Nvidia Pascal and AMD Polaris, starting with GTX 1080/1070, and the AMD 480

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168389195
RAC: 0

I have some power stats now

I have some power stats now on a system that had both an RX 480 and GTX 1070 in it, and will have further info soon on a system that has had a 7970, RX 480 and 1070. All tests were BRP6, no other tasks running. Power draw at wall with Kill-a-watt.

- RX 480 was reference XFX model with small factory overclock, tested before the power draw fix. Rated to 150w, with one 6pin connector. I returned this card and cannot do further testing.

- GTX 1070 is EVGA Gaming FTW model, overclocked 95MHz Core, 795MHz memory. Card is a higher-end model rated to max 215w, with 2 8-pin connectors (really).

- 7970 is Gigabyte Windforce OC with 3 fans, small UNDERclock. Rated to 250w, 1x 8-pin and 1x 6-pin.

--------------------------------------

SYSTEM 1: Win 10 Pro, 2x E5-2670 V1 (95w) CPU. 3x 140mm, 4x 80mm fans including CPU fans. Platinum PSU.

IDLE:

7970: can test on request.
RX 480: 130w
GTX 1070: 170w

CRUNCHING:

7970: can test on request.
RX 480 3x: 315w
GTX 1070 4x: 270w

--------------------------------------

SYSTEM 2: Win 8.1 Standard, i5-3570 (77w), 80mm fan on CPU. Bronze (I think) PSU.

IDLE:

7970: 80w
RX 480: 65w
1070: TBD

CRUNCHING:

7970 3x: 260w
RX 480 3x: 245w
1070 4x: TBD

I'll test the 1070 in the i5 box this weekend, and will get power draw for no external GPU on the 2670 system when I pull the card.

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168389195
RAC: 0

Some learning about the

Some learning about the 1070/1080 version of PrecisionX:

- Basic overclocks do not seem to apply to P0 state; when the system is doing nothing, it stays at low clock values. Only when I start E@H does it go up to the specified OC values.

- However, if you enable KBOOST (K-shaped button on lower left), it forces the card to run the overclocks full-time. In this case, the memory overclock is an offset of the base clock, so 4000+OC reported on the 1070. When you start crunching, however, the memory OC becomes an offset of 3800.

- When KBOOST is enabled, HWMonitor and GPU-Z display 0 for most GPU sensor values.

- On my i5, crunching BRP6 with KBOOST enabled, or crunching without KBOOST and starting a 3D game, results in a large voltage, clock and memory value drop as reported by Precision-X. I don't remember this happening on the 2670, and am going to try a different power supply in the i5 this weekend.

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7225718263
RAC: 1050055

RE: - Basic overclocks do

Quote:
- Basic overclocks do not seem to apply to P0 state; when the system is doing nothing, it stays at low clock values.


In Nvidia Inspector representation, the 1070 rests in P8 state when very little is going on. That has very low (on the order of one tenth speed) clock rates for both memory and core. P0 seems for the 1070 likely to be the standard state reported during game-playing, though as I don't play games this is anecdotal for me. I'm not sure, but suspect the system may pass through a transient P0 state on the way to the P2 state it normally occupies during crunching. That is troublesome for the P0 clock offset method I have been using in Nvidia Inspector, as it may mean memory clock is higher than I want during those transients.

Gamboleer
Gamboleer
Joined: 5 Dec 10
Posts: 173
Credit: 168389195
RAC: 0

RE: RE: - Basic

Quote:
Quote:
- Basic overclocks do not seem to apply to P0 state; when the system is doing nothing, it stays at low clock values.

In Nvidia Inspector representation, the 1070 rests in P8 state when very little is going on. That has very low (on the order of one tenth speed) clock rates for both memory and core. P0 seems for the 1070 likely to be the standard state reported during game-playing, though as I don't play games this is anecdotal for me. I'm not sure, but suspect the system may pass through a transient P0 state on the way to the P2 state it normally occupies during crunching. That is troublesome for the P0 clock offset method I have been using in Nvidia Inspector, as it may mean memory clock is higher than I want during those transients.

My typo. P8, not P0.

