It just crossed my mind how current CPU-only coins would handle a 51% attack coming from a supercomputer. I decided to make a quick calculation based on Monero (XMR), the current CPU coin with the biggest network hashrate. (Note: there are already GPU miners for Monero but their performance is the same as CPUs, check CPU Coin List).
At the time of writing, the network hashrate for Monero is 12.01 MH/s.
An Intel Xeon E5-2697v2 does 480 H/s.
The current biggest supercomputer, Tianhe-2, has 32,000 cores of Intel Xeon E5-2692 12C processors. Supposing a similar performance this gives it 32000×480=15360000= 15.36 MH/s. Now lets take into account the 48,000 Xeon Phi 31S1P co-processors with 57 64-bit x86 cores each. Supposing a moderate performance of 20 H/s per core we get 48000×57×20=54720000= 54.72 MH/s. It would be a total of 70.08 MH/s.
Therefore, Tianhe-2 would have no trouble doing a 51% attack on Monero. To avoid it, Monero network hashrate would have to grow 12x. Like Tianhe-2, many other supercomputers from the TOP500 list could do the attack.
One of the arguments by those who support Proof of Work (PoW) algorithms with an advantage for GPUs is that these give cryptocurrencies have better chances to withstand an attack coming from a big operation. The rational is that it is much easier for a hobbyist to build a multi-GPU system than a multi-CPU one. With roughly $2000 you can now build a 8x R7 265 miner. Tianhe-2 has 48,000 Xeon Phi which could have been substituted for GPUs. Imagining that, it would have 48000÷8= 6000 times more GPUs than the hobbyist. On the other hand, with the same $2000 the hobbyist could at most (really at most) build a 2x Xeon system. In comparison the Tianhe-2 would have 32000÷2= 16000 times more processors than the hobbyist.
If the same PoW algorithm with GPU advantage is ASIC-resistant it would also avoid the threat of big ASIC operations.
CPU PoW supporters would now argue that it is easier to get more people to mine with their PCs than to get more hobbyists to build small scale miners. I tend not to agree, specially if there is a good monetary incentive to mining.
Either way, what this ultimately shows is that a mix of PoW and Proof of Stake (PoS) continues to be the best solution found so far to secure a blockchain. You can read more about it on the Vitalik Buterin post On Stake.