50-70W? For a atom or via board? In what universe?

Had the same thing once, pulling the plug + 10 min wait magically fixed it.

+1 for MSI 890FXA-GD70
3 cards in standard cases with airspace, can do 6 cards with risers, 6-pin connector onboard so no need for powered risers.
And +1 on "don't put cards right next to each other, unless you like overheating cards and dead fans".

Well, if that's your guess... I could tell you what that's *very* unlikely, but... meh.

I'd stay away from Scythe UKs, got a box full of em. Sleeve bearings shot after a few months @ 100%.
Btw, OEM of the Scythe UK line is Young Lin Tech Co.
If you're looking for decent rad fans in the size/power category of a UK3k...
delta afb1212hhe
panaflo fba12g12u
...
or really pretty much anything 120x120x38 from a reputable mfg. with ball bearings and 5-7W.

Yeah, a FF784 package on a FF780 footprint. right.
While we're making random shit up, EP3SL70 or 110 (can't see any way to fit 120 rounds in a SL50, even 64 single-stage rounds is pushing it).

Can't set anything lower, when I try it *shows* lower but jumps back to 1375 (obvious in power consumption @ 925 core)

I'd subtract 108W or so, should get decent estimate of real card power at the wall with a 92% efficient PSU.

*shame* disregard those results, somehow managed to completely botch it.

I think I mixed up hashrates, clocks and power measurements from different sets of runs or skipped lines somehwere.
Great, have to throw all data out and start over :/

Phenom II 1100T @ 4.0GHz, cpu freq scaling disabled.
idle, display off 122W
idle @ desktop 139W
phoenix 1.7.3, -k poclbm AGGRESSION=10 WORKSIZE=256
stock V @ 50% fan
925 core/1375 mem, 554 Mh/s, 369W
925 core/1020 mem, 554 Mh/s, 360W
1070 core/1020 mem, 640 Mh/s, 386W
1170 core/1020 mem, 698 Mh/s, 407W
1.20 V @ 55% fan
1170 core/1020 mem, 698 Mh/s, 421W
1200 core/1020 mem, 714 Mh/s, 428W
1240 core/1020 mem, 736 Mh/s, 431W
1.25V @ 70% fan
1240 core/1020 mem, 736 Mh/s, 467W
1270 core/1020 mem, 753 Mh/s, 476W
1.30V @ 80% fan
1270 core/1020 mem, 753 Mh/s, 509W
1350 core/1020 mem, 799 Mh/s, 525W

not sure how I got 695/759/804, guess that was a run at 1375 mem

Now, after massive thread derailment... back to OPs question.
stock vs. stock 2*6870 easily beats 6970. 2*~290 = ~580 vs. ~390 for a 6970.

Well, iirc GPU-Z shader # is simply based off a hardcoded table of device IDs, so that's no big help either.
Atm I'm out of ideas of how to figure out what is really going on there, just seems *very* odd that your 5850 is 11% faster than what everyone else is reporting, and 111% just happens to be very close to 1600/1440.

Yeah, but the PCI device ID says it's supposed to be a 5850...
Honestly not sure WTF to make of that, all I get is "congratulations, you're the proud owner of a 5850 with 1600 enabled shaders."

It's rather simple to convert between cards of the same family, mining scales pretty much 100% with #shaders * clock.
430Mh/s / 930MHz / 1440 shaders = 0.000321087 hash/shaderclock
0.000321087 hash/shaderclock * 850 MHz = 272923.95 hash/shadersecond
272923.95 hash/shadersecond * 1600 shaders = 436678320 hash/sec.

so to convert, it's generally simply hashrateA / clockA / #shadersA * clockB * #shadersB

Nope, haven't found any better settings than what other people have been saying for Diablo, best I found was ~2.2% worse than phoenix.

430Mh/s at 930 core on a 58*50*?
That'd be 335Mh/s stock. On a 5850.
Or 436Mh/s on a stock 5870.
So... why is everyone else reporting numbers *way* below that?

Just tried it for shits and giggles... 7970 primary, 6870 secondary w/ the 7970 drivers from AMDs site for windows and 11.12 for linux. works.

Quick calc says we're already a few % higher hashrate/shader/clock than VLIWx.
I'd have to take a closer look at how GCN shader ASM looks and what engine performance is like, but wouldn't expect any miracles, maybe another 2-3% increase.
So my opinion so far "pretty much a wash, a hair better than 5xxx per shader per clock, may get another few % when more time is spent tuning kernels for GCN".
Mining is really unique in how it manages to get well > 90% peak throughput even on VLIW5

Now that I can agree with.
The work required to add OpenCL support is pretty massive, and AFAIK there's pretty much 0 done on a open source OpenCL -> VLIWx shader asm compiler, several other large parts of the puzzle are also completely missing (runtime, runtime/opengl/driver integration/...).
But saying that the open source driver can't use shaders is wrong.
Now, for something like a dedicated miner for 5/6xxx you'd "only" have to handcraft a kernel for each arch in VLIWx ASM (fully documented in public AMD docs btw) and hack the radeon driver so it allocates input/output regions and loads/runs your shader program on command. Iirc there's still a open 200BTC bounty for a miner working like that. but it's a *lot* of work for very little benefit.

1-upd!
max @ stock V: 1070 core, 695Mh/s
@ 1.20V: 1170, 759Mh/s
and for shits and giggles...
@ 1.25V: 1240 core, 805Mh/s

Tip: try phoenix with -k poclbm AGGRESSION=10 WORKSIZE=256

Err, what?
Take a look at the Xorg radeon driver and how it does several parts of 2D acceleration on 5xxx and 69xx