GTX 1660Ti Compute Performance

[gordonbb]

Anand Tech has posted their review of the $279US EVGA GTX 1660 Ti. It only show Single-Precision results but shows the card at stock clocks producing in FAHBench about the same results as a GTX 1070 FE so likely 700-800kPPD.

The reviewed model is a triple (2.75) slot single fan “short board” model not a dual-slot dual-fan model and is also the “Black” model with a 1770MHz Boost clock rather than the XC Gaming model with a 1860MHz Boost clock which is likely of more interest to those folding with multiple air-cooled cards in one rig which should produce slightly better results.

[gordonbb]

Code: Select all

            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    850  130   3.05     6.54

TLDR - 2060 is likely a better buy if cost is not a factor

[MeeLee]

From what I read, it's closer to a regular 1070 in PPD (600-700k).
The main ingredient is faster and more efficient, lower power, GDDR6 ram, and a slightly higher GPU boost clock speed.

[toTOW]

gordonbb> You're way too optimistic ... I'm seeing only 620k on a 1070 and 700-800k on a 1080 ...

[gordonbb]

gordonbb> You're way too optimistic ... I'm seeing only 620k on a 1070 and 700-800k on a 1080 ...

The original value for the GTX 1660 Ti I agree was a little optimistic and would likely be closer to 750kPPD. The values are assuming a stable overclock and not a lower power limit for stability and longevity but the results should scale for comparison.

Stable Overclock

Code: Select all

Card        Cost Yield  TDP   Value Efficiency
             US$  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

I can't speak to the stock values as the actual data is no longer available at Overclock.net just the summary so it is hard to judge how skewed the reported values are. I am running a 2060 and 2070 currently and can vouch for those numbers.

But the conclusion, even if we further reduce the PPD estimate for the 1660 Ti, is still that the 2060 is a better value and more efficient if cost isn't too much of an issue. Given the process and architecture improvements in Turing the 1660 Ti would likely only make sense if cost is an issue and one is trying to decide between a GTX 1060 6GB or 1070 and a GTX 1660 Ti.

My EVGA 1070 Ti SC Black, averaged over 7 days, is running at 883.52kPPD, 1.96GHz Shader Clock (+150MHz Offset), 145.58W Actual Power Draw, (150W Power Limit), 63.71C GPU Temperature and 44.81% (1633.15rpm) Fan Speed and has completed 41 WUs.

I'm still relatively new at Folding and have a lot I can learn. I'm trying to maximize my Folding efficiency and to be rigorous in my methodology and appreciate any and all feedback that can help me along the way.

[Nathan_P]

toTow's numbers for a reference spec 1070 and 1080 are spot on, my heavily factory overclocked 1070 and 1080 score around 700k and 850k PPD respectively.

As with all F@H numbers though, PPD is heavily dependant on the project being worked on. My 1080 can vary from 680k PPD to 890k PPD depending on the project

[MeeLee]

Still, the 1660 isn't like the 1060 or 1070 or 1080.
It can be considered a lower end model of a new generation of cards.

It has slightly lower cuda cores, but higher GPU boost frequency from factory, and GDDR6 ram, which is a big improvement on GDDR5 ram.
As a result, lower overclock percentages can be gotten from these cards than a GTX1070 or 1080, but the 1660 should be considered a lower end RTX2060 card, without ray tracing, rather than a competitor to the 1060 or 1070.
It's a lot faster than a 1060.

[gordonbb]

Still, the 1660 isn't like the 1060 or 1070 or 1080.
It can be considered a lower end model of a new generation of cards.

It has slightly lower cuda cores, but higher GPU boost frequency from factory, and GDDR6 ram, which is a big improvement on GDDR5 ram.
As a result, lower overclock percentages can be gotten from these cards than a GTX1070 or 1080, but the 1660 should be considered a lower end RTX2060 card, without ray tracing, rather than a competitor to the 1060 or 1070.
It's a lot faster than a 1060.

Folding won’t generally see a significant advantage with GDDR6 versus 5 or 5X. Nor do the improvements in the Turing architechture with respect to dedicated integer cores be able to be leveraged in the current version of OpenMM in use but that may change.

It is, however, useful to compare the performance of the new Turing cards compared to Pascal as that is what most of us are familiar with and at this point both are available in the market so the information is useful for someone deciding whether to buy a1070 or a 1660 for folding.

