7970 more than linear performance

[blazer]

Question

I noticed on the ocl-plus page that the 7970 appears to be scaling higher than linearly with peak performance with integers.

Eg.

5850 (ppi 1044.00):

MD5: ~2 000 000 000
NTLM: ~2 600 000 000

1.3x Faster for NTLM

6950 (ppi 1126.40):

MD5: ~2 300 000 000
NTLM: ~3 000 000 000

1.3x Faster for NTLM

7970 (ppi 1894.40):

MD5: ~4 700 000 000
NTLM: ~7 400 000 000

1.57x Faster for NTLM

Questions

• Whats new in the 7970 archetecture allowing a greater speed difference between MD5 and NTLM ?
  
• While the ppi did not double, the MD5 performance increased by more than 2x for 7970
  
• How come the 7970 performs so much better for other hash types such as DES when compared to a 5xxx series card. (Eg Not shown but 1x 7970 will outperform 2x 5850 on DES even though 2x 5850 has higher ppi than 1x 7970)
  

I'm just wondering whats new in 7950 allowing this, any explanations ?

[atom]

No more VLIW is the answer. No VLIW, less register usage, faster speed.

Coding for hd7970 is as simple as CUDA.