- Mar 3, 2017
- 1,777
- 6,789
- 136
Page 712 - Discussion - Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
Page 712 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
forums.anandtech.com
For this one the frequency difference is way larger so even if the ISA/arch is almost same the max frequency is going to affect the perf of C
Markfw
Moderator Emeritus, Elite Member
- May 16, 2002
- 27,171
- 16,064
- 136
You need to look at actual performance numbers. In reality, they do not see anywhere close to those numbers. Look at some benchmarks and compare to benchmarks of p-core with ecore competitions chips. Also, when the arch is the same vs WAY different even a 20% clock makes almost no different.@MarkfwPage 712 - Discussion - Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
Page 712 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
For this one the frequency difference is way larger so even if the ISA/arch is almost same the max frequency is going to affect the perf of C
View attachment 113499
You can look at these
Markfw
Moderator Emeritus, Elite Member
- May 16, 2002
- 27,171
- 16,064
- 136
Yes, and they are very close. Whats your point ?You can look at these
It's the full Zen core affecting performance if we have 2 in isolation we will know oh i just realised
@511,
the maximum clock speed limit of AMD's dense cores does not matter. The mobile CPUs with dense cores as well as the server CPUs with dense cores are power-limited, not clock speed limited.
the maximum clock speed limit of AMD's dense cores does not matter. The mobile CPUs with dense cores as well as the server CPUs with dense cores are power-limited, not clock speed limited.
Not trying to derail this thread, so if this becomes a subject, we should create a new one.
I know 1.45 is safe and I seem to recall 1.5 is safe. The issue becomes heat. The higher you go, the hotter those things get. Heat causes errors in RAM.
RAM overclocking is also a PITA. On my board, if you mess up too bad, you have to open the case, remove hardware, and hit the reset CMOS button. I added a button to the jumpers so I only have to take the side off...
At least for me, the safe mode failures almost never work and I don't have buttons on the back of my ITX board.
BTW I have one of the worst 7950Xs out there, so I'm sure others could have better results.
On Topic: That is why you shouldn't beat up AMD over the current situation. DIMMs today are very loosely binned. Many of the slowest DIMMs can run at DDR5-8000 or more. Intel and AMD are pushing the industry is catching up to this, but it takes time. On most AM5 CPUs, DDR5-8000 is easily achievable with the right kit, but speeds alone provide almost no benefit thanks to the GMI links and stuff. If they design around this, we'll see significant uplift. They are for some designs that use LPDDR5, and other designs in the future, so I'm looking forward to seeing results. I've said in the past I think Zen 5 is bandwidth bottlenecked. Hopefully we'll see something with more, or maybe with Zen 6.
Yes that is nice, I've been planning to attempt to stabilize my 64gb at CL28 soon, and they are making my result look pathetic! Some details below...
Voltages. I have 64gb, I haven't yet gotten 6400 stable, I COULD try, but I also can't get a 2200mhz IF clock to work so I am working on timings, and making great progress. At the default 1.4V of my dual rank chip, I can't do much, bumping it up to 1.45 gives sooo much flexibility. My timings are now even different from last screenshot and aida64 reports a 55-57ms latency. You can't stick to anything 1.4 and below and expect an improvement, at least for dual rank.
I know 1.45 is safe and I seem to recall 1.5 is safe. The issue becomes heat. The higher you go, the hotter those things get. Heat causes errors in RAM.
RAM overclocking is also a PITA. On my board, if you mess up too bad, you have to open the case, remove hardware, and hit the reset CMOS button. I added a button to the jumpers so I only have to take the side off...
At least for me, the safe mode failures almost never work and I don't have buttons on the back of my ITX board.
BTW I have one of the worst 7950Xs out there, so I'm sure others could have better results.
On Topic: That is why you shouldn't beat up AMD over the current situation. DIMMs today are very loosely binned. Many of the slowest DIMMs can run at DDR5-8000 or more. Intel and AMD are pushing the industry is catching up to this, but it takes time. On most AM5 CPUs, DDR5-8000 is easily achievable with the right kit, but speeds alone provide almost no benefit thanks to the GMI links and stuff. If they design around this, we'll see significant uplift. They are for some designs that use LPDDR5, and other designs in the future, so I'm looking forward to seeing results. I've said in the past I think Zen 5 is bandwidth bottlenecked. Hopefully we'll see something with more, or maybe with Zen 6.
