Ampere binning variance

[Cowboystrekk]

After testing 5 Amperecards now I begin to see a bit of variance in binning between cards. I checked for lowest load volt without vram downclock and what MHz they could run at 750mv and 850mv + vram oc potential plus consumption in SOTTR at 750mv.

1. Palit 3060ti dual oc: min 731mv, 1600@750mv, 1850MHz@850mv, 15.2GHz vram, 132fps@120W vram at stock.
2. Asus 3060 tuf oc lhr: min 743mv, 1675@750mv, 1950MHz@850mv, 17. 6GHz vram, 108fps@95W vram at stock.
3. Msi 3060ti trio Z lhr: min 743mv, 1675@750mv, 1950@850mv, 15.6GHz vram, 136fps@125W vram at stock.
4. Msi 3060ti trio Z lhr: min 743mv, 1640@750mv, 1900@850mv, 15.8GHz vram, 136fps@124W vram at stock.
5. Asus 3060ti dual oc lhr: min 737mv, 1600@750mv, 1850@850mv, 15.9GHz vram, 132fps@119W vram at stock.

Seems like higher tier models have a bit better bin, except for the vram oc on the Asus dual. An vanilla 3060 also snuck in and impressed me both with binning on core and vram (it's specced for 16GHz).

Powerconsumption stock on trios is way worse than Palit and Asus dual, they use 235-240W avg with 3% higher fps, while Palit Asus uses 195-205W avg.

 

[W1zzard]

Is this actual GPU clock under load (like in GPU-Z sensors tab) or the rated boost clock target?

 

[basco]

interesting thanks!
could the 30 watt diff be the components used on the pcb?

 

[Cowboystrekk]

Is this actual GPU clock under load (like in GPU-Z sensors tab) or the rated boost clock target?

This is reading using afterburner + rtss overlay in games and I use gpu-z running in backround for checking max values after testing. When using the undervolt values I set a flat curve after the rarget freq/volt.

interesting thanks!
could the 30 watt diff be the components used on the pcb?

Msi trio z has 2x8-pin and higher powerlimit stock, Palit/Asus dual only has 1x8-pin. From the reviews here on TPU the trio also used more power than most other 3060ti's. The 5-10W higher consumption at same voltage might be explained by higher pcb-consumption, W1zzard knows much more about this than me

Asus dual stock:

Msi trio stock:

Trio runs about 100mv higher and 100MHz higher which equals 35-40W higher consumption and 3% better perf.

Cpu is running in 45W mode due to moneromining, hence the low cpu-performance and slight bottleneck.

 

[Zaqq]

interesting thanks!
could the 30 watt diff be the components used on the pcb?

isn't that because of different power limit settings? they all have 1.09-1.11fps/Watt with 750mv except for asus that has 1.136fps/W thanks to higher vram clock

 

[Cowboystrekk]

isn't that because of different power limit settings? they all have 1.09-1.11fps/Watt with 750mv except for asus that has 1.136fps/W thanks to higher vram clock

Forgot to mention that vram was not overclocked when testing SOTTR, I just tested to see how far I could OC vram in general Palitcard has a large rgb light which might explain the extra 1watt vs Asus

 

[joemama]

I thought everyone already knew that GPU manufacturers pick the better bin chips for higher tier cards so there is more room for overclock.
The power difference should be just the power limit settings in the vbios.

 

[Cowboystrekk]

I thought everyone already knew that GPU manufacturers pick the better bin chips for higher tier cards so there is more room for overclock.
The power difference should be just the power limit settings in the vbios.

Yeah, no surprise there, but I thought I would be interesting to check how big the difference is, and it looks like it is about 50-100MHz at same voltage.

The 1x8pin cards with max pwr-limit only reaches about 215W during spikes which makes sense since pcie+8pin=225W max. The 2x8pin trio cards could spike to above 250W using max pwrlimit.

 

[delshay]

The capacitors on each variant of the 3060ti makes a big difference. You could have a golden chip & don't even know it.

