melkor_unlimitedYour results are showing that cores in threadripper do not get too much power. At higher counts they are starving.

I do not understand what you mean. These are the same classes of workstation processor between Shimada Peak and GNR-WS, and trust me, both are getting enough power as they are overclocked by professionals to squeeze as much performance as possible out of them lol. My personal tests on my 658X lineup perfectly.

Intel has been far ahead of AMD in IPC on their P-cores for many years now, and that trend will only continue. Panther Cove-X will likely obliterate Zen 6 and even Zen 7 in IPC. The only reason there has been parity on desktop is because Intel uses E-cores as well. Which, while the E-cores have gotten pretty good, the overall per core performance is diminished because of them.

ir_cowHopefully. I want Xeon back on consumer sockets. I miss the X79 days.

C2D through X99, good days... X99 could even have Rdimms on consumer boards...

PatriotC2D through X99, good days... X99 could even have Rdimms on consumer boards...

X99 and X79 were not consumer sockets (former E5-2699v3 user here). They are equivalent to today's LGA4677 and LGA4710-2. What people should be wishing for are Xeons (aka HEDT/WS) to be much cheaper than they are now. Prices for HEDT/WS have gotten insane the past 5-years due to lack of competition from Intel. That looks to be changing as GNR-WS is very competitive, but Intel needs to be incentivized to not abandon the HEDT/WS again for a couple years. They better launch a Coral Rapids-WS in late 2027/early 2028 to compete with Mustang Peak.

XeonsNOpteronsX99 and X79 were not consumer sockets (former E5-2699v3 user here). They are equivalent to today's LGA4677 and LGA4710-2. What people should be wishing for are Xeons (aka HEDT/WS) to be much cheaper than they are now. Prices for HEDT/WS have gotten insane the past 5-years due to lack of competition from Intel. That looks to be changing as GNR-WS is very competitive, but Intel needs to be incentivized to not abandon the HEDT/WS again for a couple years. They better launch a Coral Rapids-WS in late 2027/early 2028 to compete with Mustang Peak.

x99 and X79 were not sockets they were consumer chipsets for the HEDT (high end DESKTOP) segment and they shared the 2011 and 2011v3 socket with the server segment. so... no x99 and x79 were consumer even if you could slot in Xeons to consumer boards.

What people are wishing for is an end to artificial segmentation of markets, but ddr4 and 5 having the notch be different for rdimm vs ecc and electrically different enough makes that rather difficult.
for 2011v3 you could have an x99 board or a WS C612 board depending on the features you needed and not pay the workstation tax.

The explosion of core counts created such a large gap between server and consumer space that hedt got forgotten about and the workstation became a superclocked server variant.
LGA4677 and LGA4710-2 are sockets for workstations not HEDT, the distinction is the motherboard choices and purpose though intel is blurring the lines with overclocking support on workstation boards.

AMD skipped the HEDT variant of the threadripper for Zen3 and it greatly hurt that segment, but they got away with it because well... Intel hasn't been competing there.
I do also appreciate that AMD has just stuck to RDIMMS in the HEDT and WS threadrippers as unbuffered ecc has always cost more than rdimms and been capacity limited.

XeonsNOpteronsNah, my testing puts it literally on par with Zen 5 (and in more modern Cinebenchs, actually ahead of Zen 5). Lion Cove however is consistently 10-12% ahead of Zen 5. This is just some data from HWBot of Zen 5 Threadripper vs GNR-WS for some reference in Cinebench 2024. The right column is IPC.

Industry standard specint2017 testing puts RWC IPC on par with Zen 4.

(David Huang, P-cores running at Fmax, AMD lead attributed to slightly higher boost clocks).
Unfortunately skatterbench uses different variants of cinebench in his GNR OC review vs the one he did for his Zen 4 TR one, but using his GB6 ST results, you see once again that RWC there has similar IPC as Zen 4 does.
This is LNC vs Zen 5 too, for reference:

Note that the 9900x and the 285K have similar boost clocks for their P-cores, so their scores here should be directly comparable for an IPC comparison. RPL (using GLC, which is very similar to RWC) seems like it comes close in total perf for specint2017, but that's due to the 7% bump in Fmax it has over the AMD part in this chart.
Here's another comparison from C&C between Zen 5 and RWC in GNR Xeon 6:

The AMD part has a ~24% higher score in specint2017 1T, however it also boosts ~13% higher than Intel's part, meaning that IPC should be ~ >10% higher (the greater than symbol because IPC decreases at higher clocks, hence why IPC comparisons are done best at iso frequency rather than dividing boost/perf at different boost clocks).

GeddagodIndustry standard specint2017 testing puts RWC IPC on par with Zen 4.

(David Huang, P-cores running at Fmax, AMD lead attributed to slightly higher boost clocks).
Unfortunately skatterbench uses different variants of cinebench in his GNR OC review vs the one he did for his Zen 4 TR one, but using his GB6 ST results, you see once again that RWC there has similar IPC as Zen 4 does.
This is LNC vs Zen 5 too, for reference:

Note that the 9900x and the 285K have similar boost clocks for their P-cores, so their scores here should be directly comparable for an IPC comparison. RPL (using GLC, which is very similar to RWC) seems like it comes close in total perf for specint2017, but that's due to the 7% bump in Fmax it has over the AMD part in this chart.
Here's another comparison from C&C between Zen 5 and RWC in GNR Xeon 6:

The AMD part has a ~24% higher score in specint2017 1T, however it also boosts ~13% higher than Intel's part, meaning that IPC should be ~ >10% higher (the greater than symbol because IPC decreases at higher clocks, hence why IPC comparisons are done best at iso frequency rather than dividing boost/perf at different boost clocks).

There's only one important detail with such a highly synthetic benchmark such as SPEC 2017. Is AVX-512 enabled? That's the only way a 9900X tops out a 285K in any given benchmark ;)

XeonsNOpteronsI do not understand what you mean. These are the same classes of workstation processor between Shimada Peak and GNR-WS, and trust me, both are getting enough power as they are overclocked by professionals to squeeze as much performance as possible out of them lol. My personal tests on my 658X lineup perfectly.

I am not sure what those professionals do, but numbers show that with the same TDP total you will get less power per core with higher numbers of cores.
9965wx ... 350W / 24 cores = 14.58 W per core
9975wx ... 350W / 32 cores = 10.94 W per core
9985wx ... 350W / 64 cores = 5.47 W per core

And that means lower clock -> lower CB score per core
Exactly what your number are telling.

BTW ... CB MT score / number_of_cores ≠ IPC

Does that influence the release date of "52-core ready" motherboards?

Patriotx99 and X79 were not sockets they were consumer chipsets for the HEDT (high end DESKTOP) segment and they shared the 2011 and 2011v3 socket with the server segment. so... no x99 and x79 were consumer even if you could slot in Xeons to consumer boards.

They were not consumer, it was the HEDT/WS line. You literally explained they were the same socket as the top-of-the-line server chips at the time. No point in arguing if you cannot see that.

PatriotThe explosion of core counts created such a large gap between server and consumer space that hedt got forgotten about and the workstation became a superclocked server variant.
LGA4677 and LGA4710-2 are sockets for workstations not HEDT, the distinction is the motherboard choices and purpose though intel is blurring the lines with overclocking support on workstation boards.

Another misconception. There was always a large gap between server and consumer core counts. Xeon Platinum 8180 had like 5 times the core count of consumer Intel parts like the i7-8700K at the time. That is pretty much the same ratio as today with the Xeon 6980P vs Core Ultra 9 285K. You have bought into the thinking that somehow core counts have gotten much greater on the server side, but as I state, that is not the reason for the price increases. The W-3175X (which was overpriced even for the time) was essentially Intel's top server chip but unlocked and "only" came to $3000. The reason prices have gotten high is AMD hype, greed and lack of competition from Intel.

