Some more results from me:

Intel C2D 6320 slightly O.C. to 2.1GHz:  16kh/s   -t 2
Intel Celeron G1620 @ 2.7GHz:      26kh/s     -t 2
AMD Athlon 64 3200+ @ 2.0GHz:    7kh/s     -t 1

All running Ubuntu 12.04.

Ahem..
Sorry for skewing the measurement somewhat.
I must admit I tweaked the miner that Wolf and Joe had already optimized. So I think my miner is approx. 10% faster.
Differences on the OS side: My guess is that mining under Windows is slower due to all the background tasks that
seem to run incessantly and beyond user control in Windows.

My Macbook Air 2013 has an I5-4250U processor (2-core 4-thread cpu)
Running -t 4 gives 28kh/s.  MacOS (obviously) Yosemite.

Hefty,
I have one i5-4460 Has(h)well on ubuntu 12.04.
It does  ~81kh/s @ -t 4 at stock freq, stock cooler, temp goes up to 65 deg.

One Intel Centrino (Thinkpad T42)  one core running at 2Gig does  ~2.1kh/s

Nice work.  That card is running a tad hot even for a reference card.
I replaced the thermal paste on my own card and temp went down from 84 to 78 degrees, power draw went down 10W too.

Have you remembered to add the Poloniex payment id in the payment address field on the pool?
Address that the pool sees should be
xmr address.payment id

I was thinking power failure, too.
However, the original 3-card rig was drawing 460Watts at the wall thru a Corsair 750W rated PSU so I think
I had good margin.
Now the card is sitting alone in a spare rig to see if it runs stably overnight.
Thanks again for tipping about the caps and possible power issue.

Thanks for the tip.
The card was not broken.
I reverted back to kernel version 3.2.0-72 from 3.2.0-74 and reinstalled the stock driver.
And voila, the card is now working again and sgminer is hashing away.
I have had instabilities with kernel version 3.2.0-74 and higher in another rig, and this
strengthens my belief that something was broken in kernel version 3.2.0-74 and higher.

This card was sitting in a rig with 2 other R9 280X cards running sgminer 5 under Ubuntu 12.04 when suddenly the
miner hung and top command revealed  Xorg at 99% cpu.  Driver problem, I thought and rebooted the rig.
But Ubuntu refused to boot up.   I took the card out and inserted into another rig.  This time the machine booted
but refused to start X.  Stock driver under 12.04.
I took the card out and into a Windows 7 machine.  This machine booted no problem and I was able to start sgminer.
Now, however, I am not able to adjust core or memory frequency thru sgminer, only fan % is adjustable. 
I am able to adjust these parms from Sapphire Trixx, however.
Anyone has a clue what is going down here?

I read it from the pool dashboard. I should a made a screen snapshot, but got busy.. At that same time the maxminers pool dashboard
showed a pool rate of 2 MH/s something and the same net hashrate as nonce-pool.  I hope it is the net hash rate that was showing wrong
for a minute.

Right now:

Net hash rate: 15,562.11 KH/s
Pool hash rate on Magi@Nonce-Pool: 19,236.5 KH/s
Wtf?

Looks like you are building for the future
If I were to build a similar system in my country (Norway), I would have had to fork out the equivalent of $1759USD for those parts, plus
a bit for a cooler on top of that.
An i7-5930K alone costs $700 over here.
A more modest i7-4790K costs $400.
Anyways, If one is looking to buy new gear, the only game in town for serious CPU mining seems be Has(h)well-based chips,
from any I5 and upwards.

I got another 10% speed improvement on my Has(h)well I5 box by tweaking the already excellent mods done by Wolf/Joe .
From 72 KHash/s to 81 KHash/s     running 4 threads on my i5-4460 , no overclock

This tweak/hack uses a different implementation of the sha256 function and will only work on chips with AESNI/AVX2 .
Here is how I did it.  Read on if you are interested (Warning: The below will be a bit technical and applies to Linux only):

So I copied the source code from http://lxr.free-electrons.com/source/arch/x86/crypto/sha256-avx2-asm.S.
The code in here replaces the sha2_round code found in the m7/sha2.c file.
I edited it carefully by removing the line numbers and commented out the following lines
// #ifdef CONFIG_AS_AVX2
// #include <linux/linkage.h>
..
.global sha256_transform_rorx  ## ENTRY(sha256_transform_rorx)
..
//  ENDPROC(sha256_transform_rorx)
and the bottom line
// endif
Then stored the resulting as sha256_rorx.S in the m7 subfolder of the miner source.
Then edited the m7/sha2.c as follows:

// Added this line above the below C function:
extern void sha256_transform_rorx(const void *input_data, sph_u32 *digest, unsigned long num_blks);

