[r-t] Hyperthreading [was Exhausted search spaces]

[Mark Davies]

Simon Humphrey writes,

> It turns out that on my machine, with hyperthreading enabled,
> (a) a single instance of SMC32 produces 5.65m comps in 6 mins, a 2.7%
> improvement compared with HT disabled

That sounds about right. The improvement comes about purely because you 
have a couple of percent background processor usage, which without HT 
takes timeslice from SMC32.

> Since I'm now putting a new machine on the to-do list, I guess
> processor speed is the main determinant of SMC performance.  I don't
> run any other CPU intensive applications these days.
> Cheapest GHz on the market seems to be an AMD Phenom X4 3.4 at the
> moment. Any thoughts?

Yes, since any modern core has MMX, clock speed is king for SMC32. 
Having a decent-sized L1 cache is a good idea, too. If you can fit the 
entire node table in it, you can pretty much run the whole search at 
register speed, with no cache misses.

SMC32 logs the number of nodes and the number of bytes required by each 
(which it rounds up to an 8-byte boundary to improve cachability). For 
instance, in my Viking D Royal search I had 875 nodes of 136 bytes each, 
which is 119,000 bytes. That'll easily fit in L2 cache of most decent 
modern cores, but will overflow L1. However a simpler, tenors-together 
search with less falsness might typically give a node table of 30-50K. 
That has more chance of fitting in the L1 cache of the Phenom X4 (64KB) 
than my i7 860 (32KB).

So yes, a fast Phenom X4 sounds ideal. I had good experiences with my 
old Athlon X2 - lots of SMC32 power for the clockspeed. If you buy one, 
let me know how you get on. The Full Monty is the key benchmark, with 
1h49 to beat!

MBD