AMD ZEN3 for EPYC Milan Processors to gain 20% performance

I'm going to build a Zen 3 system for my son. His 4-core Intel system is in need of an upgrade. All the 3700x and X570 systems i have been building for work and friends are humming along nicely.

 

Maybe it's time to upgrade the Threadripper 1950x soon

 

Since 15% is from IPC, only 5% comes from clock => power efficiency will be same or better.
I intend to keep my "gaming" / "productivity" CPU till 2nd generation of AM5. (Till prices for DDR5 settle a bit.)

But if there is serious performance boost from better memory access with DDR5...

 

I'm having to upgrade my rig so its looking like a 8 core zen3, 15% gains on single thread performance in gaming would be a huge win for AMD. I'm sure Intel will still be ahead if you want 1080p gaming at 240hz when they release their next stuff but Zen 3 is looking like a great set of cpus

 

Personally i'm excited for Zen 3 but not to buy, i'm excited to see how much better it is, and to anticipate how good Zen 4 with DDR5 and a overall new system will be bare minimum.

 

When I moved to Zen2, I moved from 8C/16T to 12C/24T, as improvement per core was not enough alone, but AMD delivered serious MT improvement at that price.
One generation is definitely not enough for me while keeping same number of cores. (And I do not need more than 12C/24T for now.) But 2 generation leap is quite big.
When there is going to be comparison between Zen1 and Zen3, it will be really jaw dropping. Zen1 were good CPUs at time for given price. Same applies for Zen2. I hope Zen3 will not be overpriced. But high cost of moving to new platform is to be expected.

At time we get to compare Zen2 to Zen4 it may be like 30~45% faster per 1thread->core and that's serious upgrade. And if they improve SMT on top of that 8C/16T will do more work than my 12C/24T in total, while having much better ST performance too. (I would likely not upgrade to that, as I would not want to upgrade just for ST performance. But still...)

 

Yeah i'm curious how DDR5 well affect Zen's performance considering how much Zen like a balanced ram speed.

 

If this comes out at least that well as it looks right now I will say goodbye to my current Intel build of mine.
Zen3 seems to be the best of cpu user can use with DDR4 for one more and maybe last time.
As DDR5 will be heavily overpriced for a generation or more Zen3 could be the one to 'survive' that waiting.

 

Thinking the same thing, but coming from a 1800x. Just have to decide if its a TR on just normal one.

 

Unless that Warhol cpu will be Zen3+ or Zen4 and will/could use still DDR4, if it is even a mainstream cpu.
https://www.guru3d.com/news-story/a...-with-codenames-leak-van-gogh-and-warhol.html

 

I just wish I knew how CPU demanding next-gen games would be. Games tend to revolve around the capabilities of consoles, so, I only care about getting whatever can handily keep up and nothing more. I'm sure Zen3 will be more than good enough to outpace consoles, but in the likely event my X370 board won't offer an upgrade path, I might as well just wait for DDR5/AM5 so I'm not stuck on a dead-end platform.

Zen2 doesn't seem to care as much about RAM speed as the first two gens. I wouldn't be surprised if Zen3 isn't really any different.

 

I am wondering if these will be 7nm on EUV or DUV. If they moved to EUV then the all core clock speeds should be lovely as that quad patterning they have used prior only yields 1-2 cores that can hit the boost clocks. We may have a reason to OC again. They may not have and simply will move to EUV with Zen4 when they use TSMC's 5nm.

I also expect a bit better latency due to the 8-core CCX's especially in the 8-core and under CPU's that have a singe CCX. This sort of feeds into IPC but I mention it because its a reason Intel winds at very high FPS gaming

I'm on a Gen1 Ryzen 1800x so i'm probably going to get a Zen3 chip as I have 32 GB of some very nice Samsung b-die memory. That should hold me for 2-3 generations.

 

I do not really think big OC is to be expected. Most of Zen2 CPUs can be clocked to 4.7GHz. But voltage required increases in very exponential way past 4GHz.
If used process allows similar clock at lower voltage, then chips can evade meltdown and be OCed.
At highest voltage I dared to use, 1CCX was stable with all cores under load @ 4.35GHz and another on 4.45GHz. And no, I did not like power draw. Because entire 12C/24T (under full load) can eat just around 65W when undervolted and locked to 4GHz.

 

Thats where I'm going once they move to EUV the voltage needed to hit higher clocks will drop significantly. If you take the best core on your 12C chip It needs less voltage than most other cores and will be the one getting boosted. I expect every core to be slightly better than that on EUV. You will see it become more on-par with Intel's 14nm but with less voltage required. One of the big downsides of quad patterning is the "lines" are not as precise which causes inconsistencies. This is why only 1-2 cores can hit those boost speeds at the sane voltages AMD has specked. No idea if they are moving to EUV but I really hope so as that should make Zen3 a fun chip to play with.

 

Zen 4 is DDR5. The leak you posted is about APUs

 

We see a curious paradigm shift atm, where an important segment of TR/workstation candidates, are "upgrading" to AM4.

.... except the lane constrained - right?

well maybe them too?

A pcie 4 gpu from amd or Nvidia, gets the same 16GB/s bandwidth from 8 lanes as the former pcie 3 gets from 16 lanes, netting the vast majority (even 8GB/s is rarely exceeded afaik), a spare precious 8 lanes.

 

According to the leak the Van Gogh and Cezanne are the APUs, though Warhol is still named as Zen3 without gpu. Will see.

 

So while waiting on those upcoming ZEN3 Cpus from AMD i made a final CPU overclock session for my 5820k, after updating BIOS against Spectre, thought i had already done it.

In stress-test benchmark, my CPU is now pulling 220W, for a 6 cores/12 threads mind you, i'm beyond 3950X territory by 80Watts here.

As Haswell-e doesn't have "Intel® Turbo Boost Max Technology 3.0" (Intel please...), you can't define per core ratio while overclocking..... Unless !
I just find out after 5 years that using "Intel Extreme Tuning Technlogy" i can set a per core overclock.

Long story short, my 5 years old CPU on 22nm now has its first core at 4.6Ghz and the 5 others at 4.5Ghz, with BCLK at 100.7 MHz, closing in on 8700K which is 3 years younger.

https://valid.x86.fr/bench/y8qggt/12

Now i hope it'll last at least 6 months, and then the whole rig is gonna hit the dust, even the H100i is starting to struggle.

edit edit: https://valid.x86.fr/bench/7xz29h

 

While I used 4.5GHz for all cores with 2500k, I had option to set turbo per number of cores under load.
I did 4.7/4.7/4.6/4.5 depending on number of cores used. (Not pleasant voltage for occasional 4.7GHz when CPU was mostly under full load in newer games.)
But per core boosting was available even for Sandy.