Speculation: Ryzen 3000 series

[Abwx]

Intel's claims point to ICL having a bigger increase over SKL than Zen 2 does over Zen/Zen+.
Leaked Geekbench results support those claims. The gap between the two is unclear, but I think it's pretty certain that Sunny Cove will have the edge compared to Zen 2, small or not.

From Zen to Zen+ there is 3% stated by AMD, dunno if you made a typo but that makes your estimation being relevant...

As far as I know, AMD never gave an 18% figure. They said 15% in Specint 2006 and 13% in Cinebench, not sure whether that's compared to PR or SR.

They said 15% for Zen+ to Zen (hence the 18%) in various workloads, they quoted Cinema 4D but also, wich is more interesting, Spec int 06 wich is more relevant for basic software than renderers.

Intel's 18% is an average that comes from multiple benchmarks, like Cinebench, Geekbench, Spec 2017, and several others.
Your accusations seem unfounded. If anything, I'd sooner believe that AMD changed their slide to point out the 15% Specint 2006 increase instead of the 13% Cinebench increase to look better compared to Sunny Cove. We have older slides that used the smaller Cinebench increase that were changed "last-minute" to use 15%...I also see nothing wrong with Intel including workloads where AVX512 plays a part.

Several others that are not known but we know that they used Web Xprt, an Intel originated bench and also Sysmark, wich is telling that it s not 18% if they need to use those quasi in house tests.

Notice that the uop cache has 2.25k ops stored, while Zen 2 is at 4k ops, this was pointed by AT s Ian Cutress, and is a likelyhood that Intel s uarch is not as efficient as Zen 2, indeed AMD said that the latter was designed to be up to ICL in 10nm iteration ..

https://www.computerbase.de/2019-05/intel-ice-lake-ueberblick/

 

[Schmide]

I also see nothing wrong with Intel including workloads where AVX512 plays a part.

I guess. If intel were to include AVX512 in all their segments? Yet traditionally AVX512 has only existed in the low power netbook and the high end server market.

Such segmentation leads to segmentation of results.

 

[HurleyBird]

Several others that are not known but we know that they used Web Xprt, an Intel originated bench and also Sysmark, wich is telling that it s not 18% if they need to use those quasi in house tests.

It's also potentially telling that they used R15 instead of R20, and ran two different versions of SPEC, basically doubling the weighting of SPEC in that 18%. Think Intel would do that if SPEC showed <18% improvement?

That said, I expect ICL might still have a slight advantage in single threaded IPC, growing as AVX-512 gets wider adoption, and also a moderate advantage in gaming. I expect Zen 2 will have a slight-to-moderate advantage in multi-threaded IPC.

But at the end of the day the question is somewhat academic if AMD is near launching Zen 3 by the time Intel gets their process mess sorted out.

 

[Abwx]

It's also potentially telling that they used R15 instead of R20, and ran two different versions of SPEC, basically doubling the weighting of SPEC in that 18%. Think Intel would do that if SPEC showed <18% improvement?

With Spec int 06 and Spec fp 06 they can take advantage of AVX512 as well as some new instructions, all this with the bench optimised with ICC for the uarch, and then there is Spec iNT_rate and Spec FP_rate wich are essentialy data manipulation dependant of bandwith....

That said, I expect ICL might still have a slight advantage in single threaded IPC, growing as AVX-512 gets wider adoption, and also a moderate advantage in gaming. I expect Zen 2 will have a slight-to-moderate advantage in multi-threaded IPC.

But at the end of the day the question is somewhat academic if AMD is near launching Zen 3 by the time Intel gets their process mess sorted out.

AVX512 as a saving grace..?.When even AVX1/2 are still marginal..?..
Anyway new instructions are only a timely solution since they can be implemented by the competition in a matter of two year, and still, they wouldnt be used broadly even after all this time.

 

[DrMrLordX]

It doesnt matter. I wont judge the IPC balance based on we all know is an outlier. I am really looking forward to get a skylake 4,4GHz+10% performance with reasonable real wattage.
Handbrake can use AVX2, lets see what that r3x got...

Glad you brought that up! @.vodka has you covered.

It's not that a working 10nm alone would do the trick, it's that 10nm being a failure has been holding that "next-gen" back.

Intel has missed not one but two uarch updates. They've also missed 2 years of process optimization on 10nm that should have been ongoing had 10nm worked out-of-the-gate in 2017. Clockspeed regressions could have been fixed for high-performance desktop parts (or whatever else Intel had in mind). Woulda coulda shoulda.