// Then the following changes to this function.
static void
sha2_round(const unsigned char *data, sph_u32 r[8])
{
// Comment out these 3 lines
//#define SHA2_IN(x)   sph_dec32be_aligned(data + (4 * (x)))
//        SHA2_ROUND_BODY(SHA2_IN, r);
//#undef SHA2_IN
// Add this line:
   sha256_transform_rorx(data, r, 1);
}

saved the above file, then ran the following commands:
run 'make' in the m7 folder.
Then run 'gcc -c  sha256_rorx.S'
Then 'cd ..'
Then:
gcc -std=gnu99  -O2 -pthread -fuse-linker-plugin   -o minerd minerd-cpu-miner.o minerd-util.o minerd-sha2-old.o minerd-scrypt.o minerd-m7mhash.o minerd-haval.o minerd-keccak.o minerd-ripemd.o minerd-sha2.o minerd-sha2big.o minerd-tiger.o minerd-whirlpool.o m7/sha256_rorx.o -ljansson -lpthread  -lgmp -lcurl -lm

There are other versions of the above, so for example you might try the sse3 version if you only have a core 2 duo, for example.

Great work again Wolf0 and Joe,

However, the new minerd core dumped in my Xubuntu 12.04 box running on an Athlon 64 bit cpu.
I have never liked the -Ofast compiler option, so I removed it from the makefiles.  Then I ran make clean and reran
make and after that the new minerd runs fine.  Speed increase from 5.03 KHash/s -> 7.23 KHash/s.  43%!!

Hi Joe,
Great work!
Could you post a Linux version of the miner - either the source or a statically linked generic version?
Your posted generic version brought my Athlon on Windows 7 from 5.03 to 6.3 kHash/s
EDIT: Not 6.3, but 7.41 whoa

I compiled Wolfs version of the miner on Ubuntu 12.04, i5-4460 @ 3.2GHz and my girlfriends MacBook Air i5-mobile something
Running 2 threads for now on my Ubuntu i5:  14.5 kHash/s each thread = 29 kHash/s.
The original version ran at 21 kHash/s, so almost 40% improvement!!
Same %age improvement on the mac.   Hope she wont notice the extra heat - she doesnt know i hacked her machine..
Thanks Wolf for your great work.  Expect some coin coming your way soon. You deserve it for the effort ;-)

Has anybody tried the optimized  avx2 sha256 routine that can be found at
http://lxr.free-electrons.com/source/arch/x86/crypto/sha256-avx2-asm.S
??

Hi again,
Now I tried the "improved" kernel on my own 280X rig.
3 cards, all running 1000/1500 core/mem.  550Watts at the wall 
Orig neoscrypt kernel (Kh/s)
  301
  296
  287
My "improved" kernel
  295
  289
  276
Yiikes!  I got worse performance on the 280X! 
Sorry guys.  This "improvement", as it stands, seems to come to the 290 only.

Those cards must be screaming 
How much of a %age improvement did you get?

Here is a small neoscrypt kernel improvement for free, since I am mostly doing X11 anyway.
It gave me a 5.8% speedup on my reference R9 290 card (with Stilt bios),
from 290.2 to 307Kh/s  at 800/1500 core/mem freq on Ubuntu 12.04 with stock drivers.
I didnt try it on my R9 280x cards, so please post your results if you try this.

You will have to mod the kernel as per the code below.
The bottleneck in this kernel is the way it stores the 128 intermediate results of chacha and salsa in global memory.
By doing the change below you are reducing stalls/latency by not making read/writes to same/adjacent memory banks.

Change:
void ScratchpadStore(__global void *V, void *X, uchar idx)
{
   ((__global ulong16 *)V)[idx << 1] = ((ulong16 *)X)[0];
   ((__global ulong16 *)V)[(idx << 1) + 1] = ((ulong16 *)X)[1];
}

void ScratchpadMix(void *X, const __global void *V, uchar idx)
{
   ((ulong16 *)X)[0] ^= ((__global ulong16 *)V)[idx << 1];
   ((ulong16 *)X)[1] ^= ((__global ulong16 *)V)[(idx << 1) + 1];
}

To:
void ScratchpadStore(__global void *V, void *X, uchar idx)
{
   ((__global ulong16 *)V)[idx] = ((ulong16 *)X)[0];
   ((__global ulong16 *)V)[idx + 128] = ((ulong16 *)X)[1];
}
void ScratchpadMix(void *X, const __global void *V, uchar idx)
{
   ((ulong16 *)X)[0] ^= ((__global ulong16 *)V)[idx];
   ((ulong16 *)X)[1] ^= ((__global ulong16 *)V)[idx + 128];
}