Is The GTX 1660 Ti The New Folding Best Value?

[Gorgon]

Nvidia launched their newest GPU today the GTX 1660 Ti. AnandTech in their review tested the Single-Precision FAHBench on a EVGA's lowest end model the 1660 Ti XC Black Gaming and it produces about the same as the 1070FE so about 700-800kPPD.

 

The model tested is a 2.75-slot "short board" not the likely better performing XC Ultra Gaming that features dual fans and a 2 slot width more compatible with multiple GPU setups on most z370/390 and x-series boards.

 

So at MSRP of 279US$ for the XC Black and $309US$ for the Ultra models do these represent the best bang for the buck for folding these days given the Ray Tracing Tax built into the RTX models?

 

Looking at Value (kPPD/$) and Efficiency (kPPD/W) we see:

            Cost Yield  TDP   Value Efficiency
Model       MSRP  kPPD   W   kPPD/$   kPPD/W
RTX 2080 Ti 999   2400  260   2.40     9.23
RTX 2080    699   1400  225   2.00     6.22
RTX 2070    499   1275  185   2.56     6.89
RTX 2060    349   1050  160   3.01     6.56
GTX 1660 Ti 279    750  130   2.69     5.77

So the 1660 Ti does not have a value advantage over the 2060 and is less efficienct based on estimated performance. If real world testing confirms the yield at 700-800kPPD then the 2060 is still the better buy if cost is not a limiting factor.

[mattheginger]

Correct me if I'm wrong, but the 2060 averages around 800k ppd also right? So on a dollar or efficiency basis, the 1660 Ti would have to produce under 670-680ish k ppd to be worse value.

[mattheginger]

On a wattage basis 130/160 = 0.813

 

0.813 x 800,000 ppd (average performance of 2060) = 650,000 ppd. 

 

Of course you also would need to take into account the power consumption of the system running the card, and when you add that to the kw/ppd then I guess it may tip towards the 2060's favour?

[Gorgon]

On 2/26/2019 at 10:26 PM, mattheginger said:

Correct me if I'm wrong, but the 2060 averages around 800k ppd also right? So on a dollar or efficiency basis, the 1660 Ti would have to produce under 670-680ish k ppd to be worse value.

Stock settings (clocks, powerlimits, fan) under Ubuntu 18.04.1 with ppa 410 Driver in a PCIe3 x16 running at x8

I’d say over 1MPPD 1.1 with some manual adjustments.  Haven’t clocked this bad boy past a +75MHz graphics offset yet but if it’s anything like my 2070 (and this is the EVGA XC Ultra whereas my 2070 is just an XC) it should be stable at +175.

Edited April 30, 2019 by Gorgon
Inserting Graphs

[Gorgon]

5 minutes ago, mattheginger said:

On a wattage basis 130/160 = 0.813

 

0.813 x 800,000 ppd (average performance of 2060) = 650,000 ppd. 

 

Of course you also would need to take into account the power consumption of the system running the card, and when you add that to the kw/ppd then I guess it may tip towards the 2060's favour?

Don’t forget the Quick Return Bonus which applies an inverse square root bump in points for how much faster than the time limit you finish the WU. Faster cards get a disproportionately bigger bonus.

[ProDigit]

On 2/26/2019 at 10:26 PM, mattheginger said:

Correct me if I'm wrong, but the 2060 averages around 800k ppd also right? So on a dollar or efficiency basis, the 1660 Ti would have to produce under 670-680ish k ppd to be worse value.

I'll correct you, and no. My 2060 averages around 1040kPPD, and in some cases reaches as high as 1.3 to 1.4MPPD.

 

But I'll tell you where the 2060 makes 870k PPD.

When you limit it to run at 125Watts in Linux.

Most of the time the NVidia drivers will overclock the card higher, so you'll average more like at 970kPPDs.

 

But if you set the 2070 to 125Watt, it reaches to 1.1MPPD. and the 2080 1.4 MPPD (at the same 125Watt).

Now that's efficient! Especially when you pay your own power bills...

[Gorgon]

Initial testing of a EVGA 1660Ti XC Ultra (the 2-slot, dual fan card) stock under Linux has shown:

So 731.3 kPPD. This is "stock" settings as the lower card in a dual-card rig on a PCIe3 x16 slot at x8 at a manual fan of 60% to keep it from cooking the card above it too much and it was running at 53C and an average of 109.43W with the default power limit of 160W.

