Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[igor_kavinski]

@igor_kavinski
Why 9999W and not 256W?

Based on the assumption that even going to 9999W may do nothing more than burn the CPU out. Meaning, just stop at 253W and don't burn any more power.

 

[PJVol]

Based on the assumption that even going to 9999W may do nothing more than burn the CPU out. Meaning, just stop at 253W and don't burn any more power.

I meant the power is unlimited in both cases, which could be indicated on your diagram.

 

[igor_kavinski]

I meant the power is unlimited in both cases, which could be indicated on your diagram.

Hey maybe we'll see 9999W in UEFIs now instead of the boring 4096W

 

[Timmah!]

Hey maybe we'll see 9999W in UEFIs now instead of the boring 4096W

Is 9999W one watt under the fabled 1 iggawatt? 😛

 

[yottabit]

So now the Osborne effect has moved on to the “special” X3D chips right? I know I’ll be willing to wait until September to learn more.

It seems like I’ll be doing a Zen5 X3D or Zen6 upgrade for my 5950x

Still pretty happy with what we’ve seen of Zen5 performance though

 

[yottabit]

Does anyone have any good resources on the 1:1 and 2:1 modes and memory tuning for Zen in general? I'm completely oblivious to that topic and would like to start learning more since it seems to be coming up a lot here

 

[Hail The Brain Slug]

Does anyone have any good resources on the 1:1 and 2:1 modes and memory tuning for Zen in general? I'm completely oblivious to that topic and would like to start learning more since it seems to be coming up a lot here

I don't have any resources offhand but the core concept is pretty easy to understand.

Ryzen has a memory controller that can only run in one of two clock modes: 1:1 with memory speed or 1:2 with memory speed. So, for example, DDR5 6000 runs the memory controller in 1:1 at 3000MHz.

The upper limit we've observed for this is around 3200MHz, or DDR5 6400. In order to run faster ram, like 7200/7600/8000 it has to be changed to 1:2 mode, which runs at 2000MHz at DDR5 8000.

The memory controller running at lower frequency does impair performance somewhat, so clocking the ram very high, say 8000MT/s is necessary to benefit over say, 6000MT/s in 1:1 mode.

 

[igor_kavinski]

The memory controller running at lower frequency does impair performance somewhat

In terms of latency or bandwidth?

 

[Hail The Brain Slug]

In terms of latency or bandwidth?

Can't say, I wasn't able to successfully get either of my systems to run at a higher memory clock so I haven't had any hands-on experience with it. Maybe @Det0x knows more about it.

 

[Fjodor2001]

a very large custom cooling rig with a optimus signature v3 block together with a MO-RA3 radiator

Any idea what the Zen5 temps are if more ”normal” cooling solutions are used?

 

[igor_kavinski]

Any idea what the Zen5 temps are if more ”normal” cooling solutions are used?

Nope. But can't be much different than Zen 4. On air temperature during heavy AVX-512 workloads, that's an open question at this point.

 

[Josh128]

I disagree here-- I think we will see a markedly different temperature / power ratio compared to Zen 4. AMD hinted at it at CES. What changes they made to achieve this and to what extent, remains to be seen.

 

[Det0x]

In terms of latency or bandwidth?

Can't say, I wasn't able to successfully get either of my systems to run at a higher memory clock so I haven't had any hands-on experience with it. Maybe @Det0x knows more about it.

Will copy my post over from from a other forum as a response:

So over the last few weeks ive been working on a memory performance comparison between the following maxed out daily memory profiles

• Profile1 = SR 2x16gigs adie @ 6600MT/s CL26-37-32-30-62 + 2200mhz FCLK 1:1 mode
• Profile2 = SR 2x16gigs adie @ 8000MT/s CL32-45-40-44-84 + 2200mhz FCLK 2:1 mode
• Profile3 = DR 2x32gigs adie @ 6600MT/s CL28-38-36-36-72 + 2200mhz FCLK 1:1 mode
• Profile4 = DR 2x32gigs adie @ 8000MT/s CL34-46-44-60-104 + 2200mhz FCLK 2:1 mode

The CPU used for this purpose is my newly acquired SP99 7800X3D @ daily settings, running on the ASUS GENE motherboard



My main performance metric for this comparison have been Clam cache/memory benchmark and/or Karhu ramtest, but i have also included AIDA64 and hwinfo in my screenshots as i know people in this thread mostly like to look at them. My criteria for being a fully stable daily memory profile and being added to this performance comparison is being able to survive atleast 6 hours in karhu and over 1 hours Y-cruncher all memtests only.