I actually find my 2070 to be a strong overclocker. I can run +180MHz stable Folding at 2025MHz for weeks. I haven’t pushed the 2060 yet to see how far it goes yet but I’m mostly focused on efficiency which is where Turing really seems to shine with the process and architecture improvements.

[HaloJones]

I've read that Turing's compute pipeline is more efficient so whereas Pascal switches between integer and floating-point, Turing can run them simultaneously.

[toTOW]

Which is pretty useless since FAH is floating point operations only ...

[gordonbb]

toTow's numbers for a reference spec 1070 and 1080 are spot on, my heavily factory overclocked 1070 and 1080 score around 700k and 850k PPD respectively.

As with all F@H numbers though, PPD is heavily dependant on the project being worked on. My 1080 can vary from 680k PPD to 890k PPD depending on the project

I just started running a "Stock" Baseline for my cards. i.e. No overclocks, Default Power-Limit, Auto Fan. Granted I'm just a few WUs in but my EVGA GTX 1070 SC Gaming Black ACX 3.0 (08G-P4-5671-KR) is still yielding 895.14kPPD as the lower card in a dual-card Rig. It's not a higher-end FTW model so it's base clock is 1607 and Boost is 1683MHz but it still manages to Boost up to 1974-87MHz at 67C and 57% Fan so I'm not certain why your yield is so low with your 1080 if your running it Stock with no Power Limit Adjustments and your not Thermally limited by your chassis or Ambient Temperature. I'll have better numbers for yields at Stock for my hardware at the end of the week.

[Theodore]

Folding won’t generally see a significant advantage with GDDR6 versus 5 or 5X. Nor do the improvements in the Turing architechture with respect to dedicated integer cores be able to be leveraged in the current version of OpenMM in use but that may change.

I beg to differ on that.
Due to the way the pascal cards are made, not much overclocking can be done on the GPU, but quite an amount can be done on the Vram.
I think this is what most people have been doing with these cards to get higher PPD (when limited by the purchase of faster hardware).
The GDDR5 used in the 1060, 1070, and 1080 is limited to 8Gbps; or about 7,6GB/s.
It is probably very easy to overclock these cards to 8,6GB/s, and a noticeable performance improvement is detected while folding.
I think many overclocking articles in EVGA or overclock.net forums will concur.
The GDDR6 used in 1660 and RTX cards, is limited to 14Gbps or 13.6GB/s; which is nearly twice the speed.
Definitely a boost in performance for folding that far outpaces overclocking GDDR5 RAM on a now 'older gen card' (the 1060, 1070, and 1080 cards are now 1 generation old), and performance of faster RAM is definitely measurable in FAH.

The exact amount of benefit GDDR6 gives over GDDR5, is probably best tested by safely overclocking the GDDR5 RAM as high as possible; and/or underclocking the GDDR6 RAM to speeds equal that of a GDDR5 card.
It will probably give a closest estimate on the performance impact in FAH.

[Joe_H]

You can "beg to differ" all you want, but your belief runs counter to extensive testing by folders showing speed and amount of VRAM has a minimal effect on GPU folding speed. The GPU clock speed for the shaders is more important. In fact, such testing has shown that excessive VRAM speed leads to increased thermal load and can either bring on errors, or slow down the shaders.

Now if you want to do careful testing and post the results, others may be interested or be able to show problems with your test setup.

[gordonbb]

Folding won’t generally see a significant advantage with GDDR6 versus 5 or 5X. Nor do the improvements in the Turing architechture with respect to dedicated integer cores be able to be leveraged in the current version of OpenMM in use but that may change.

I beg to differ on that.
Due to the way the pascal cards are made, not much overclocking can be done on the GPU, but quite an amount can be done on the Vram.
I think this is what most people have been doing with these cards to get higher PPD (when limited by the purchase of faster hardware).
The GDDR5 used in the 1060, 1070, and 1080 is limited to 8Gbps; or about 7,6GB/s.
It is probably very easy to overclock these cards to 8,6GB/s, and a noticeable performance improvement is detected while folding.
I think many overclocking articles in EVGA or overclock.net forums will concur.
The GDDR6 used in 1660 and RTX cards, is limited to 14Gbps or 13.6GB/s; which is nearly twice the speed.
Definitely a boost in performance for folding that far outpaces overclocking GDDR5 RAM on a now 'older gen card' (the 1060, 1070, and 1080 cards are now 1 generation old), and performance of faster RAM is definitely measurable in FAH.

