Vcore and vdimm with A64

[mike]

Sticky Please!!

[Kameleon]

Sticky Please!!

This thread has been a sticky since it was posted

[101]

I thought in OPB's case he didn't even get into the BIOS yet. If that's the case he lost his CPUs at stock voltages. What are the implications of the RAM being tied to the 5V rail for the jumperless Vdimm on the new DFI NF4? Isn't their crossfire also configured this way?

Perhaps on boot there are some floating bits that can cause the Vdimm to spike before the bios is fully loaded into RAM creating a huge Vtt/Vcore delta. Thus breaking down any clamping diodes in the CPU with Vcore=Vtt=~2.5V? Sounds like that would let the genie out of the die. Also if it's a control issue that occurs before the bios loads, there's not much chance in hell of it being fixed with an update.

All just speculation on my part obviously, but sounds feasible.

[Hybrid]

Ok let me see if i got this right, please correct me if im wrong. The diodes are acting as a "surge" protector. If the potential across the diodes get too high, they will break down and let a lot of current through which kills the CPU. So to prevent this, you would have to increase the Vcore in order to lower the potential across the diodes. If this is right, it doesnt seem like a bios update can fix it, unless the bios update limits the voltage range. Could this have been prevented if DFI used more diodes?

[101]

The protection diodes are on the die

[high5]

What are the implications of the RAM being tied to the 5V rail for the jumperless Vdimm on the new DFI NF4? Isn't their crossfire also configured this way?

vdimm regulation is completely different on new boards. new ones (expert & crossfire) have switching regulation, while old ones have plain linear.

[mascaras]

i have vcore=1.45v and VTT= 1.60 (vdim 3.20v) , so i think im ok .

[Flib]

Ok, I think I got it...
You have to keep the difference between vcore und vDimm as low as possible.
I want to run 3,5V or 3,6V through my BH-5 for 24/7.
But which vCore would be safe? Would 1,5V be enough? Also for 3,6vDimm?
I am on watercooling.

This is very interesting but without an official statement it is difficult to say which vDimm requires which vCore. Isn't it?

[high5]

for starters, i wouldn't consider 3.5-3.6v to be safe 24/7 vdimm.

[ccokeman]

This is all pretty interesting and will make me sit on the sidelines and watch to see what happens.

[ozzimark]

First of all, the CPU doesn't have to do directly with the VDD, but with Vtt/Vref, wich should vary from 1.25 to 2V when you adjust the RAM from 2.5 to 4V. Maximum JEDEC Vdd voltage is 2.85V, and lowest voltage for A64 CPU's (CnQ activated) is 1.1V, so we know for sure that a difference of 0.325V between Vcore and Vtt is safe. This means that for a 1.35V Vcore voltage, we can have a safe Vdd of at least 3.35V, and for 1.5V Vcore -> 3.65V Vdd.
Probably the real danger is when those diodes go near the opening (conduction) voltage, wich should be ~0.6V for Si-based junction; if this is true, 1.35V Vcore and above 3.9V Vdd should kill the CPU instantly ... anyone wanna try ?

that's good logic.. however, that would imply that 3.45v is fine with 1.4v vcore (3.45/2 = 1.725 - 1.4v = .325)

remember that power draw increases with the square of the voltage

[high5]

remember that power draw increases with the square of the voltage

in theory yes. real world - not quite.

[ozzimark]

well yes, esp since silicon has an interesting habit of decreasing resistance, thus increasing current as temps go up. instead of going into the complexities of it all, i think making a rough guide by leaving it simple would work better


however, with that in mind, it would make sense that ram like micron, that runs hot because it draws more power, would also be a higher strain on those diodes at a given voltage, no? and that a single sided stick won't be nearly as bad as a double sided stick, and that 4 dimms will be worse than 2. there's just a hell of a lot of variables in this.

[uOpt]

Based on Micutzu's conclusion that 0.325 V differential should be safe (which I think makes a lot of sense), here is a small table listing what Vcore should be safe for which Vdimm:

ETA: this is at default drive strength. Higher drive strength will require less differential for safety. (Thanks, Ryder).

