GTX 970 Benchmark

[TheJoSko]

I spend a lot of time cracking NetNTLMv2 hashes.

Brute-force and or mask attacks I get what I consider great throughput:

/cudaHashcat-1.31/cudaHashcat64.bin --gpu-loops=1024 --gpu-accel=256 --remove -o /Hashes.cracked -a 3 -m 5600 /Hashes.txt ?u?l?l?l?l?l?a?a

Hash.Type......: NetNTLMv2
Speed.GPU.#1...: 221.6 MH/s
Speed.GPU.#2...: 221.1 MH/s
Speed.GPU.#*...: 442.7 MH/s
Recovered......: 0/1 (0.00%) Digests, 0/1 (0.00%) Salts

However with a large wordlist I am seeing what I would consider abysmal throughput:

/cudaHashcat-1.31/cudaHashcat64.bin --gpu-loops=1024 --gpu-accel=256 --remove -o /Hashes.cracked -m 5600 /Hashes.txt mega2.txt

Hash.Type......: NetNTLMv2
Speed.GPU.#1...: 4348.1 kH/s
Speed.GPU.#2...: 4356.3 kH/s
Speed.GPU.#*...: 8704.4 kH/s
Recovered......: 0/1 (0.00%) Digests, 0/1 (0.00%) Salts


Are my expectations inappropriate, or is there something that needs to be addressed?

The wordlist is stored on a Samsung EVO 850 SSD which I perceive to have adequate io:
# hdparm -Tt /dev/sda
/dev/sda:
Timing cached reads: 9826 MB in 2.00 seconds = 4916.81 MB/sec
Timing buffered disk reads: 1562 MB in 3.00 seconds = 520.47 MB/sec