Is it normal to not be able to hit x44 Cache Ratio on a 4790K?

[Quad5Ny]

So my 4790K running at 4.4GHz (all cores) crashes almost instantly if I set the cache ratio above x43.

In order to get x44 stable I need to add another 100mV (going from 1.2V to 1.3V) - which seems a bit overkill. Is there another voltage I should be adjusting instead?

Rig in question: Asus Z87-Pro, BIOS 2103| 32GB 1866 (9-9-9-27)

---

On a side note the CPU cores are happy with a -67mV VID offset which gets me a stock VID of 1V (1.008V measured) and very cool running chip. :hmm:

 

[Bubbleawsome]

Cache ratio should stay around stock. Don't bother with it. Overclock the core as high as you want/can, and then go until cache is unstable. Cache ratios can hurt a core OC without helping performance.

That also looks like it could be a nice chip. Want to give it to me?

 

[crashtech]

I settled on x40 for mine. It doesn't seem to be worth the trouble to try and push it.

 

[Lepton87]

My HW-W does 1:1 ratio and some benchmarks see marked improvement but how it translated into real world performance is unknown to me. I list some improvements from stock 3.0GHz to 4.25GHz, at this speed it was clocked at full core clock just like in SB.
Sisoft Sandra

Stock to OC: Memory bandwidth 49.15 to 54
Latency: 24.4ns to 22.9ns that's actually a 7% lower latency
Cache Bandwidth L3 268GB/s to 324GB/s that's pretty disappointing scaling of cache performance, only 20% more and the increase in cache frequency is 41%. VERY low efficiency.

Memory Transaction Throughput

MTT: 7.44 to 7.99

HLE: 6.68 to 8.3

SFT: 5.41 to 6.71

I didn't do any real world test because it turned out that the CPU is just too hot to OC properly on AIR cooling so I started assembling a custom loop and it took a lot of time to buy all the parts not to mention money(for the price of the CPU and the loop I could easily buy 5960X) and even then there was one screw missing which delayed the deployment for at least a day.

 

[Ryanrenesis]

VRIN and RING voltage need to be upp'd when dealing with high processor cache multipliers.

 

[crashtech]

I don't know if other users have experienced this, but I found that when pushing the ring multi too far, my machine would not simply BSOD or freeze, rather it would spontaneously reboot. I got a little tired of it.

 

[Lepton87]

I forgot to answer your question, yes it's very normal to not be able to hit 4.4GHz on cache. The best samples hit 4.6 and some don't even hit 4 and what's more how well cache overclocks is independent of how well your core clocks overclocks. The core clock is vastly more important if it wasn't they would change it back from SB to the way it was in Nehalem.

 

[wilds]

Are there any games or real world applications with a noticeable improvement with cache overclocking?

 

[Quad5Ny]

Cache ratio should stay around stock. Don't bother with it. Overclock the core as high as you want/can, and then go until cache is unstable. Cache ratios can hurt a core OC without helping performance.

That also looks like it could be a nice chip. Want to give it to me?

My current clock/voltage settings are:
• Core: 4.6GHz , -10mV Offset (As measured: 1.248V/1.264V AVX)
• Cache: 4.4GHz, 1.270V Adaptive (As measured: 1.291V)
• Temps @ 24c Ambient w/Noctua NH-U14S (Variable RPM): P95_S.FFT: 71c, P95_S.FFT-AVX: 75c, IBT: 89c

^Which are completely stable with a mix of P95, IBT, AIDA64 & POV-Ray for 8 hours. As of right now I have no reason to push it higher, x46 does everything I need. ^_^

I was just confused when I couldn't do x44 Cache at stock voltage. Some light Google searches had me to believe x44 was the default - even some of Asus's blog post mentioned their auto rules set the cache to 4.4GHz by default.

VRIN and RING voltage need to be upp'd when dealing with high processor cache multipliers.

VRIN, you mean VccIN (the voltage fed to the FiVR)? I've got that at 1.808V w/ LLC at level 7, it doesn't fluctuate at all.

I don't know if other users have experienced this, but I found that when pushing the ring multi too far, my machine would not simply BSOD or freeze, rather it would spontaneously reboot. I got a little tired of it.

Same here, although sometimes I would get a blue screen and then the motherboard would go NOPE we've lost control and power off.

 

[Quad5Ny]

I forgot to answer your question, yes it's very normal to not be able to hit 4.4GHz on cache. The best samples hit 4.6 and some don't even hit 4 and what's more how well cache overclocks is independent of how well your core clocks overclocks. The core clock is vastly more important if it wasn't they would change it back from SB to the way it was in Nehalem.

Thanks

Are there any games or real world applications with a noticeable improvement with cache overclocking?

Not really but how many people with a 4.4GHz haswell part are overclocking because they ACTUALLY need more performance.

See: Haswell Overclocking Guide [With Statistics] - Scroll down to "Ring Bus Doesn't Matter [Evidence]"

 

[Quad5Ny]

Just to derail for a second, I wonder how Intel decides the voltage bins for CPU's? Do you guys think they have a really expensive oscilloscope jig that can do 4+ GHz and they keep increasing the voltage until the waveform is stable?

 

[Lepton87]

Thanks




See: Haswell Overclocking Guide [With Statistics] - Scroll down to "Ring Bus Doesn't Matter [Evidence]"

I didn't see many games tested, some games are very sensetive to cache like WoW where Xeon E5 at 3.1 is faster than 3960X at 3.5GHz because it has 20MB of cache and 3960X has 15MB I would like to see those games tested on HW-E because its cache is clocked very conservatively at default at only 3GHz that's barely more then nehalem and HW-E cache clocks almost or as good as the regular HW cache so 3GHz vs 4.5GHz cache tests would be interesting to see, alas no one has done such tests in a wide variety of games. I would really like to see that tested. I could do it if I had more games and more experience in testing games, not all games have a built-in benchmark, in fact much more games don't have it than have. It's a sad state of matters because it would really come in handy for many uses like listing system requirements in both generally useless cpu generation like i7 or Intel Quad at 2GHz and then listing the performance needed in that test. Then the producer could easily create a database of computers and people could easily check how their computer fares. Much easier than comparing an Intel Quad at 2GHz to an i3 655.

 

[Quad5Ny]

I didn't see many games tested, some games are very sensetive to cache like WoW where Xeon E5 at 3.1 is faster than 3960X at 3.5GHz because it has 20MB of cache and 3960X has 15MB I would like to see those games tested on HW-E because its cache is clocked very conservatively at default at only 3GHz that's barely more then nehalem and HW-E cache clocks almost or as good as the regular HW cache so 3GHz vs 4.5GHz...

I completely forgot about the E parts.

The 5920K and 5930K stock cache frequency is 3GHz? That seems very low (700MHz lower than Turbo). As for the 5960X I think its fair considering the base frequency is only 3GHz.