per-core manner. Essentially, they wrong as much performance as possible out of a few cores as opposed to having as many cores as possible active.

Bingo

 

[Win2012R2]

Many server applications charge licensing on an annually on a "Per Core" basis. This cost so totally and completely eclipses the server hardware cost to the extent that it becomes flat out stupid not to buy the best server chip money can buy.

Personally I think it's flat out stupid to get sucked into having to pay for software on per-core basis to begin with, and especially annually, but I guess for some big enterprises the choices are pretty limited in that respect. Having said that the most of demand for EPYCs must be coming from high-core-per-watt-loving hyperscalers who are smart and big enough not to get sucked into the above trap.

Still, AMD is wrong not to throw at least some crumbs into retail - giving middle finger like this isn't very smart strategy.

P.S. Would rather have 10% more cores or 10% discount than having SMT/HT enabled, but might change my mind after testing Turin on our workloads.

 

[igor_kavinski]

but might change my mind after testing Turin on our workloads.

Suppose when Turin is finally available for sale to mere mortals, Shimada Peak also arrives. Would you choose the latter, since it would have higher clocks than Turin, though just max 8 channels?

 

[StefanR5R]

F SKUs?

The first one in the EPYC family was 7371. It had the same amount of chiplets and CCXs and last level cache as the top-of-the line 7601 (4/ 8/ 64MB), but only half of the cores and threads enabled (16c/32t instead of 32c/64t), yet +11% TDP (200 W instead 180 W, for 3.1/3.6/3.8 base/all core/max clock, versus 7601' 3.2 GHz max boost clock).

Since then, the frequency optimized EPYCs have become more varied, and the "F" became part of the model name (as second digit in the Rome generation = 7F#2, as third digit in the Milan generation = 7#F3, and now as suffix with Genoa and Turin = 9##{4,5}F.) As previous posters said, the "F" generally means "frequency optimized", but I am adding that sometimes it additionally stands for "WTF". Take the EPYC 9175F: It has 16 cores/ 32 threads like its ancestor 7371, but they are implemented by means of 16 CCDs = one active core per CCD while having Turin's top-of-the-line CCD count. (The rest of its specs: 320 W TDP, 400 W cTDP, 4.2/4.55/5 GHz base/all core/max clock, 16× 32 MB level 3 cache.)

 

[StefanR5R]

Frankly, it doesn't matter

Except when it does. For example, take a count of enabled E cores in Raptor Lake-E (LGA1700 Xeon E-2400).

 

[Win2012R2]

Suppose when Turin is finally available for sale to mere mortals, Shimada Peak also arrives. Would you choose the latter, since it would have higher clocks than Turin, though just max 8 channels?

Hard to say without details, it depends on price/perf and wattage. Last time we considered using ThreadRippers it transpired that due to power usage we'd have to use 4U boxes instead of 3Us which we've got plenty of, but now Turin is so hot that it's a problem with it also. For some workloads we have fast clocks and reasonable amount of mem (size mostly, rather than bandwidth, but the latter is always welcome too!) will do better, so could be scope for this too, depends on price. I've had water cooled personal PCs for a decade, but not keen to use it in DC, high power usage is high running costs too, so high wattage is not just a thermal problem.

For some workloads we need a large X amount of RAM per core, so 12 channels are attractive and almost necessary -
1) to avoid paying over the odds for RAM by going higher than 64 GB DIMMS (having said that 96 GBs might become sensible next year, hopefully!) - this is one of the main reasons why I don't believe in SMT/HT, it would require a lot more RAM to run all at the same time, totally not feasible when your alternative is to go 128 GB DIMMs, cheaper to buy 2nd server
2) to avoid losing big time DRAM speed with 2DPC (if that's even available in single socket MBs)

I also don't want to even consider water cooling in DC - so 400W seems practical max at the moment, possibly 320-360W in 3Us.

My sweet dream is getting* 9575F

* reasonably priced, not for 12 grand!

 

[igor_kavinski]

For some workloads we need a large X amount of RAM per core, so 12 channels are attractive and almost necessary -

In your existing AMD servers, have you tried benchmarking the difference between NPS1 and NPS4 settings?

Explanation here: https://dl.dell.com/manuals/common/dell-emc-dfd-numa-amd-epyc-2ndgen.pdf

Could have an effect on performance if your workload has separate threads in the different CCDs so NPS4 would guarantee that they get data as quickly as possible from their local memory channels instead of the NPS1 setting introducing delays as the threads wait for data from the distant memory channels.

 

[Win2012R2]

In your existing AMD servers, have you tried benchmarking the difference between NPS1 and NPS4 settings?

