Discussion Zen 5 Builders thread

[Det0x]

In that video only half of the boards were hitting 8000

View attachment 121577

"with todays more mature agesa/bios"
How close what this first video to the X870E release ?

Will be fun to see their new updated results 👍

 

[igor_kavinski]

Can complete pretty much all benchmark @ golden ratio 8800/2200 CL32 (8844/2211), but karhu will fail it after awhile, so the promised land is still alittle out of reach for fully stable daily settings atm

CL32 seems pretty low and ambitious for 8800. Why not relax it up to CL38 for stability?

 

[Det0x]

CL32 seems pretty low and ambitious for 8800. Why not relax it up to CL38 for stability?

If i run CL32 or CL50 dont really matter, IMC on the cpu is the limting factor, not the memory

But for those interested in how much there is to gain from going "synced" (mostly get lower latency compared to 8744MT/s quoted above)
Some benchmarks at golden ratio 8800/2200 (actually 8844/2211 since 100.5mhz baseclock)

• 9950X3D with core tuning = legacy running PBO
• ROG CROSSHAIR X870E APEX @ baseclock 100.5mhz
• G.SKILL 6000MT/s CL26 @ 1.78VDD 1.55VDDQ cooled by custom heatsinks + 2x60mm noctua fans
• FCLK @ 2211mhz / memory @ 8844MT/s CL32 (2200 /8800 multipliers with 100.5mhz baseclock)
• Nitro 1-3-1 with synced tPHYRDL @ 38/38
• Large custom watercooling with a mora

 

[igor_kavinski]

4 easy installments @ 0% interest and...



Sorry to some poor bloke in the UK. He was too slow to grab this CPU. Saw it on my local Amazon being shipped by Amazon UK and the chances of prices going higher seemed higher so had to bite. Damage = $872

The above screenshot is at FCLK 2133 UCLK 1800 7200C34. It booted at 2200 but Windows installation froze. Then at 2167, everything seemed fine but y-cruncher failed so had to settle for 2133.

2167 7200C60 benchmarks below:

Spoiler

Subjectively, it boots REALLY fast. Mobo is ASROCK B850M RS Pro Wifi.

 

[DrMrLordX]

Mobo is ASROCK B850M RS Pro Wifi.

ASROCK? Uh oh. Have they patched out all the failed X3D problems yet?

 

[igor_kavinski]

ASROCK? Uh oh. Have they patched out all the failed X3D problems yet?

I'm on BIOS ver 3.15 and no intention of upgrading unless I encounter an issue.

 

[igor_kavinski]

Compared to the 245KF, installing the AIO on AM5 socket was much, much easier and straightforward, not to mention quicker. With the 245KF, I needed to keep one hand behind the mobo to hold and keep the backplate stable while I screwed from the front. It was challenging, to say the least.

 

[In2Photos]

Compared to the 245KF, installing the AIO on AM5 socket was much, much easier and straightforward, not to mention quicker. With the 245KF, I needed to keep one hand behind the mobo to hold and keep the backplate stable while I screwed from the front. It was challenging, to say the least.

Tape is your friend in situations like these! Put some on the backplate to keep it from moving while attaching the AIO. Remove it when finished.

 

[Markfw]

Tape is your friend in situations like these! Put some on the backplate to keep it from moving while attaching the AIO. Remove it when finished.

using some anti-static foam or bubble wrap as a spacer to the case is better. tape can leave residuals or take or damage some of the motherboard traces when you remove it.

 

[In2Photos]

using some anti-static foam or bubble wrap as a spacer to the case is better. tape can leave residuals or take or damage some of the motherboard traces when you remove it.

Blue painters tape, electrical tape, or Kapton tape will not leave a residue or remove any traces.

 

[igor_kavinski]

On my particular 9950X3D sample, I'm getting fixed 5.35 GHz (task manager) frequency on the cache CCD and 5.34 GHz on the frequency CCD, in an LLM workload that pegs the CCD utilization at almost 94%.

Anyone with good Ryzen tuning experience wanna chime in on how I can persuade the frequency CCD to hit close to 5.6 or even 5.7 GHz?