 

[nguyen]

After testing 5 Amperecards now I begin to see a bit of variance in binning between cards. I checked for lowest load volt without vram downclock and what MHz they could run at 750mv and 850mv + vram oc potential plus consumption in SOTTR at 750mv.

1. Palit 3060ti dual oc: min 731mv, 1600@750mv, 1850MHz@850mv, 15.2GHz vram, 132fps@120W vram at stock.
2. Asus 3060 tuf oc lhr: min 743mv, 1675@750mv, 1950MHz@850mv, 17. 6GHz vram, 108fps@95W vram at stock.
3. Msi 3060ti trio Z lhr: min 743mv, 1675@750mv, 1950@850mv, 15.6GHz vram, 136fps@125W vram at stock.
4. Msi 3060ti trio Z lhr: min 743mv, 1640@750mv, 1900@850mv, 15.8GHz vram, 136fps@124W vram at stock.
5. Asus 3060ti dual oc lhr: min 737mv, 1600@750mv, 1850@850mv, 15.9GHz vram, 132fps@119W vram at stock.

Seems like higher tier models have a bit better bin, except for the vram oc on the Asus dual. An vanilla 3060 also snuck in and impressed me both with binning on core and vram (it's specced for 16GHz).

Powerconsumption stock on trios is way worse than Palit and Asus dual, they use 235-240W avg with 3% higher fps, while Palit Asus uses 195-205W avg.

So it seems GA104 + GA106 work better at higher voltage than GA102

My Asus TUF 3090 OC sample is tested to be stable at 1725mhz@750mV (+300mhz core) and 1935mhz@850mV (+240mhz core), minimum voltage 737mV

 

[Cowboystrekk]

So it seems GA104 + GA106 work better at higher voltage than GA102

My Asus TUF 3090 OC sample is tested to be stable at 1725mhz@750mV (+300mhz core) and 1935mhz@850mV (+240mhz core), minimum voltage 737mV

Sounds like a golden sample I would also guess that the highest tier full dies like 3090 and 3070ti would be better binned than 3080 and 3060ti

 

[Ferd]

Well at least they’re not charging more for no reason

 

[Cowboystrekk]

Well at least they’re not charging more for no reason

It`s good to know that you don`t just pay for a better cooler\heatsink A few percent extra OC-potential is always handy on top tier models!

 

[nguyen]

It`s good to know that you don`t just pay for a better cooler\heatsink A few percent extra OC-potential is always handy on top tier models!

Yeah my 3090 sample can reach 2205mhz until power/thermal throttle kick in, probably not possible for other Ampere GPU.
Though for 4K gaming or Timespy, the stock 375W TDP is enough for 1950-2010mhz, with 2205mhz requires 500W+ power draw

 

[Cowboystrekk]

Got a 3060ti tuf lhr and this seems like the binning champ. I can do 1650@731mv and 1700@750mv.

Comparison in mining ravencoin, both 15.6GHz vram OC:
Asus 3060ti dual 1560@737mv 56C 65% fan, 155W
Asus 3060ti tuf 1650@731mv 55C 65% fan, 166W

The extra fan and maybe a bit more complex pcb seems to use a bit more power. Consumption of tuf is very similar to MSI trio, but binning on tuf is better and powerconsumption stock is 35-40W lower for 1% performance, trio is really thirsty stock, but okay when tamed (undervolted ).

 

[Vayra86]

I thought everyone already knew that GPU manufacturers pick the better bin chips for higher tier cards so there is more room for overclock.
The power difference should be just the power limit settings in the vbios.

This is true, but different generations (and especially nodes) have a different spread of bins. Pascal was very tight for example. You could get north of 2000 mhz on almost anything and it was almost always temp limited before voltage under air, lower voltage could sometimes even get you more clock stability.

 

[chrcoluk]

There is heavy variance on ampere chip quality even within same sku's on FE cards.

I can achieve 1905mhz at 882mv, but others can do easily at 850mv. (3080 FE)

We need more focus on undervolting as thats the thing on these new cards.

The stock voltages are so high wasting lots of power.