 

[TheGiant]

Stock R5 3600, ~49.25 FPS

https://www.reddit.com/r/Amd/comments/c5n8r9

8700k @ 4.7GHz, ~46.76 FPS

My R7 1700 @ 3.9GHz, maxed out, ~ 44.62 FPS

• dram calc tuned timings (14-15-14-14-28 1T)
• 1.7GHz/3400MHz IF/memory
• perfbias AIDA64/GB3 (+3% IPC, Zen+'s tighter cache latency)
• relaxed EDC throttling disabled

Yeah, that's 6 Zen2 cores @ 4.2GHz (unknown memory/IF speed) casually beating 8 maxed out Zen1 cores @ 3.9GHz in SIMD heavy workloads... and beating 6 Skylake cores at a 500MHz deficit. That doubled and improved SIMD hardware on top of the IPC improvements are really paying themselves off.

The higher end models at higher clockspeeds will rip through x264/x265 like they're nothing.

I dont use that benchmark, so uncle google told me

https://hwbot.org/benchmark/hwbot_x...=processor_5519&cores=6#start=150#interval=50

there is a big variability in results...sa again I am waiting for a proper review
but it definitely doesn't look bad

 

[.vodka]

I dont use that benchmark, so uncle google told me

https://hwbot.org/benchmark/hwbot_x...=processor_5519&cores=6#start=150#interval=50

there is a big variability in results...sa again I am waiting for a proper review
but it definitely doesn't look bad

Of course there's variability, you're linking to a ranking of HWBOT overclocked samples. How they tune their systems will obviously have an impact. This is why you're required to upload a screenshot of the run + all relevant system settings.

If anything this kind of testing on an encoder makes the results even more valid, as you're always using the same x265 version, and the same preset on the same input file.

raw  [info]: output file: run0-1080p.hevc
x265 [info]: HEVC encoder version 2.0+24-df559450949b
x265 [info]: build info [Windows][GCC 6.2.0][64 bit] 8bit
x265 [info]: using cpu capabilities: MMX2 SSE2Fast SSSE3 SSE4.2 AVX AVX2 FMA3 LZCNT BMI2
x265 [info]: Main profile, Level-4 (Main tier)
x265 [info]: Thread pool created using 12 threads
x265 [info]: frame threads / pool features       : 3 / wpp(17 rows)
x265 [info]: Coding QT: max CU size, min CU size : 64 / 8
x265 [info]: Residual QT: max TU size, max depth : 32 / 1 inter / 1 intra
x265 [info]: ME / range / subpel / merge         : hex / 57 / 2 / 2
x265 [info]: Keyframe min / max / scenecut       : 23 / 250 / 40
x265 [info]: Lookahead / bframes / badapt        : 15 / 4 / 0
x265 [info]: b-pyramid / weightp / weightb       : 1 / 1 / 0
x265 [info]: References / ref-limit  cu / depth  : 3 / on / on
x265 [info]: AQ: mode / str / qg-size / cu-tree  : 1 / 1.0 / 32 / 1
x265 [info]: Rate Control / qCompress            : CRF-20.0 / 0.60
x265 [info]: tools: rd=2 psy-rd=2.00 rskip signhide tmvp fast-intra
x265 [info]: tools: strong-intra-smoothing lslices=6 deblock sao

It doesn't look bad, not at all. Actually, it looks awesome. The 3900x and 3950x should eat video encoding for breakfast

 

[lobz]

You'll be able to say it confidently too once reviews are out.

Fair enough, then I'm trying not to call you out on this one

 

[coercitiv]

So the focus has switched from "Zen 2 won't beat the holy 5Ghz CFL" to "Zen 2 won't beat Sunnycove in IPC". Impressive considering Zen 2 hasn't even launched yet and AFAIK Ice Lake will not compete with Ryzen 3000 in desktop.

We are witnessing the birth of the IPC Messiah, which is quite ironic considering ICL is currently in fmax hell, with top frequencies so limited that 9th gen products may yet match it in overall performance.

People get excited about a CPU trailer while ignoring the elephant in the script - Intel avoids talking about overall CPU performance when comparing 10nm ICL to 14nm CFL.

 

[OTG]

Robert Hallock on memory latency-
"From the perspective of the operating system, these are essentially monolithic designs"

 

[Markfw]

Robert Hallock on memory latency-
"From the perspective of the operating system, these are essentially monolithic designs"

No closed captions. Can you elaborate ?

 

[exquisitechar]

So the focus has switched from "Zen 2 won't beat the holy 5Ghz CFL" to "Zen 2 won't beat Sunnycove in IPC". Impressive considering Zen 2 hasn't even launched yet and AFAIK Ice Lake will not compete with Ryzen 3000 in desktop.

We are witnessing the birth of the IPC Messiah, which is quite ironic considering ICL is currently in fmax hell, with top frequencies so limited that 9th gen products may yet match it in overall performance.

People get excited about a CPU trailer while ignoring the elephant in the script - Intel avoids talking about overall CPU performance when comparing 10nm ICL to 14nm CFL.

It's not really the focus and I'm not saying it to rain on Zen 2's parade - someone just said that Intel couldn't compete clock for clock with 10nm, which is likely not true. I already gave my thoughts on Ice Lake overall, it's irrelevant due to its clock regression and low volumes.

