the nv board design is such that vrm is expecting one 12v blob. the vrm therefore only sees 1 rail so it cant do anything to change the situation other than shut down. the asus 3 shunt solution can only monitor and notify you if there is a problem because the vrm still doesnt see the 3 rails after they merge back into 1 rail.
monitoring and alerting like the asus astral is the 2nd weakest protection as you have to install the alert software and be willing to stop your gaming as soon as it chimes. monitoring and shutdown before you even try to game is better as it actually tells you your cables arent plugged in correctly. the vrm balancing multiple "rails" is part of the monitoring and has some ability to deal/correct with the problem.People keep acting like balancing is a high priority, but all you can really depend on, is monitoring, alerting, and then shut down/go into limp mode.
If you have a situation like Der8auer had in his testing with 20 amps on one wire and 2 amps on another, you can't balance that.
monitoring and alerting like the asus astral is the 2nd weakest protection as you have to install the alert software and be willing to stop your gaming as soon as it chimes. monitoring and shutdown before you even try to game is better as it actually tells you your cables arent plugged in correctly.
the vrm balancing multiple "rails" is part of the monitoring and has some ability to deal/correct with the problem.
This was brought up and addressed earlier. The 40/50 series were a departure from Ampere which had 3 shunt resistors that capped load to 200w each. 40/50 series now have just one blob taking in 600w which can easily overload any of the 6 power leads to the GPU if improper contact.Agreed. Monitoring and shutdown/limp mode should be card level, not dependent on PC software.
Tell me how you balance when one cable has 20 amps and one has 2 amps?
This was brought up and addressed earlier. The 40/50 series were a departure from Ampere which had 3 shunt resistors that capped load to 200w each. 40/50 series now have just one blob taking in 600w which can easily overload any of the 6 power leads to the GPU if improper contact.
View attachment 121310
Are you sure on that? I believe Ampere split the 12 pin to three separate 12V regions to power different phases.This is still only a about monitoring. It does nothing to balance a 2 amp vs 20 amp situation (hint you can't balance that).
It feeds the groups of pins into different (groups of) power phases on the card. GPUs with multiple power connectors have been doing this for many years, in this instance you have 3 groups of 2 pins. The "balancing" comes from the lowered max amperage that can end up on a single pin in a worst case scenario, max load per pin goes down linearly with the number of groups.This is still only a about monitoring. It does nothing to balance a 2 amp vs 20 amp situation (hint you can't balance that).
It feeds the groups of pins into different (groups of) power phases on the card. GPUs with multiple power connectors have been doing this for many years, in this instance you have 3 groups of 2 pins. The "balancing" comes from the lowered max amperage that can end up on a single pin in a worst case scenario, max load per pin goes down linearly with the number of groups.
Not if you have 4 PCIe 8-pins. You're getting at most 12.5a per connector.Unless you get down to individual pins, you can still have something like the 2 amp and 20 amp situation that Der8auer had, even with groups of two. All you need is one bad pin/connector/wire and it will try to pull all the current through the other one of the pair, and at 600W there is a lot of current through each wire.
With a three-way split of power phases on groups of 2 pins you get only 17A max per pin in case the other one makes no contact at all. That's 5A less than we saw in the Der8auer video, which was not even the worst case possible for that setup.All you need is one bad pin/connector/wire and it will try to pull all the current through the other one of the pair, and at 600W there is a lot of current through each wire.
If you create 6 separate 12V regions on the GPU, they're all going to pull ~8A with a 5090 load. If one of the contacts has a high contact resistance there's no other path and it's still going to pull it's normal current or even slightly more due to the voltage drop across the contact, and it'll very quickly start to overheat.
200w big difference vs 600w. 2 wires handling 200w is still potentially less problematic than 6 wires handling 600w where any single wire can more easily melt when carrying the entire load due to poor contact on other 5 wires.Going in groups of two won't save you from this. What if you have two bad connections in group, and you still try to force the 200 watts through them. 200 watts through through a bad connection can still melt/catch fire.
200w big difference vs 600w. 2 wires handling 200w is still potentially less problematic than 6 wires handling 600w where any single wire can more easily melt when carrying the entire load due to poor contact on other 5 wires.