ExtraTerrestrial Apes
ExtraTerrestria...
Joined: 10 Nov 04
Posts: 770
Credit: 578596871
RAC: 199189

Why would you think there's a

Why would you think there's a transient state? This P8 - P2 - P0 power level scheme has been like this since quite some time. And nVidia already has many operation states available in P2 and P0 by choosing different boost values, so I don't see any benefit of some special transition sate (apart from the fact that any clock & voltage adjustments always involve some transistion at the physical level).

MrS

Scanning for our furry friends since Jan 2002

ravenigma
ravenigma
Joined: 20 Aug 10
Posts: 69
Credit: 80558946
RAC: 199

RE: Why would you think

Quote:
Why would you think there's a transient state?

Without much knowledge on the subject, I can testify that I've seen clock speeds boost way up (equivalent to P0 speeds) before settling down to P2 speeds when crunching.

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7225718263
RAC: 1050055

The release date/time for GTX

The release date/time for GTX 1060 reviews was less than an hour ago. Also product orders have started to be accepted.

Some differences are obvious compared to the 1080 and 1070 releases.

1. Review NDA release this time is the same day as initial order acceptance, rather than a week or so earlier.
2. This time the Founders Edition is only offered by Nvidia, not the other brands.
3. There is at least some availability at the stated MSRP.
4. Nowinstock seems so far to be totally useless for this product, while it was active from the first moments for at least four major sites for the previous two.

I'm in a hurry to see how this works, and am pleasantly impressed with the character of the fan noise on my Founders Edition 1070, so against all prudent advice I've ordered a GTX 1060 FE this morning.

mmonnin
mmonnin
Joined: 29 May 16
Posts: 291
Credit: 3412016540
RAC: 3441367

PNY is actually setting 1060s

PNY is actually setting 1060s at MSRP out of the gate @$250.

Jim1348
Jim1348
Joined: 19 Jan 06
Posts: 463
Credit: 257957147
RAC: 0

RE: I currently predict the

Quote:
I currently predict the GTX 1060 to be slower than the RX 480 on Einstein BRP6/CUDA55, perhaps rather substantially, but to be rather a low lower in system level power consumption and system resource requirement (with the current applications and drivers). I don't have an RX 480 on order now, so probably will not provide a direct comparison on the same host for that, but do plan to swap some 1060 (maybe even the one I just pre-ordered, if that comes through) into the host I used to provide 1070 Einstein data.


I am sure you have a basis for your statement, but the GTX 1060 is said to be equivalent to a GTX 980 in most respects, and I doubt that the RX 480 is better. Let us know what measurements you do get though, and thanks for the tests.

archae86
archae86
Joined: 6 Dec 05
Posts: 3157
Credit: 7225718263
RAC: 1050055

RE: GTX 1060 is said to be

Quote:
GTX 1060 is said to be equivalent to a GTX 980 in most respects, and I doubt that the RX 480 is better.


I think current results here at Einstein generally have AMD cards getting a comparative advantage vs. Nvidia cards if you take the Einstein performance ratio compared to the generally observed game performance ratio.

Whether that is more because of some AMD computational architectural advantage in suitability to this use over game suitability, or whether it is because the developers here are more successful at exploiting the potential of the AMD chips than of the Nvidia (for example, not compiling with the higher CUDA levels which have specific architectural accommodations for Maxwell and Pascal) is something I lack the data to take a position on.

Beyond that blithering generalization, we have already some specific posted RX 480 and GTX 1070 elapsed times on BRP6. That taken with a reasonable guess as to how much the GTX 1060 is likely to fall below the GTX 1070 is the primary source for my prediction.

At the price I paid for a Founders Edition 1060 (to get one fast, and to get fan sound I had confidence in) the GTX 1060 is OK. But in a few weeks one should be able to get one delivered in the US for under $260. At that price the performance per unit investment cost will be rather nice indeed, and the credit productivity per unit electric power consumption will be very fine.

Now if only the memory leak problem had a solution.

As I think this thread gets rather more readers than does the Pascal memory leak on Windows 10 thread, let me remind readers here who might be considering a Pascal card purchase with Einstein production as a significant usage that there is a currently unsolved very serious problem in running Pascal cards on Windows 10 on the BRP6 application (whether CUDA55 "test" or the CUDA32 " production). So far the only clear workaround is to reboot the system about every day or two or to restrict one's Einstein GPU work to the BFP4G Arecibo GPU work.

Comment viewing options

Select your preferred way to display the comments and click "Save settings" to activate your changes.