All will be revealed soon enough, IPC wise...

 

[Tuna-Fish]

No closed captions. Can you elaborate ?

Just that the design is not NUMA. Every core has common latency access to all memory, if not each other's caches.

 

[IEC]

No closed captions. Can you elaborate ?

All the CCXs have common latency access to the memory (via central IO die). Common cache access as well. So from the perspective of the OS, the chip might as well be a monolithic design.

 

[Markfw]

All the CCXs have common latency access to the memory (via central IO die). Common cache access as well. So from the perspective of the OS, the chip might as well be a monolithic design.

So does that mean it has much lower latency than Zen 2000 series ?

 

[OTG]

No closed captions. Can you elaborate ?

Sure! Gordon asked if there would be any performance difference between 1- and 2-CPU chiplet designs (3600x and 3900x, say).
As IEC already posted, because the IMC is on the IO die, all CCX's have the same memory latency and cache access latency.

 

[IEC]

Also per Robert Hallock re: AM4 socket design and any speculated differences between X370/X470/X570:

"We planned this 5 years ago"
You have faster chips that perform better using less power = you have MORE headroom for overclocking, even on X370. Per RH, no anticipated difference in manual OC performance (other than due to VRMs and maybe memory OC?)

Also, AnandTech's quote of Lisa Su saying with Ryzen 3000 series bringing core counts "up" and that Threadripper would have to go "up up" is apparently not a typo, as Robert Hallock used that exact phrase to describe core counts re: Threadripper 3000.

 

[moinmoin]

Also, AnandTech's quote of Lisa Su saying with Ryzen 3000 series bringing core counts "up" and that Threadripper would have to go "up up" is apparently not a typo, as Robert Hallock used that exact phrase to describe core counts re: Threadripper 3000.

So TR3 with 64 cores pretty much confirmed. Thanks!

 

[Markfw]

So TR3 with 64 cores pretty much confirmed. Thanks!

From here: https://www.anandtech.com/show/14579/all-ryzen-qa-with-amd-ceo-dr-lisa-su

It says:
"Tarinder Sandhu, Hexus: Given that you’ve got 24+ threads now in mainstream Ryzen (the 12-core), can it be argued at all that it’s kind of stepping on Threadripper’s toes?


Lisa Su: Threadripper is still an important step up. You will see future generations of Threadripper from us. Now obviously if mainstream is moving up, Threadripper is going to have to move up up."

Exact quotes, and this was from before Computex .

So not only does that mean more than 32 cores confirmed, with the IO die, no more memory handicaps with more than 16 cores. My 2990wx's and 2970wx's will be obsolete when these come out.

 

[.vodka]

So not only does that mean more than 32 cores confirmed, with the IO die, no more memory handicaps with more than 16 cores. My 2990wx's and 2970wx's will be obsolete when these come out.

Yeah. Zen2 TR fixes every and any issues with 1xxx and 2xxx TR, especially the 2990WX and 2970WX and their weird topology. It'll be a HUGE upgrade for these two particular models.

How much do you all think the 64 core part will cost? $3000? On the other hand, I suppose the 32 core replacement of the 2990WX will come at the same cost.

 

[Markfw]

Yeah. Zen2 TR fixes every and any issues with 1xxx and 2xxx TR, especially the 2990WX and 2970WX and their weird topology. It'll be a HUGE upgrade for these two particular models.

How much do you all think the 64 core part will cost? $3000? On the other hand, I suppose the 32 core replacement of the 2990WX will come at the same cost.

I would have to say in that neighborhood. Since I need at least 5 upgrades, that $15,000 for me to fix my farm. And 1950x's are already down to $400, so I won't get much relief there selling those 5. And the 2990wx is already down to $1200 and the 2970wx down to $1000. Somebody knows whats coming.

 

[Kenmitch]

All the CCXs have common latency access to the memory (via central IO die). Common cache access as well. So from the perspective of the OS, the chip might as well be a monolithic design.

The reviews should be interesting.

AMD's next challenge is applying similar technology to their gpu lineup.

 

[scannall]

I would have to say in that neighborhood. Since I need at least 5 upgrades, that $15,000 for me to fix my farm. And 1950x's are already down to $400, so I won't get much relief there selling those 5. And the 2990wx is already down to $1200 and the 2970wx down to $1000. Somebody knows whats coming.

Mark, sell your story. It's a great one. Really. I'm old as dirt myself, but you are going way and above addressing yours. Hats off. It might even get you a few 64 core Threadrippers.

 

[DeathReborn]

A TR with 48+ cores would be quite useful to me, should reduce my local server upgrade costs compared to Xeon/Epyc.

 

[Accord99]

8700k @ 4.7GHz, ~46.76 FPS

When set to the same Realtime setting as used by AMD, my 8700K at 4.7 GHz gets this: