Madshrimps Forum Madness - [WORKLOG]

Madshrimps Forum Madness (https://www.madshrimps.be/vbulletin/)

- Hardware Overclocking and Case Modding (https://www.madshrimps.be/vbulletin/f10/)

- - [WORKLOG] - HD3300 overclocking (https://www.madshrimps.be/vbulletin/f10/worklog-hd3300-overclocking-64000/)

[WORKLOG] - HD3300 overclocking

First post of the worklog. Here, I will post my most up to date results :-).

: 3DMark01 - 29713

: 3DMark03 - 9038

: 3DMark05 - 7048

: 3DMark06 - 3149

: 3DMark Vantage - 716

: Aquamark - 76223

a copy of http://www.madshrimps.be/vbulletin/f...3-720be-63393/ ? or more extreme OC planned? :)

More extreme planned. I could use some of the posts of that thread, though :p

126 points away from Gold Cup:)

Biggest problem is the high memory latency

what cooling on the GPU currently?

Still on air cooling.

I'm trying to reduce the memory latency by reducing the memory divider from the very slow 1066 to 800, but it seems that the increase in HTT frequency (240 -> 325) makes the sideport memory to act up. I have a hypothesis about the problem.

My motherboard is capable of running 533MHz on the sideport memory. Anything higher than 533 (which is the 667 option in the bios) results in an overly artifacted screen, which is a symptom of the memory being overclocked too high. It seems to come down at 667MHz being too much for the sideport memory.

Now, the bios options offered by Asus (533/667/700/...) come across as fixed frequencies. In other words: we think we can set either 533MHz or 667MHz or and so on. Now, I have good reasons to believe that these frequencies are, as everything of the AM2 platform, derived from the HTT frequency, which can also be described as the base frequency. So instead of the three-digit values, we have to see the values as a fraction: 533/200, 667/200 et cetera.

Now, I'm unable to use the following combinations (sideport-htt):
*533/325 => 866MHz real
*533/242 => 645MHz real
*700/200 => 700MHz real

This was part 1 of the puzzle. Part 2 is the actual reason why I'm unable to run 325MHz HTT in combination with (any divider) sideport memory.

Looking at the screenshot I've already posted above, I think you should all be able to guess the reason :)

Might be usefull to add this piece of information to this thread. Originally, I found this issue while testing the M3A78-T, which is the same motherboard I'm using in this overclocking journey.

Quote:

Yesterday, I had a few benchmarks scoring incredibly low for no apparent reason. When I checked the CPU-Z memory details, however, I saw that the memory was clocked at ~480MHz CL5, instead of the 640MHz I set in the bios. Apparently, it was not possible to run with the 1067 divider.

Today, I spent a few hours figuring out what was exactly causing this bizare issue. I went over the different timings, HT Link frequency settings, different combinations and finally ... I figured it out.

To get the 1067 divider working properly at high memory MHz overclocks, enable 2T mode. Even if you're certain the memory could run 1T, it seems to be a technical issue more than a memory bottlenecked issue. If you disable the 2T mode, the bios automaticly resets the memory to the 800MHz divider, even if you've manually put 1067MHz.

By the way, in the proces of figuring out this issue, I also figured out a golden setting for benchmarking .

Two posts I made over at XS which contain a few observations, which might come in handy some day. For reference only.

Quote:

Yes, the HT Link frequency is extremely important. Never have I seen such impressive scaling of a, at first sight, insignificant frequency. When I was benching the IGP a few weeks ago, I had a boost of 1K in 01 just by increasing the HT Link multiplier by 1.

Before you start raising the HTT frequency, please think about all the frequencies that are affected by this base clock frequency. There's a reason why I handicap myself by using the 1066MHz memory divider instead of the more performant 800MHz one. And yes, I am able to run 325MHz HTT.

I've actually thrown out my 720BE and replaced it with my 940BE because it's so bad in overclocking. 3500MHz and 2500MHz NB is just too little for this IGP ... YES, the cpu was bottlenecking me :p:

I'm running Ludicrous voltages on air, to be honest, but it seems that everything is running just fine, so no worries. 1.4v vHT to answer your question.

By the way, can you set the voltage of the IGP seperate from the voltage of the actual northbridge? I can't, so I always figured I was running 1.6v through the IGP, but since you're only at 1.36v and I've been reading voltages from the board that say ~1.3v I suspect the vIGP is derived from the vNB somehow on this motherboard.

http://i154.photobucket.com/albums/s...man-/29265.png

Lods actually make a giant difference with these cards. Even a small increase can make you score go up quite a bit.

Quote:

So you are running Vnb at 1.36v then? Problem is that I can only control the Vigp through the Vnb (~your setup), but the voltage of the Vnb is around 0.3v higher than the Vigp, which means I have to go up a lot in Vnb to run 1100MHz core frequency. The heat itself isn't really a problem, but I'm not sure the NB likes that much voltage ... and I still have to perform the Vnb voltage modification.

At first, I was amazed by the CPU bottleneck as well, but it actually makes a lot of sense ... we are still benching an AMD platform here :p:.

On AM3, it's not really that interesting to search for raw memory frequency since the platform prefers latency over bandwidth. Tony made a nice comparison in the AMD section where you can see that 1333CL5 outperforms 1600CL7. Add to that the fact it's quite difficult to get DDR3 upto 900MHz without a superior IMC on your cpu. For benching purposes, I'd stick with the 1680CL6 you have running now since it seems (at least in my tests) even more performant than 1500CL5, both bandwidth and latency-wise. Could you run an Everest benchmark for me next time you run 01? I'm pretty sure you're below 38ns latency with that configuration, which blows away my 41.3ns due to the 1066MHz memory divider.

The black slots are the ones closest to the socket? If yes, I have the same observations on my board: those furthest away provide less overclockability.