 

[igor_kavinski]

Having a good day today. Curve offset at -37 and my system is happily crunching away in LM Studio at fixed 5.35 GHz 6200C34 with FCLK 2133/UCLK 3100.

I really like how easy it is to tune using Ryzen Master. Unlike XTU, it sets a global setting and overrides the BIOS ones. If the Ryzen Master settings turn out to be unstable and fail to load Windows, you can boot into BIOS and change any setting to tell the mobo to override the Ryzen Master settings and you get your Windows back. I had some serious and scary BSODs but so far, have avoided trashing my Windows installation completely so that's been awesome. I really love the convenience of not having to go into BIOS to tweak and tune.

Somewhat unhappy to report that V-cache does not improve performance in LM Studio

Pinned to V-cache CCD @ 5.35 GHz: 1.47 tokens/sec

Pinned to Freq CCD @ 5.34 GHz: 1.46 tokens/sec

 

[Makaveli]

Somewhat unhappy to report that V-cache does not improve performance in LM Studio

Pinned to V-cache CCD @ 5.35 GHz: 1.47 tokens/sec

Pinned to Freq CCD @ 5.34 GHz: 1.46 tokens/sec

If shouldn't the only way to improve performance is with gpu.

using the processor for anything in LM studio is slow including Epyc processors.

It looks like you recorded this with a cell phone?

Does that intel gpu not allow you to record video?

 

[igor_kavinski]

Does that intel gpu not allow you to record video?

I was having some internet issues (PC was out of wifi range. now fixed with USB extension cable) so decided not to use the Snipping tool (would've had to use flash drive to transfer to laptop that is working fine with internet).

I tried with the ARC A770 GPU. Unfortunately, it's slower than the CPU. The only advantage seems to be to use the GPU memory to load a larger model in case your system RAM is insufficient. But I suspect LM Studio is badly optimized for any non-Nvidia GPU. It doesn't even support AVX-512 which I can see llama.cpp supports. LM Studio uses the llama.cpp runtime but I think it's using AVX2 build only.

 

[MS_AT]

I'm getting fixed 5.35 GHz (task manager) frequency on the cache CCD and 5.34 GHz on the frequency CCD, in an LLM workload that pegs the CCD utilization at almost 94%.

Task manager is less than ideal for this. It will show you high utilization even though the cores are waiting for memory. As it the case here, so the frequency is high.

Anyone with good Ryzen tuning experience wanna chime in on how I can persuade the frequency CCD to hit close to 5.6 or even 5.7 GHz?

I am no overclocker but you should first fix your cooling, then tweak current / power limits. But if you are interested in just seeing 5.7 GHz, just write a program to run a series of no_ops

Somewhat unhappy to report that V-cache does not improve performance in LM Studio

You are looking at token generation which is bottlenecked by memory bandwidth as it is a streaming workload. L3 won't help much.

And you should get few percent higher score if you split the load to both CCDs to maximize memory bandwidth.

The only advantage seems to be to use the GPU memory to load a larger model in case your system RAM is insufficient.

That sounds completly opposite to how it should be Read around llama.cpp github, they should have hints there how to properly split the workload between cpu/gpu. Unfortunately I don't have any links at hand to share.

 

[igor_kavinski]

I am no overclocker but you should first fix your cooling, then tweak current / power limits. But if you are interested in just seeing 5.7 GHz, just write a program to run a series of no_ops

Cooling fixed already (as much as I could). It's no longer being annoyingly loud. I guess it was the pump complaining because it likes being on top

It's a mATX case so putting it on top was out of the question (no space). So got a movable case trolley for ~$10, put the case on its behind on that and now the AIO cooler is happy. Going to post more benchmarks soon. I'm up to 1.77 tokens/second (it should be higher in less demanding models. I'm using the 73B Athene V2 Chat). Previous best was 1.73 and it fluctuated a lot. I guess this will have to do because I don't want a bigger case and don't want to spend extra until I can get a great deal on sub-ambient cooling