Understanding the actual issue with 51% attacks of PoW: The problem/solution is not "how much hash power the network has"... But rather it is "how decentralized is the hash power". Even if a network had 100 Th/s nethash.. if that's all controlled by 1 body/pool.. you are screwed anyways.
So this has nothing to do with one algo or another, but rather how many independent solo mining operations are happening around the world that has adopted one protocol over another. Bitcoins history itself was close at one point to having all the hash power centralized. And through self moderation eliminated the risk.
The reason why people like GPU/and Scrypt (at least before these next round of ASICs is coming out) for the last year, is that it seemed harder to centralize Scrypt, than SHA256 ASIC farms. Because of the decentralized nature of GPUs distributions.. and that theory is absolutely still true today..(ahead of ASICs that are just about to come out for Scrypt).
GPUs have limited supply, and more more distributed than any other type of processing available, whether it be ASIC or CPUs...
Please do not mistaken total nethash rate with distributed nethash rate. Those are related, but different issues.
PoW part of bitcoin is just a lottery system.. so the question of security is not how much hash power any protocol can do.. but rather how decentralized a particular a PoW technology/algo. In that sense GPUs will always win out (at least today), in terms of decentralization.
How many GPUs do you need to make solo mining of e.g. litecoin or DOGE worthwhile?
As everyone who has not enough tends to join a centralised pool.
So thousands of people who all have too few GPUs to feel solo mining to be worthwhile end up all flocking to a very few pools, often not even seeming to care that the pool they chose has dangerously high percentage of the hash power.
Do more GPU miners solo mine than ASIC miners who solo mine?
Maybe another way to compare would be how many litecoin p2pool nodes are there and how many bitcoin p2pool nodes are there?
-MarkM-