200w x3 worked well with Ampere and no reports of molten connectors.
Seeking the better choice is almost irrelevant when none is used with the current Nvidia designs. The basic and imperfect power "balancing" from past generations is still orders of magnitude safer than what we have today.That's better than attempting to push more current through bad connections through load balancing.
There's no reason you can't do both. Pretty much every card from the dawn of supplemental power has involved some form of per connector current balancing, and while I've melted my share of PCIe 6 and 8 pin connectors it was pretty much a non-issue overall. Similarly with Ampere and the 3x2 arrangement we didn't see the same issues with melting connectors even without monitoring.You shouldn't attempt to pull more current through any connection that has a current drop. This is why I've always been arguing for alert/shutdown over load balancing.
Because load balancing is likely to attempt to pull more current through the connections with the most resistance, if you attempt to balance them.
Going in groups of two won't save you from this. What if you have two bad connections in group, and you still try to force the 200 watts through them. 200 watts through through a bad connection can still melt/catch fire.
The only reasonable approach is alert and shutdown.
Good safety is usually obtained through layers. Good user experience also tends to follow this rule.There's no reason you can't do both.
In general I agree, though the 4090Ti had 2 12VHPWR/12v2x6 connectors which isn't as bad as the single-connector 4090 or 5090 cards (especially not as bad as the 5090FE).Excluding the 3090 Ti, all other implementations of the 12VHPWR from AMD or NVIDIA are not safe; and should not be recommended.
The lack of safety margin compared to traditional 8-pin should be alarming.
This is no longer an Nvidia only issue; with Sapphire showing off their stupidity too. I fear someone’s house is going to have to burn down for any meaningful change to occur.
Techpowerup sums it up nicely:In general I agree, though the 4090Ti had 2 12VHPWR/12v2x6 connectors which isn't as bad as the single-connector 4090 or 5090 cards (especially not as bad as the 5090FE).
The GeForce RTX 4090 Ti is a graphics card by NVIDIA, that was never released.
Power Pin Current Rating: (Excluding sideband contacts) 9.2 A per pin/position minimum with a limit of a 30 °C T-Rise above ambient temperature conditions at +12 V VDC with all twelve contacts energized.
Due to variations in contact resistance, an individual pin may see more than 9.2A of current depending on cable contact resistance nonuniformity, but the total current for the assembly shall not exceed 55A RMS in each direction. See section 9.3.1.1 for contact resistance variability measurements and requirements.
Contact resistance variability within a cable assembly causes current imbalance between contacts and may result in individual contacts and/or wires exceeding the current per pin specified in section 9.1. In addition, this current imbalance may be increased due to cable bending and/or side loading of cable assembly mated to the Add-in Card. The specific wire, connectors and manufacturing process used for a cable assembly must be designed to accept the current imbalance due to contact resistance variability and side loading. Side loading is defined here as a load applied in each direction defined in Figure 9-10 perpendicular to housing bodies.
To measure low-level contact resistance and qualify that a mated header and cable assembly design can control contact resistance variability under side loading conditions, the following methodology is provided. LLCR measurements shall be conducted according to EIA-364-23, Low Level Contact Resistance Test Procedure for Electrical Connectors and Sockets.
- Secure the Add-in Card connector to a fixture.
- LLCR of the cable assembly is measured from the footprint of the receptacle on the top of the Add-in Card PCB to 50mm from the point where the wire exits the body of the plug. The purpose of controlling the measurement point is to ensure that wire length and its contribution to contact resistance is repeatable. This does not imply a restriction on specific cable assembly implementation.
- Perform 30 mating cycles between the connector and the cable assembly.
- Record LLCR of each conductor of the cable assembly in the unloaded condition.
- Apply side load of 20N in each direction as defined in Figure 9-10. Load must be applied to the wire bundle and beyond any cable tie or strain relief feature of the assembly if present. Record LLCR of each conductor once the value has reached a stable value.
- Calculate average contact resistance of pin groups 1-6 and 7-12 independently for each side load condition.
- The LLCR shall not vary on any pin more than 50% from the average of that pins respective group in each test condition. A maximum LLCR of 6mohm/contact is required on each conductor.