GeddagodIndustry standard specint2017 testing puts RWC IPC on par with Zen 4.

I honestly do not care for these benchmarks much, far too irrelevant to real tests. At least Cinebench is doing an actual rendering. I have many more benchmarks from OpenFOAM, and other Cinebench versions which justify my statement. Redwood Cove is on par with Zen 5, and Lion Cove is ~10% ahead of Zen 5.

melkor_unlimitedI am not sure what those professionals do, but numbers show that with the same TDP total you will get less power per core with higher numbers of cores.
9965wx ... 350W / 24 cores = 14.58 W per core
9975wx ... 350W / 32 cores = 10.94 W per core
9985wx ... 350W / 64 cores = 5.47 W per core

Those chips are not drawing their TDPs lol. They are overclocked, easily drawing 600W+. I use them as those professional overclockers tune everything, or at least try, to get the best performance out of them.
Here is one of the 9985WX submissions for you to take a look at. He is drawing 880W:
hwbot.org/benchmarks/cinebench_-_2024/submissions/5869478

melkor_unlimitedBTW ... CB MT score / number_of_cores ≠ IPC

If you bothered to look at the table, and looked at the numbers, you would realize I divided by the all-core clock speed for those submissions. Sorry to sound harsh, but you seemed snarky in your comment when you are clearly not understanding the table.

XeonsNOpteronsThey were not consumer, it was the HEDT/WS line. You literally explained they were the same socket as the top-of-the-line server chips at the time. No point in arguing if you cannot see that.

You are calling chipsets sockets. Fairly basic stuff, cant even read the word desktop without calling it a server. Truly your intellect is dizzying.

Another misconception. There was always a large gap between server and consumer core counts. Xeon Platinum 8180 had like 5 times the core count of consumer Intel parts like the i7-8700K at the time. That is pretty much the same ratio as today with the Xeon 6980P vs Core Ultra 9 285K. You have bought into the thinking that somehow core counts have gotten much greater on the server side, but as I state, that is not the reason for the price increases. The W-3175X (which was overpriced even for the time) was essentially Intel's top server chip but unlocked and "only" came to $3000. The reason prices have gotten high is AMD hype, greed and lack of competition from Intel.

Perhaps another misconception by you... There has not always been a large gap between server and consumer, For the longest of times, desktops had 1 core, and servers had 1 core (abiet multiple cpus/server but that hasnt changed.)
If you want to move the goalpost to coffee lake I can move it back... but with sandybridge 1155 4c 2011 8c, ivybridge 4c, 2011, 12c , 15c for 2011-2 but that was its own thing.
the v3 and v4 chips exploded up to 18/22 cores on 2011-3 while consumer stayed the same but it was very clearly not always as large a gap... but fairly consistent consistent from Pentium through sandybridge.

but if you want to measure the current gap with AMD at 16 cores, and intel at 24cores compared to Turin 192c and Xeon6 288 cores I think you will find the gap has grown from 1x to 12x.

I honestly do not care for these benchmarks much, far too irrelevant to real tests. At least Cinebench is doing an actual rendering. I have many more benchmarks from OpenFOAM, and other Cinebench versions which justify my statement. Redwood Cove is on par with Zen 5, and Lion Cove is ~10% ahead of Zen 5.

Please post your benchmarks and your test settings, trust me bro after your display of "knowledge" is not going to hold up. You are very dismissive of the expertise of others while showing none.
You seem to be arguing with 5 people from the peak of mt..... time to knowledge up and humble up.

XeonsNOpteronsI honestly do not care for these benchmarks much, far too irrelevant to real tests.

Dr. DroThere's only one important detail with such a highly synthetic benchmark such as SPEC 2017.

It's because they are relevant to "real tests", that they are the industry standard.

XeonsNOpteronsAt least Cinebench is doing an actual rendering.

Yes, spec2017 has a subtest for rendering, as well as many other workloads:

XeonsNOpteronsI have many more benchmarks from OpenFOAM, and other Cinebench versions which justify my statement. Redwood Cove is on par with Zen 5, and Lion Cove is ~10% ahead of Zen 5.