 

I have just figured out how to set the Graphics Clock Offset for Turing and have set a initial offset of 3 15MHz "bins" (+45MHz) and will report back on stability and any change in performance.

[Gorgon]

On 4/28/2019 at 1:48 PM, Gorgon said:

Initial testing of a EVGA 1660Ti XC Ultra (the 2-slot, dual fan card) stock under Linux has shown:

So 731.3 kPPD. This is "stock" settings as the lower card in a dual-card rig on a PCIe3 x16 slot at x8 at a manual fan of 60% to keep it from cooking the card above it too much and it was running at 53C and an average of 109.43W with the default power limit of 160W.

 

I have just figured out how to set the Graphics Clock Offset for Turing and have set a initial offset of 3 15MHz "bins" (+45MHz) and will report back on stability and any change in performance.

Did some profiling with FAHBench. +150 was unstable so running at +135 for now. 751kPPD so far

[Gorgon]

Seeing an average of 778.9 kPPD now on the 1660 Ti. It is running at a cool 55C as the lower card in a dual-slot system at 60% Fan, 115-125W Power.

I had one WU fail so I've bumped the GPU O/C now down to +105MHz (5 15MHz Turing "bins") yielding a 2055MHz Graphics Clock.

 

I'll let it run for a couple more days to finish the "Profiling" verifying a stable overclock and maximum power draw. The next step will be to test the card's yield at 90% of the FE Power Target (108/120W) which I've found to be a sweet spot for production and efficiency.

 

Though Turing and Pascal cards appear to have a linear increase in efficiency as the Power Limits are lowered there is a relatively fixed "System Overhead" from the CPU, VRM, Chipset, Chassis Fans, Power Supply efficiencies etc. that results in better Overall "System Efficiency" (PPD/Watts at the Wall) at higher Power Targets.

[ProDigit]

What are the settings in Linux for minimum power draw on the 1660 ti?

For RTX cards, the minimum is 125W, but for optimal settings,
The 2060 needs to run at about 128-129W, as long as the card can stay around 60C.
The 2070 needs to run at 130-133W,

The 2080 needs to run at 136-139W,

The 2080 ti needs to run at ~170-178W.

 

Lowering the power curve, allows for slightly higher overclocks.

 

The procedure for overclocking RTX cards (and GTX cards) is simple:
1- Start the card cold (sub 50C).
2- Increase power limit to maximum (usually a few watts above the factory settings)

3- Check how high initial boost frequency goes to. Write that number down. It's your OC aim.

4- Lower power limit to the lowest setting.
5- Try to overclock until you're close to stock boost clock rates.
6- If WU fails, lower; if WU seems stable, dial back by 5Mhz and save.

7- Over time, GPUs will wear out, and overclocking margins will lower.

 

The 1660ti at this moment isn't the best buy yet, unless if they lower the price by a lot.

If the 1660 ti gets 750k PPD, and the 2060 gets 1M PPD, you'll need 4x 1660 tis to match the performance of 3x RTX 2060s.
The cost price of 4x 1660 tis, is about $30 higher than 3x 2060s.

 

Aside from cost price, the 2060s run best at 129 Watts. x3 = 387 W.
Just for reference, I'll presume that the 1660 ti's run best at 117 Watts. x4 = 468W.
For roughly the same performance.

 

2060s are still the sweet spot. Especially if you can get second hand ones.

 

[ProDigit]

PS, if you're interested in the 2060 vs 2080:

2060 = 1M PPD for $350, x3 = 3M PPD for $1050

2080 = 1.5M PPD for $700 x2 = 3M PPD for $1400

 

2080 costs $350 more initially.

 

Electric savings:

3x 2060 @ 129W = 387W

2x 2080 @ 136W = 272W

=============== 115W

 

115W savings when folding 24/7 = $115/Yr on savings.

 

$350 initial cost difference divided by the $115 annual electricity savings = 3 years.

 

2080 GPUs will start to benefit when folding 24/7 for more than 3 years.

That number is 75 years if you compare the RTX 2060 with the RTX 2080 ti.