Nice 01, by the way! Looking at the subtest details, I'm trailing but quite a lot in the CL and CH tests, in spite of the 200MHz cpu advantage. It's been a while since I tweaked ATI videocards, so it's possible my tweaking is off by miles, but it's also possible the high memory latency is doing it's work (or rather: doesn't) here. Judging from the LL and LH scores that are pretty much on par (LL is equal to mine :p: ) I'd say the memory bandwidth is alright on this DDR2 setup. What LOD are you using for Nature? At 1140MHz core you should be over 114-115FPS.

Playing with lod at the moment.

CH-DH-LH-DL-LL-N-CL
ALL at 2.9
CL at 3.2
N at 5.7

Testing 3DMark03 again: lods of course. LOD at 2.9 gives me a bit under 9k, at 14 a bit over 9k. The boost is located in GT3, which gains about 1-2FPS. GT1 decreases by a few FPS, though (around 4FPS)

1st spot reclaimed :)

Having an increase CPU frequency is totally pointless. The only gain is in CH (220 -> 240), but with a normal increase in GPU frequency (hello cold + volts) you could gain a lot, a lot more.

Time for finding the voltage modifications for the IGP.

AREA1

First, I measured the voltage on each pin and searched for those that change when I change the voltage in the bios:

Code:

COLOR -- HTV -- C/PCIV
 

1.2/1.1 -- 1.3/1.2 -- 1.34/1.44
 

Red = 1.14v -- 1.24v -- 1.48v

Blue = 1.50v -- 1.60v -- 1.64v

Yellow = 1.03v -- 1.08v -- 1.26v

Green = 5.07v -- 5.07v

Brown = 12v -- 12v

Bordeaux = 1v -- 1v

Grey = 11v -- 11v

Purple = 7.65v -- 7.42v -- 6.92v

Inserting this information in a chart to see the scaling better:

Or, more mathematical, correlations:

BLUE = HT Link voltage, altough the voltage measured with a DMM = Bios value +0.3v. Not sure if this is correct.
RED = NB Core voltage, not entirely sure, but I assume the core voltage is higher than the IGP voltage
YELLOW = IGP Voltage, seems pretty correct since when I set the voltage in the bios at 1.6v, I can hit 1.1GHz core and other people do the same frequency at around 1.38v Vigp.
PURPLE = ???, no idea what this does. It scales with the NB voltage setting in the bios, though.

AREA 2:

Red spot on picture underneath is Sideport memory readout (scaling with voltage options in the bios):

AREA 3:

This chip seems to be the one that's modified on HD3xxx and HD4xxx cards to increase Vgpu, so I added it. Not sure what's it doing on the board, though ...

DATASHEET SIDEPORT: http://www.qimonda.com/download.jsp?...rev064_www.pdf

(hoping Geoffrey sees this thread)

I just did, though too informational for me to quickly read through :p
Any requests?

Voltage modifications. Only post #17 is FYI :)

Area3 is Vmem, by the way :)

Go go for them golden cups !!!

What BIOS voltage do you increase to get higher VGA clocks?

Increasing the voltage of the NB core / Pci-e (which seems to be link, at least according to the bios).

Nu, Vigp != Vnb. 1.6v in de bios, is -wellicht- ongeveer 1.35v Vigp (andere mensen kunnen de Vigp wél appart instellen en kunnen 1.1GHz core doen op 1.38v.

And there are probable no datasheet available?

Not of the CZ-AD 71A IC :(

AS324M: http://www.alldatasheet.net/datashee...MI/AS324M.html

What volts do you read at the MOSFET legs in area 2 and 3?

None of the voltages seems to be related to NB/IGP/VDIMM/HTLINK

Quote:

Originally Posted by Massman (Post 238925)

So that's for onboard mem correct? DDR2?

Quote:

Originally Posted by Massman (Post 238925)

None of the voltages seems to be related to NB/IGP/VDIMM/HTLINK

What do you read on them small mosfet and coils. Should be Vdimm around that area. Set it high enough in BIOS so you're not getting confused with onboard mem voltage.

Quote:

Originally Posted by geoffrey (Post 238929)

So that's for onboard mem correct? DDR2?

Onboard memory is DDR3 and was set at 1.5v; increasing it to 1.7v doesn't change anything on the mosfet voltages

Quote:

Originally Posted by geoffrey (Post 238929)

What do you read on them small mosfet and coils. Should be Vdimm around that area. Set it high enough in BIOS so you're not getting confused with onboard mem voltage.

Can't get to the back of the motherboard since the northbridge cooler isn't fixed to the board. Two of the smaller mosfets read 2.12v (bios: vdimm set at 2.2v) and two measure a bit above 5v.

Quote:

Originally Posted by Massman (Post 238930)

Can't get to the back of the motherboard since the northbridge cooler isn't fixed to the board. Two of the smaller mosfets read 2.12v (bios: vdimm set at 2.2v) and two measure a bit above 5v.

Okay, then, DDR3, let's set it at 1,8V so there is nothing to get confused with. Then try to map the area with voltage readouts :)

Set DDR2 to 1.8v or? And map what area exactly?

For onboard memory, I already mapped voltage readout points:

They lead to the AS324M IC, which seems to be an amplifier. 8 inputs and 4 outputs:

http://pdf1.alldatasheet.net/datashe...MI/AS324M.html

1 - 4.28v
2 - 4.48v
3 - 2.44v
4 - 10.88v

Just slammed the Mach2 GT unit to the NB ... and booting at 1.3GHz core frequency :)

Beat them scores !!!

will not be for today

Just tested with the Mach2 again, there's a small gain, but not in this way it justifies firing up the SS for each bench session. As far as I can see, voltage level is much more important that temperature.

Back to the voltmodding table :)