Does Thuban benefit from lower CAS latency or higher frequency RAM?

[DrMrLordX]

Depends on how many threads are spawned by the application in question. Heavily-threaded apps may work a bit better on the Thuban versus Lynnfield or even LGA1136 Nehalem (think i7-920). On anything that relies on four or fewer cores, you can probably expect better performance from Lynnfield or Nehalem.

 

[Chicken76]

@DrMrLordX
Ouch! That bad, eh?

 

[DrMrLordX]

Well, when you consider how strong an i7-920 is even today, a chip that can trade blows with the 920 on MT workloads (video encoding, for example) really isn't that bad at all. You can still do a lot with a 4 ghz Thuban. AM3+ often helps Thuban users achieve higher clockspeeds as well versus the original AM3 socket that played host to the chips. But really, the chips are ~4 years old now, and the industry has moved on to other things. Thuban lags behind on support for newer instruction sets and so forth and so on.

 

[Chicken76]

@DrMrLordX
But talking single thread performance, a 4 GH Thuban/Zosma is only comparable to a 3 GHz Lynnfield?

 

[FlanK3r]

the best for Thuban is around 1866-2000 MHz with most tight timings what can do+CPUNB jump 2800-3000 MHz.

 

[DrMrLordX]

@DrMrLordX
But talking single thread performance, a 4 GH Thuban/Zosma is only comparable to a 3 GHz Lynnfield?

Eh, depends on the app. SuperPi? Definitely. Other apps, well, Thuban is hit-or-miss depending on what you are doing. If you want an idea of what happens when Thuban faces off against Nehalem, check Anandtech bench for a decent idea. While imperfect, Bench does serve as a decent archive of CPU matchups of ages past. Look at Cinebench R10 single-thread and you'll see that Thuban represents itself pretty well there, even in single-threaded performance. Sure, it has a small clockspeed advantage, and turbo is in effect, so the comparison needs to be taken with a grain of salt.

When overclocking, I would definitely take 4 ghz Nehalem over 4 ghz Thuban for single-threaded performance, but I don't think it's going to be 33%, especially not if the Thuban has a tuned NB and the right RAM. Both Nehalem and Thuban have performance available from NB/uncore tuning and RAM tuning, but overall I think you can squeeze more out of the Thuban chip in this fashion. Look at what Flank3r posted to see where a Thuban tune should wind up.

 

[Chicken76]

@DrMrLordX
I see. I'm just don't know where to place a 4GHz Thuban/Zosma in a lineup of Intel CPUs, all stock. I understand overclocked Nehalem is faster. What about stock Nehalem, Lynnfield? How much behind is it compared to a stock Sandy Bridge i5?

 

[SlowSpyder]

@DrMrLordX
I see. I'm just don't know where to place a 4GHz Thuban/Zosma in a lineup of Intel CPUs, all stock. I understand overclocked Nehalem is faster. What about stock Nehalem, Lynnfield? How much behind is it compared to a stock Sandy Bridge i5?

I think PhII had slightly better than C2D IPC, generally. But not a huge difference.

 

[NTMBK]

Out of curiosity, how much performance am I leaving on the table with slow RAM? I have a Phenom II 965 paired with 6GB of DDR2-800 at 6-6-6-18. The timings kind of suck because I have a mismatched set of cobbled together RAM (a 2x2GB kit, and half of a 4x1GB kit). If I were to replace the whole lot with an 8GB DDR2-1066 kit, would I notice any appreciable performance improvement?

 

[SlowSpyder]

Out of curiosity, how much performance am I leaving on the table with slow RAM? I have a Phenom II 965 paired with 6GB of DDR2-800 at 6-6-6-18. The timings kind of suck because I have a mismatched set of cobbled together RAM (a 2x2GB kit, and half of a 4x1GB kit). If I were to replace the whole lot with an 8GB DDR2-1066 kit, would I notice any appreciable performance improvement?

I had my Thuban on AM2+ with DDR2 @ 866MHz, 4-4-4-12 though. I don't have any exact numbers, but everything I saw when looking at this showed the difference between DDR3 on PhII and DDR2 on PhII was very little.

 

[DrMrLordX]

DDR2 really isn't going to hold you back that much, so long as you keep the timings tight. About the only problem with DDR2 is that DDR3 has had much more time to mature as a technology. You can probably get much better overall latency from DDR3 platforms, and you can do it with larger amounts of RAM to boot (assuming the board supports it).