Vdimm - Vcpu(min)
----------------------------
2.9 V --- 1.12 V (min)
3.0 V --- 1.17 V (min)
3.1 V --- 1.22 V (min)
3.2 V --- 1.28 V (min)
3.3 V --- 1.33 V (min)
3.4 V --- 1.37 V (min)
3.5 V --- 1.42 V (min)
3.6 V --- 1.47 V (min)
3.7 V --- 1.53 V (min)
3.8 V --- 1.58 V (min)
3.9 V --- 1.62 V (min)
4.0 V --- 1.67 V (min)

Absolutely NO WARRANTY for this. This is a knee-jerk approach based on loose findings we have. Use at your own risk.

And again, it is minimum safe CPU voltage based purely on the problem of voltage differential between CPU and RAM. The above table does not take any other voltage considerations into account. Or in other words: the voltage that is the minimum with regards to the Vdimm differential might already be over the upper limit of your CPU - and blow it up.

[RyderOCZ]

Keep in mind also that Drive Strength on a board is associated with varying increases or decreases in resistance to the ram from the memory contoller.

Lower resistance could result in that safe Delta being lower than we think.

[wittekakker]

Probably the real danger is when those diodes go near the opening (conduction) voltage, wich should be ~0.6V for Si-based junction; if this is true, 1.35V Vcore and above 3.9V Vdd should kill the CPU instantly ... anyone wanna try ?

Some more details on how this works would be nice, I like to understand it 100%.
It makes much sense

[Ma7z0]

Sticazzi!

Anyway I run stable 24/7 for 2/3 months UTT@3.22v with C&Q enabled. That means going down with vcore to 1.088v (CPU-Z) with VRAM 3.22v.
I don't have any issue.

[uOpt]

if the 0.6 V differential is safe, this would be the safe values:

Vdimm - Vcpu
2.9 V --- 0.85 V (min)
3.0 V --- 0.90 V (min)
3.1 V --- 0.95 V (min)
3.2 V --- 1.00 V (min)
3.3 V --- 1.05 V (min)
3.4 V --- 1.10 V (min)
3.5 V --- 1.15 V (min)
3.6 V --- 1.20 V (min)
3.7 V --- 1.25 V (min)
3.8 V --- 1.30 V (min)
3.9 V --- 1.35 V (min)
4.0 V --- 1.40 V (min)

[Micutzu]

I didn't say the 0.6V margin is safe, that was just a shot in the dark as i really don't know what's going on in there; that was just the first assumption i could make given the known facts.
Also, we should take into consideration the signal that travels trough those buses, that is spec'ed at +/-0.76V for Winbond BH/CH chips with default Vdd/Vtt.

[uOpt]

However, AMD seems to say 1.45 Vcore is safe.

If they had 4.0 Volts Vdimm in mind, then that would fit the second table.

We need some volunteer to blow up his CPU. Why don't we collect some money and execute the most disappointing Venice we have?

[STEvil]

if you read the amd white papers you will notice 3.0v is the maximum rated memory voltage, or thats what it was last time I read them (s939 wasnt out yet).

I would assume keeping a delta of no more than .3v is a good idea assuming stock voltage (1.35?) vs max selectable memory voltage in most boards (2.85-2.9v) is a good guide. Of course thats just speculation and could change as drive strength settings are changed as mentioned earlier...

[fareastgq]

Good thing I always ran 1.65-1.7 volts through my chips, hahaha. But really, that would explain why I have never fried a chip myself. Plenty of ram, but zero chips. Thks for the info Tony.

[uOpt]

Folks, I tried to summarize this issue here and plan to spread it a little more:
http://cracauer-forum.cons.org/forum...ic.php?p=83#83

It would be great if you could have a look to see whether
- my summary is accurate
- whether I give proper credits to the people involved

[OV3RCLOCK3R]

Do the new opterons suffer from this problem?

I'm running Ocz VX atm at 3.4v

[Fr3ak]

Does damaging the diodes mean that the CPU dies instalntly or does it still run, but very unstable?