Oh, the main reason I came here was to say this: Looking forward to see the 8800S performance compared to the mobile 3070. Strix halo is looking like an amazing upgrade to my Razer laptop, which has a 5900hx and a 3070.
Not in the Zen 5 generation.
And for Zen 6, it remains to be seen whether or not dense cores play a role in top-of-the-line (highest power) desktop SKUs, but this belongs into a different thread.
Edit, there is also a dedicated thread to discuss Intel's vs. AMD's approaches towards inhomogeneous CPUs.
Last edited:
adroc_thurston
Diamond Member
- Jul 2, 2023
- 6,552
- 9,221
- 106
all the MRDIMM-8k goodness and Granite Rapids is a joke even in HPC. Too bad!
Win2012R2
Golden Member
- Dec 5, 2024
- 1,183
- 1,213
- 96
Looks like my next upgrade, pity that cheapMD did not give both chiplets with 3D cache - $999 would have been fair price.
Win2012R2
Golden Member
- Dec 5, 2024
- 1,183
- 1,213
- 96
"For HPC workloads AMD recommends
1) disabling SMT
2) engaging the "high performance" power profile,
3) running in performance determinism mode
4) running the respective CPU at its maximum configurable TDP (cTDP) value
5) running with four NUMA nodes per socket (4 NPS)"
#1-4 are obvious even without any guides (had them for long time on older EPYCs), will give #5 a go once I get my hands on new stuff, been suggested on here too.
Those results suggest power determinism is faster.
Win2012R2
Golden Member
- Dec 5, 2024
- 1,183
- 1,213
- 96
Here is what I want in top 16 core product (and that's not negotiable):
1) both chiplets of same top quality - none of the "second chiplet is from the B-team"
2) both chiplets to have 3D cache
$999 product, yes, easy to make and will be plenty of willing 5090 owners to get the best of the best - right now I had to disable one chiplet in 7950X3D to deal with crappy scheduling, plus this baby will work out great in a small power efficient server.
Give bloody choice to people to buy stuff they want AMD!
Win2012R2
Golden Member
- Dec 5, 2024
- 1,183
- 1,213
- 96
Win2012R2
Golden Member
- Dec 5, 2024
- 1,183
- 1,213
- 96
Why would it be worse if 3D chiplets 1T perf is very close if not equal in 9950X3D to 2nd chiplet.
People (like me) bought 7950X3D with far worse delta, pick the best chiplets for Halo product and it will sell - there will be millions of 5090s sold and people who buy them will want top end CPU, buying 8 cores is 2025 is really not acceptable proposition and if someone buys 16 they should be top notch stuff, even if it's 9999X3D version for $999.99.
This is especially important in light of Zen 5 clearly being mem bandwidth starved - both chiplets with 3D should help nicely distributed workloads.
It is not. X3D chiplets are only marginally higher clock rate than last generation, they're still behind.
Win2012R2
Golden Member
- Dec 5, 2024
- 1,183
- 1,213
- 96
How do you know?
In 9800X3D people get 3D chiplet to clock 5.4-5.6. What makes you think a more premium product won't have 100-200 mhz bump for 3D chiplet?
That's already close anyway, I'd rather have a 3D chiplet clocking 5.4 ghz, than a non-3D clocking 5.7 ghz - especially for mem bandwidth starved Zen 5.
biostud
Lifer
- Feb 27, 2003
- 19,859
- 6,959
- 136
For the same price I would choose dual vcache over a slight speed bump on the single CCD. But as I've said before I'm sure AMD could sell a dual vcache 9970X3D for +$150-200 over a 9950X3D.
Overclocking does not correlate with what AMD will ship.
TRENDING THREADS
-
Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)
- Started by DisEnchantment
- Replies: 25K
-
Discussion Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads
- Started by Tigerick
- Replies: 22K
-
-
Facebook
Twitter