Perhaps you should add the qualifier "RWC is on par with Zen 5, in the benchmarks I care about" then. Because by the metric companies and architects usually use to measure IPC, that is not the case.
Also, I'm curious, is there being a 33% gap in IPC between the different Zen 5 skus in your table not ringing any alarm bells?
And to a lesser extent, since when was has IPC usually been measured using nT benchmarks? Because by that metric, with LNC not having SMT, I doubt LNC is even ahead of RWC iso clock.

GeddagodIt's because they are relevant to "real tests", that they are the industry standard.

What I am saying is that Zen 5 has AVX-512 and none of the Intel CPUs bar the Xeon 6 do, which means math heavy benchmarks that are AVX-512 aware will zip on the 9900X. In most other benchmarks limited to AVX2, this scenario does not repeat itself

Dr. DroWhat I am saying is that Zen 5 has AVX-512 and none of the Intel CPUs bar the Xeon 6 do, which means math heavy benchmarks that are AVX-512 aware will zip on the 9900X. In most other benchmarks limited to AVX2, this scenario does not repeat itself

Sorry, I meant to reply to the rest of your quote about AVX-512 and it totally slipped my mind by the time I got to the end of that reply.
Well, as you said, the Xeon 6 CPU does support AVX-512, so the specint2017 comparison to Zen 5 in the last picture in the reply you were responding to there is "fair".

GeddagodSorry, I meant to reply to the rest of your quote about AVX-512 and it totally slipped my mind by the time I got to the end of that reply.
Well, as you said, the Xeon 6 CPU does support AVX-512, so the specint2017 comparison to Zen 5 in the last picture in the reply you were responding to there is "fair".

The clock speeds are very low in comparison, though, an iso clock measurement would be required. Intel chips will hold their own on any synthetic restricted to 256-bit vectors, and yes, I realize this is "intentionally gimping" the Zen 5 chip, it's that at the same time the only modern processor that supports it is the Ryzen. Rocket Lake is easily over five years old, this was a regression introduced with Alder that stuck by until now.

XeonsNOpteronsX99 and X79 were not consumer sockets (former E5-2699v3 user here). They are equivalent to today's LGA4677 and LGA4710-2. What people should be wishing for are Xeons (aka HEDT/WS) to be much cheaper than they are now. Prices for HEDT/WS have gotten insane the past 5-years due to lack of competition from Intel. That looks to be changing as GNR-WS is very competitive, but Intel needs to be incentivized to not abandon the HEDT/WS again for a couple years. They better launch a Coral Rapids-WS in late 2027/early 2028 to compete with Mustang Peak.

X58, X79, X99 and X299* were consumer platforms that shared the socket with workstations and low-end & mid-range servers (Xeon E5 series). Similarly LGA115x was used in low-end servers (Xeon E3 series) but that doesn't mean Z77 or Z97 isn't a consumer platform.
During that time high-end servers (Xeon E7 series) used LGA1567, LGA2011-1 and LGA3647 which is the line that the LGA4677 etc. are successors to.

*Low-end & mid-range server CPUs were moved to the high-end socket so LGA2066 was only used for (cheaper) workstations and consumer HEDT.

I think at this point it would make more sense to give the mainstream platform quad channel memory and more PCIe lanes (40-48 lanes from the CPU). It's really unlikely for Intel to ever lower the WS platform prices to the level where the old consumer HEDT platforms were. Nowadays even mainstream Z890 motherboards are mostly in the price range where consumer HEDT motherboards used to be.

XeonsNOpteronsIf you bothered to look at the table, and looked at the numbers, you would realize I divided by the all-core clock speed for those submissions. Sorry to sound harsh, but you seemed snarky in your comment when you are clearly not understanding the table.

I bothered myself.
Twice.
And result is the same.
You mixed things you should not mix and got something what in not what you claim it is.