Nvidia Ampere Discussion [2020-05-14]

[Man from Atlantis]

2080Ti Power limit@109%

3080

 

[trinibwoy]

How are we supposed to interpret those numbers? Chip power draw on the 3080 is lower when overclocked? That can’t be right.

 

[Benetanegia]

How are we supposed to interpret those numbers? Chip power draw on the 3080 is lower when overclocked? That can’t be right.

From source:

The GPU was only overclocked by 70 MHz and memory by 850 MHz.

Big memory OC paired with little GPU OC on a total board power seemingly limited to 317w in both cases. Makes sense to me, that memory is taking from GPU power in this case.

 

[techuse]

Where is that 134 watts on top of the GPU core going? Seems like an awful lot of power. Almost triple the 2080ti.

 

[pharma]

Ampere Details: Error Detection and Replay Simplifies GDDR6X Overclocking
September 13, 2020

With the GDDR6(X) storage controller, NVIDIA is introducing a new technology called Error Detection and Replay (EDR). This plays an important role, especially in the interaction with overclocking.

The GDDR6X memory works on the GeForce RTX 3080 Founders Edition with a clock of 1,188 MHz. It is not yet known how far it can be tacted. However, Error Detection and Replay is intended to make overclocking the memory easier, since errors in the transfer of memory are detected (error detection) and the data is transmitted until it arrives (replay). Instead of displaying artifacts, the transfer errors are detected and the storage controller tries to compensate for them. It is a Cyclic Redundancy Check (CRC) and therefore a method that works with a test value for the data. If the check value is not correct, the transmission failed.
...
The GeForce RTX 3080 and GeForce RTX 3090 uses GDDR6X storage. The GeForce RTX 3070, on the other hand, does not. This is where the GA104 GPU is used. It is currently unknown whether the GeForce RTX 3070 will also support Error Detection and Replay.

https://www.hardwareluxx.de/index.p...d-replay-vereinfacht-gddr6x-overclocking.html

 

[trinibwoy]

Big memory OC paired with little GPU OC on a total board power seemingly limited to 317w in both cases. Makes sense to me, that memory is taking from GPU power in this case.

Yeah but avg GPU power is lower while clocks are higher In the second pic. Something is off there.

 

[Benetanegia]

Yeah but avg GPU power is lower while clocks are higher In the second pic. Something is off there.

It is lower, likely because the memory is taking up more and the board is limited. We don't know GPU load of each case, and I don't even mean the 0-100% load metric that GPUz typically shows (tho missing there), I mean actual GPU load across all of its units.

 

[Scott_Arm]

It is lower, likely because the memory is taking up more and the board is limited. We don't know GPU load of each case, and I don't even mean the 0-100% load metric that GPUz typically shows (tho missing there), I mean actual GPU load across all of its units.

now I’m even more curious about undervolting.

 

[kalelovil]

Ampere Details: Error Detection and Replay Simplifies GDDR6X Overclocking
September 13, 2020

https://www.hardwareluxx.de/index.p...d-replay-vereinfacht-gddr6x-overclocking.html

Any idea on how this differs from the error detection and replay present since the HD 5000 series?
https://www.anandtech.com/show/2841/12

 

[pcchen]

Any idea on how this differs from the error detection and replay present since the HD 5000 series?
https://www.anandtech.com/show/2841/12

GDDR6's error correction feature is very similar to GDDR5, only because it has twice the burst data length, so the checksum is now 8+8 bits instead of 8 bits. The CRC algorithm used is also the same. I don't know about GDDR6X but I guess it's probably similar.

Note that this feature does not provide detection for on-memory error, but only bus error. Therefore it's probably more important on GDDR6X as it uses PAM4 which can be more error prone.

 

[agent_x007]

3080

So : 0,975V @ 1873MHz by default, and 0,95V @ 1935MHz after slight undervolting...
Question is : What die quality is this ?
I think 2080 Ti (FE) had similar quirk (ie. need of undervolt, to get higher clocks useful/not power walled) ?
Still, good to know it has some headroom for undervolting + OC (70C at 320W is good, but I guess fan speed isn't supported yet in GPU-z... so no info on how loud it was).

 

[CarstenS]

2080Ti Power limit@109%
3080
https://cdn.videocardz.com/1/2020/09/NVIDIA-GeForce-RTX-3080-Memory-OC-Test.png

183,3+75,5+85,9 = 344,7
175,1+76,1+87,9 = 339,1

That's at least what MS's Win10 calculator tells me. So I'm guessing those numbers in that screenshot are not the whole truth, some of GPU, MVDCC and PWR_SRC maybe factoring into one another. The power from the individual rails does line up much better though.

 

[dorf]

So : 0,975V @ 1873MHz by default, and 0,95V @ 1935MHz after slight undervolting...

0.9514V is an average voltage from a longer period of monitoring (small green text says avg). 0.975V is a single reading (missing the green text). Not comparable.

 

[Scott_Arm]

Do nvidia GPUs lower clocks when they hit power limit? I’ve never actually had a gpu hit power limit before.

 

[dorf]

Do nvidia GPUs lower clocks when they hit power limit? I’ve never actually had a gpu hit power limit before.

Yes, they move to a lower point in the voltage frequency curve.

 

[LordEC911]

Strike that. My bad

 

[Scott_Arm]

Yes, they move to a lower point in the voltage frequency curve.

Figured. Seemed the most obvious.

It looks like both of those GPU-Z images are overclocked, maybe. I don't really know how the boosting behaviour is supposed to work, but I've heard some people say Nvidia gpus will boost the clock above the listed boost clock. The mem clocks in both of these images are different, so I'm not sure if these are AIBs with factory overclocks, or if it's user overclocks. Really curious to know if they'd be power limited out of the box.

 

[DegustatoR]

I don't really know how the boosting behaviour is supposed to work, but I've heard some people say Nvidia gpus will boost the clock above the listed boost clock.

All NV GPUs do this since Maxwell. Listed boost clocks can be seen mostly on low end models with insufficient cooling, the rest are running at +100+200 MHz to that as a norm.

 

[Dictator]

All NV GPUs do this since Maxwell. Listed boost clocks can be seen mostly on low end models with insufficient cooling, the rest are running at +100+200 MHz to that as a norm.

Yeah boost of a standard RTX 2080 Ti as listed by NV is like 1545mhz, when even their stock model under extreme load will be in the 1800s. Other partner cards easily in 1990-2000+ range.

 

[Man from Atlantis]

RTX 3080 Founders Edition FLIR capture

https://streamable.com/o7l5ta