I have read about it but have not tried that because I doubt very much it will make any difference to our custom workloads which typically use CPU affinity to stick to the core they are supposed to be on (don't trust really any OS to do it, this minimizes context switching anyway) and mostly work on their own self-contained data processing task with not a lot of synchronizing communication with other cores, out stuff is typically not THAT mem bandwidth bound.

If Turin got something like this then I'd definitely give it a try tho, much faster cores, might make a difference, especially with SMT.

 

[Gideon]

I'm not sure what the appropriate thread is (as there is no recent AMD + ARM thread) but interesting dibit from here:

Intel on the Brink of Death | Culture Rot, Product Focus Flawed, Foundry Must Survive

Intel’s board is incompetent and its horrible decisions over the decades are going to push it towards death. The decision to fire Pat Gelsinger, put in charge a CFO + career sales and marketing lea…

semianalysis.com

Qualcomm released Snapdragon X for Windows PC in 2024, and there will be more entrants. Nvidia and MediaTek are both independently working on Arm client PC chips; more details on these chips later. AMD, despite being a beneficiary of the x86 ecosystem, sees the writing on the wall and is also developing an Arm-based CPU for Microsoft as a semi-custom chip.

I wonder if Soundwave ends up as the Xbox handheld chip.

 

[GTracing]

Nvidia and MediaTek are both independently working on Arm client PC chips

Tangent, but it annoys me how often I see "both" and "each" used incorrectly. It should be "both together" or "each independently". "Both independently" is a contradiction in terms.

 

[igor_kavinski]

"Both independently" is a contradiction in terms.

It could mean that they may have a mutually shared product (CPU from Mediatek and GPU from Nvidia) but they also could be working on their "in-house" SoCs, so as not to put all their eggs in the partnership based SoC basket.

 

[Thunder 57]

Tangent, but it annoys me how often I see "both" and "each" used incorrectly. It should be "both together" or "each independently". "Both independently" is a contradiction in terms.

Nothing to loose your mind over . That's my grammar nazi trigger. It's amazing how common it happens. I was watching a professional video with two lawyers on youtube and text popped up about "loosing" something or other.

 

[igor_kavinski]

I was watching a professional video with two lawyers on youtube and text popped up about "loosing" something or other.

I was chatting with this Canadian female school teacher that I had known for several years. She was expressing her anger about some people and called them loosers. I burst out into laughter. She asked me what's funny? I told her, DID NOT expect that from you. She again asked, WHAT??? I told her that's not the correct spelling. And she tells me that this is the way she's always used it.

Get this. She taught her students English!

 

[DrMrLordX]

I was chatting with this Canadian female school teacher that I had known for several years. She was expressing her anger about some people and called them loosers. I burst out into laughter. She asked me what's funny? I told her, DID NOT expect that from you. She again asked, WHAT??? I told her that's not the correct spelling. And she tells me that this is the way she's always used it.

Get this. She taught her students English!

She should have called them hosers. What kind of Canadian doesn't do that?

 

[Doug S]

She should have called them hosers. What kind of Canadian doesn't do that?

Don't you mean hoosers?

 

[soresu]

Don't you mean hoosers?

Do you mean hoosiers?

Hoser is a Canadian slang term:

a foolish or uncultivated person.

 

[desrever]

Do you mean hoosiers?

Hoser is a Canadian slang term:

whoosh

 

[DrMrLordX]

Don't you mean hoosers?

Take off, eh!

 

[511]

Many server applications charge licensing on an annually on a "Per Core" basis. This cost so totally and completely eclipses the server hardware cost to the extent that it becomes flat out stupid not to buy the best server chip money can buy.

This is also why I feel that Intel's loss of SMT is a much bigger deal in DC than people understand. Sure, they are able to put lots of little cores on a die that do pretty well against AMD's Zen 5 in single thread, but once you throw MT at it, Zen 5 gets a 40% boost that Skymont doesn't.

If AMD continues to boost performance in DC applications (and Zen 6 appears ready to do exactly that), I think it is an excellent financially driven architectural decision.

The N2 versions of Zen 6 that will be used in EPYC will likely up the core count (double is my guess). The clock speeds may well be limited by the 1KW rumored per-socket power limit for the new EPYC's in 2026.

What might be interesting is Clearwater Forest with its Darkmont cores on 18A up against EPYC D Zen 6c on N2. I am guessing that software licensing isn't an issue in this market (or a less powerful "c" Zen vs full Zen wouldn't make sense).

Sadly there is no Zen6/6C before 2H26 at the least

 

[fastandfurious6]

Sadly there is no Zen6/6C before 2H26 at the least

that's a very long wait and still not 2nm? apple will definitely release 2nm in 2025

 

[FlameTail]

that's a very long wait and still not 2nm? apple will definitely release 2nm in 2025

No.
2026.