With all that out of the way, we can start looking at some numbers

SR 2x16gigs adie @ 6600MT/s CL26-37-32-30-62 + 2200mhz FCLK 1:1 mode


SR 2x16gigs adie @ 8000MT/s CL32-45-40-44-84 + 2200mhz FCLK 2:1 mode


DR 2x32gigs adie @ 6600MT/s CL28-38-36-36-72 + 2200mhz FCLK 1:1 mode


DR 2x32gigs adie @ 8000MT/s CL34-46-44-60-104 + 2200mhz FCLK 2:1 mode


Results in Clam cache/mem benchmark:

Latency ranking:

1. SR 2x16gigs @ 6600MT/s 1:1 mode= 68.75ns
2. DR 2x32gigs @ 6600MT/s 1:1 mode =70.17ns
3. SR 2x16gigs @ 8000MT/s 2:1 mode = 70.24ns
4. DR 2x32gigs @ 8000MT/s 2:1 mode = 71.84ns

Bandwidth read-modify-write (ADD) ranking:

1. SR 2x16gigs @ 8000MT/s 2:1 mode= 97.11GB/s
2. DR 2x32gigs @ 8000MT/s 2:1 mode = 92.87GB/s
3. SR 2x16gigs @ 6600MT/s 1:1 mode = 91.23GB/s
4. DR 2x32gigs @ 6600MT/s 1:1 mode = 87.34GB/s

A few comments in random order to my findings above

A single 8core Zen4 CCD can take advantage of the higher bandwidth afforded by 2:1 mode vs 1:1 mode, even if the common misconception on many forums is that there is no benefit because they can hardly see any difference in gimmicky AIDA64 memory bench. (its also easy to double check this in other benchmarks such as y-cruncher / GB3 membench which will show the same)

The next question would naturally be what's the "best memory setup", 1:1 mode with its lower latency or 2:1 with its higher bandwidth. There is no easy answer for this as it all depends on what benchmark/game you comparing the numbers in.. Some will prefer latency while others bandwidth, so you just have to check on an individual basis.

But what i can say is that i pretty much always think higher memoryspeed is better, be it in 1:1 mode or 2:1 mode... From time to time i see some limit themself to something like 6000/6200MT/s because they think its faster in games than say 6400MT/s for some reason (?)

My next observation is that i did not find any bandwidth benefit from the "dual rank" (quad) in Clam cache/mem benchmark, but karhu is seemingly showing higher mb/s. But i suspect this is because the higher memory size tested, not increased bandwidth from DR. I will do some more DR karhu runs where i limit used memorysize to same as SR and check if the numbers change. (y) edit Its also possible the forced GDM enabled with DR is eating up the bandwidth benefit compared to SR

Have also seen some complains about some ppl having a hardtime tuning memory on the 1.1.7.0 PatchA FireRangeP AGESA, i can only say that is working pretty good for me on the ASUS GENE, even if i'm using a beta bios. But be warned, stabilizing DR 64gigs @ 8000MT/s is still insanely hard, think i spent like 5x the time on this profile compared to all others combined... Its really on a razors edge, +-5 mv on some rails and you can forget about 10k karhu.

 

[Mopetar]

The whole concern over Zen 4 temps has been overblown. It doesn't matter if it's 35 or 95 degrees, it's still dissipating the same amount of heat.

Meanwhile we're leaning that 13k and 14k chips from Intel that were operating at lower (more usual) temperatures have been having stability issues or that the silicon was degrading.