The exact amount of benefit GDDR6 gives over GDDR5, is probably best tested by safely overclocking the GDDR5 RAM as high as possible; and/or underclocking the GDDR6 RAM to speeds equal that of a GDDR5 card.
It will probably give a closest estimate on the performance impact in FAH.

Hmm - when I tested it I saw a 2% increase in PPD:

Code: Select all

Power	GPU	Mem	GPU	GPU	Cha	Fan	FAHBench	FAHBench 	FAHBench	Inc.	Total
Limit	Clk	Clk	Clk	 T	  T		Score		Scaled		Atoms		Change	Change	
 (W)	O/S	O/S	(MHz)	(°C)	(°C)	(%)	(ns/day)	(ns/day)			 (%)	 (%)
217	185	0	2025	66	44	70	118.7782	184.5906	35206		
217	185	50	2025	66	43	70	119.0690	185.0425	35206		0.24%	0.24%
217	185	100	2025	66	43	70	119.2506	185.3247	35206		0.15%	0.40%
217	185	150	2025	66	43	70	119.4067	185.5673	35206		0.13%	0.53%
217	185	200	2025	66	43	70	119.6039	185.8737	35206		0.17%	0.70%
217	185	250	2025	66	43	70	119.7430	186.0898	35206		0.12%	0.81%
217	185	300	2037	67	43	70	120.1965	186.7948	35206		0.38%	1.19%
217	185	350	2025	66	44	70	120.1370	186.7023	35206		-0.05%	1.14%
217	185	400	2037	67	44	70	120.4953	187.2591	35206		0.30%	1.45%
217	185	450	2025	67	44	70	120.5535	187.3495	35206		0.05%	1.49%
217	185	500	2037	67	43	70	120.7269	187.6189	35206		0.14%	1.64%
217	185	550	2025	67	44	70	120.7805	187.7023	35206		0.04%	1.69%
217	185	600	2025	67	43	70	120.8827	187.8611	35206		0.08%	1.77%
217	185	650	2025	67	43	70	121.0401	188.1056	35206		0.13%	1.90%
217	185	700	2025	67	43	70	121.1519	188.2795	35206		0.09%	2.00%

GTX 1070 Ti
Ambient: 24C
Stable GPU Temp reached by 60% of 5min run
nvidia-smi -i 1 -pm 1
nvidia-smi -i 0 --power-limit=217
DISPLAY=:0 XAUTHORITY=/run/user/122/gdm/Xauthority nvidia-settings -a [fan:0]/GPUTargetFanSpeed=70
DISPLAY=:0 XAUTHORITY=/run/user/122/gdm/Xauthority nvidia-settings -a [gpu:0]/GPUGraphicsClockOffset[3]=185
DISPLAY=:0 XAUTHORITY=/run/user/122/gdm/Xauthority nvidia-settings -a [gpu:0]/GPUMemoryTransferRateOffset[3]=0
nvidia-smi -i 0 -l 1 --format=csv,noheader --query-gpu=temperature.gpu,power.draw,clocks.current.sm,fan.speed
~/projects/FAHBench-2.3.2-Linux/bin# ./FAHBench-cmd -w wu-11713 --run-length 300

When I tried it in Production I saw no noticeable increase in PPD. When I stopped after a month I saw no noticeable decrease.
What increase are you seeing and what is your test methodology?

[MeeLee]

You can "beg to differ" all you want, but your belief runs counter to extensive testing by folders showing speed and amount of VRAM has a minimal effect on GPU folding speed. The GPU clock speed for the shaders is more important. In fact, such testing has shown that excessive VRAM speed leads to increased thermal load and can either bring on errors, or slow down the shaders.

Now if you want to do careful testing and post the results, others may be interested or be able to show problems with your test setup.

I don't think that information is correct.

I can concur with Theodore, overclocking just the VRAM by 700Mhz on a mid-tier card, increases PPD by about 10-15%, not the 2% mentioned above.
My stock 1060 runs at 335k PPD, but with a VRAM overclock it runs at about 375k PPD.

Likewise, when I would underclock the VRAM to the maximum underclock, the PPD count is nearly halved of fully working.
If it were the case that folding wouldn't benefit vram overclocking, everyone would underclock their VRAM to idle, and save a lot of energy in the process.

Vram overclocking definitely makes a difference on my cards!