 

[NTMBK]

Oh, I won't be upgrading to DDR3- this platform is too old to bother sinking money into a new mobo.

 

[schmuckley]

Thanks guys.
Do you think it's worth overclocking the IMC past 2400 ? I achieved 2400 1year+ stable with a very-very small voltage increase. 2600 might require a bit more voltage.

Yes! 2800-3200Mhz is the goal!
I set to 1.37v on cpu/nb straightaway.

o..ddr2..hmm..eh

 

[DrMrLordX]

Oh, I won't be upgrading to DDR3- this platform is too old to bother sinking money into a new mobo.

Agreed, not really worth it. It's just an academic consideration, perhaps for someone with a DDR3-capable AM3/AM3+ board hosting a Thuban or something perhaps.

 

[Chicken76]

Another related question: is it worth increasing the HT-Link? From what I understood it connects to the PCI-E bus, and unless you're using multiple GPUs or a top-end one the result is neglijable, but with a significant increase in power consumption and heat output.

 

[DrMrLordX]

No. HT-Link mostly controls available bandwidth to off-die components such as the southbridge. Maybe if you are running SLI or Crossfire, sure, you could use a few extra mhz there . . . maybe. But mostly it's just a source of instability that provides no gain.

Keep your HT speed between around 1800-2000 mhz or so and you should be fine. Sometimes you might have to fudge a bit if you are overclocking with htt. If you are overclocking by multiplier, then just set it to 10x multiplier for 2000 mhz HT and leave it alone.

 

[lopri]

The bottom line for optimal performance for Thuban is: Rated memory frequency x 2 > CPU-NB frequency.

For 2000 MHz CPU-NB, DDR3-1066 or higher.
For 2500 MHz CPU-NB, DDR3-1333 or higher.
For 3000 MHz CPU-NB, DDR3-1600 or higher.

Then tighten the heck out of timings.

It is better to use HTT (system bus) overclock than CPU multipliers for overall performance if your hardware can handle it. Note CPU-NB frequency should not exceed CPU frequency.

 

[DrMrLordX]

The bottom line for optimal performance for Thuban is: Rated memory frequency x 2 > CPU-NB frequency.

For 2000 MHz CPU-NB, DDR3-1066 or higher.
For 2500 MHz CPU-NB, DDR3-1333 or higher.
For 3000 MHz CPU-NB, DDR3-1600 or higher.

Then tighten the heck out of timings.

As an addendum, let me add that there is a practical minimum below which timings can not go. Usually your BIOS won't let you go below CAS6 or 5 with DDR3, for example. The lowest CAS I've ever actually achieved with DDR3 is CAS6. So, if you see that your timings are rock bottom and the RAM is still stable, the only avenue left for continued memory performance improvement is to raise RAM clockspeed.

It is better to use HTT (system bus) overclock than CPU multipliers for overall performance if your hardware can handle it. Note CPU-NB frequency should not exceed CPU frequency.

Darn tootin. And yeah, don't try to push your NB to a higher clock than your CPU. It just doesn't work. Believe me, I've tried!

 

[Chicken76]

It is better to use HTT (system bus) overclock than CPU multipliers for overall performance if your hardware can handle it.

I don't understand this. Can you explain? Isn't everything a multiple of this frequency? I only see one advantage: reaching values between what the multipliers offer, for example reaching 3050 on the core instead of either 3000 or 3100. If so, the advantage is minimal for the core and small for the NB.

 

[SlowSpyder]

I don't understand this. Can you explain? Isn't everything a multiple of this frequency? I only see one advantage: reaching values between what the multipliers offer, for example reaching 3050 on the core instead of either 3000 or 3100. If so, the advantage is minimal for the core and small for the NB.

Using the system bus to overclock does give you the ability to fine tune a bit. At stock speeds, even a half multiplier adjustment means 100MHz, so there may be some more clockspeed available somewhere between 0-100MHz, using the system bus allows someone to try and find that speed.

But that isn't the main advantage. The big advantage is that when you use just the multipliers to overclock, the only thing being adjusted is the CPU's cores. When you use the system bus to overclock, you speed up the CPU and everything it communicates with. As the CPU goes faster, at some point getting data in and out faster will benefit you too. The problem with using the system bus is that there are a lot more variables.

I think the best approach is to leave everything but the CPU multiplier stock and see how far the cores can go. Then, lower the core multiplier and see how far you can take the NB/L3 and memory separately. Once you have those values, you can find some combination of multiplier, NB/L3, memory speed that is within those maximum values but close to it.