 

[Det0x]

Will copy my post over from from a other forum as a response:

So over the last few weeks ive been working on a memory performance comparison between the following maxed out daily memory profiles

• Profile1 = SR 2x16gigs adie @ 6600MT/s CL26-37-32-30-62 + 2200mhz FCLK 1:1 mode
• Profile2 = SR 2x16gigs adie @ 8000MT/s CL32-45-40-44-84 + 2200mhz FCLK 2:1 mode
• Profile3 = DR 2x32gigs adie @ 6600MT/s CL28-38-36-36-72 + 2200mhz FCLK 1:1 mode
• Profile4 = DR 2x32gigs adie @ 8000MT/s CL34-46-44-60-104 + 2200mhz FCLK 2:1 mode

The CPU used for this purpose is my newly acquired SP99 7800X3D @ daily settings, running on the ASUS GENE motherboard

View attachment 102953View attachment 102954

My main performance metric for this comparison have been Clam cache/memory benchmark and/or Karhu ramtest, but i have also included AIDA64 and hwinfo in my screenshots as i know people in this thread mostly like to look at them. My criteria for being a fully stable daily memory profile and being added to this performance comparison is being able to survive atleast 6 hours in karhu and over 1 hours Y-cruncher all memtests only.

With all that out of the way, we can start looking at some numbers

SR 2x16gigs adie @ 6600MT/s CL26-37-32-30-62 + 2200mhz FCLK 1:1 mode
View attachment 102955View attachment 102956

SR 2x16gigs adie @ 8000MT/s CL32-45-40-44-84 + 2200mhz FCLK 2:1 mode
View attachment 102957View attachment 102958

DR 2x32gigs adie @ 6600MT/s CL28-38-36-36-72 + 2200mhz FCLK 1:1 mode
View attachment 102959View attachment 102960

Results in Clam cache/mem benchmark:

Latency ranking:

1. SR 2x16gigs @ 6600MT/s 1:1 mode= 68.75ns
2. DR 2x32gigs @ 6600MT/s 1:1 mode =70.17ns
3. SR 2x16gigs @ 8000MT/s 2:1 mode = 70.24ns
4. DR 2x32gigs @ 8000MT/s 2:1 mode = 71.84ns

Bandwidth read-modify-write (ADD) ranking:

1. SR 2x16gigs @ 8000MT/s 2:1 mode= 97.11GB/s
2. DR 2x32gigs @ 8000MT/s 2:1 mode = 92.87GB/s
3. SR 2x16gigs @ 6600MT/s 1:1 mode = 91.23GB/s
4. DR 2x32gigs @ 6600MT/s 1:1 mode = 87.34GB/s

A few comments in random order to my findings above

A single 8core Zen4 CCD can take advantage of the higher bandwidth afforded by 2:1 mode vs 1:1 mode, even if the common misconception on many forums is that there is no benefit because they can hardly see any difference in gimmicky AIDA64 memory bench. (its also easy to double check this in other benchmarks such as y-cruncher / GB3 membench which will show the same)

The next question would naturally be what's the "best memory setup", 1:1 mode with its lower latency or 2:1 with its higher bandwidth. There is no easy answer for this as it all depends on what benchmark/game you comparing the numbers in.. Some will prefer latency while others bandwidth, so you just have to check on an individual basis.

But what i can say is that i pretty much always think higher memoryspeed is better, be it in 1:1 mode or 2:1 mode... From time to time i see some limit themself to something like 6000/6200MT/s because they think its faster in games than say 6400MT/s for some reason (?)

My next observation is that i did not find any bandwidth benefit from the "dual rank" (quad) in Clam cache/mem benchmark, but karhu is seemingly showing higher mb/s. But i suspect this is because the higher memory size tested, not increased bandwidth from DR. I will do some more DR karhu runs where i limit used memorysize to same as SR and check if the numbers change. (y) edit Its also possible the forced GDM enabled with DR is eating up the bandwidth benefit compared to SR

Have also seen some complains about some ppl having a hardtime tuning memory on the 1.1.7.0 PatchA FireRangeP AGESA, i can only say that is working pretty good for me on the ASUS GENE, even if i'm using a beta bios. But be warned, stabilizing DR 64gigs @ 8000MT/s is still insanely hard, think i spent like 5x the time on this profile compared to all others combined... Its really on a razors edge, +-5 mv on some rails and you can forget about 10k karhu.

As a follow up to my post above:
The same is true with a dual CCD cpu, the bandwidth difference is only bigger here

• 32gigs a-die @ 2:1 mode 8000MT/s GDM disable CL32-45-38-44-82 (synced)
• Memory sticks are cooled by barts custom copper + two noctua 60mm fans @ 1.74v VDD + 1.6v VDDQ (have ran these voltages for ~16 months now)
• FCLK @ 2200mhz (seems no amount of voltage finetuning will allow this cpu to run higher FCLK)

Stability tests completed in same boot instanse:

• ~42k karhu @ 341.760mb/s
• 27 cycles testmem5 1usmus cfg (minimum 90min runtime in options)
• 1 hour OCCT memory stresstest (think i saw some comments about it above so i decided to include it
• 1 hour y-cruncher all tests

Some performance numbers included also:

• 826 GFLOPS average in linpack extreme with daily tune
• Clam cache and memory benchmark latency = ~72ns
• Clam cache and memory benchmark ADD bandwidth = ~97 GFLOPS (seems like there is no difference between dual and single CCD in this test)
• Hwinfo open since i know some really values these

------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

• 32gigs a-die @ 1:1 mode 6600MT/s GDM disable CL26-37-32-30-62 (synced)
• Memory sticks are cooled by barts custom copper + two noctua 60mm fans @ 1.85v VDD + 1.6v VDDQ (have ran these voltages for ~16 months now)
• FCLK @ 2200mhz (seems no amount of voltage finetuning will allow this cpu to run higher FCLK)

Stability tests completed in same boot instanse before system froze

• ~21k Karhu @ 317.558mb/s
• ~8.5 hours RunMemTestPro @ 1420mb/s
• Completed 12 cycles Testmem5 before NVMe died
• Performance numbers for Clam latency and ADD bandwidth

 

[StefanR5R]

@igor_kavinski, thanks for the data. Here is another way of plotting it.



(My representation of the tail end may not be accurate.)

 

[gaav87]

But what i can say is that i pretty much always think higher memoryspeed is better, be it in 1:1 mode or 2:1 mode.

Very interesting read im still on two 5800x3d @4000cl16 since day1
Wonder how 1:1 vs 2:1 compare in gaming heavy RT/CPU games.
Will need to test this myself on 9700x i guess. Would be awesome to see some jedi survivor / dragon dogma 2 if you have them.

 

[igor_kavinski]

• Completed 12 cycles Testmem5 before NVMe died

High FCLK killed it?

 

[tsamolotoff]

No.

Well, if you don't believe me you welcome to do what ppl at i2hard (as well as derbauer) did, they actually have multiple CPUs and delidded them. Same stuff happens to intel cpus, sometimes the difference is even greater, like 20C or so.

Here's another 7950x

 

[igor_kavinski]

View attachment 102970

Guess classroom and junkshop are pretty heavy scenes and don't respond that well to more watts.

 

[Det0x]

High FCLK killed it?

No it was a agesa/bios problem..NVMe simply just went on a spring break, luckly its back alive now

You can check out these 3 posts if you want to learn more:

AMD DDR5 OC And 24/7 Daily Memory Stability Thread

www.overclock.net

AMD DDR5 OC And 24/7 Daily Memory Stability Thread

www.overclock.net

AMD DDR5 OC And 24/7 Daily Memory Stability Thread

www.overclock.net

 

[Abwx]

Meanwhile we're leaning that 13k and 14k chips from Intel that were operating at lower (more usual) temperatures have been having stability issues or that the silicon was degrading.

You can check the max temp of AMD and intel CPUs here, 13xxx and 14xxx reach higher temps than AMD chips, and that s with stock RAM speeds :

Intel Core i9-14900KS im Test: Anwendungs-Benchmarks, Leistungsaufnahme & Temperatur

Intel Core i9-14900KS im Test: Anwendungs-Benchmarks, Leistungsaufnahme & Temperatur / Testsystem und Testmethodik für Anwendungen

www.computerbase.de

 

[tsamolotoff]

Can't say, I wasn't able to successfully get either of my systems to run at a higher memory clock so I haven't had any hands-on experience with it.

In a Navier-Stokes sparse matrix calculator that I use:
1:1 mode 6400 / 2167 - 16:05 min to fully calculate a model
1:2 mode 7800 / 2167 - 14:11 min to do the same, primary timings are roughly the same, just more frequency.

 

[igor_kavinski]

primary timings are roughly the same, just more frequency.

Surely the primary timings for 6400 could've been lower than for 7800?

 

[tsamolotoff]

Surely the primary timings for 6400 could've been lower than for 7800?

Same in terms of nanoseconds, numerically they are different, naturally. Like CL30 at 6400 is equivalent to CL37 at 7800 and so on.