03-16-2015, 05:42 AM
295X2 doesn't just throttle for heat, it also throttles for power consumption. 4096 cores on GCN with the die shrink will likely want to draw somewhere between 375W - 425W at full load, but the card is rumored to only have a TDP of 300W. That means it will throttle. How much it throttles all depends on what they do with PowerTune in the 390X's firmware.
HBM is great and all, but likely won't translate into any significant performance increase for hash cracking. It might mean less parasitic loss for multi-hash cracking, but that remains to be seen.
No, it didn't take Nvidia 3 years to catch up to GCN. On the contrary, GCN looks an awful lot like an Nvidia GPU. Nvidia has always produced superior GPUs, they just haven't historically haven't shown good performance for hash cracking until Maxwell. There may be some confusion as to why that is, and what exactly Maxwell changes, so let's review that.
Nvidia has always opted for fewer cores that run at a much higher clock, whereas AMD has always opted for more cores at much lower clocks (though by ditching VLIW and adopting a more Nvidia-esque design, it enabled them to increase their clock rates substantially) This meant that AMD was able to perform more integer operations per second than Nvidia, which matters for hash cracking.
But the real reason AMD has had the edge over Nvidia for hash cracking is that AMD has had instructions for BFI_INT and BIT_ALIGN_INT, which are heavily used in hashing algorithms. This means that we can reduce the instruction count for most algorithms on AMD vs Nvidia, since Nvidia was lacking these instructions in their ISA. Maxwell changes this because it introduces LOP3.LUT, which is a single instruction that can do everything BFI_INT and BIT_ALIGN_INT can do, plus more.
This is why Maxwell with 1664 cores @ 1250 Mhz can now match the hash cracking performance of GCN with 2816 cores at 1000 Mhz. And they can do this while drawing *half* the power that GCN draws.
HBM is great and all, but likely won't translate into any significant performance increase for hash cracking. It might mean less parasitic loss for multi-hash cracking, but that remains to be seen.
No, it didn't take Nvidia 3 years to catch up to GCN. On the contrary, GCN looks an awful lot like an Nvidia GPU. Nvidia has always produced superior GPUs, they just haven't historically haven't shown good performance for hash cracking until Maxwell. There may be some confusion as to why that is, and what exactly Maxwell changes, so let's review that.
Nvidia has always opted for fewer cores that run at a much higher clock, whereas AMD has always opted for more cores at much lower clocks (though by ditching VLIW and adopting a more Nvidia-esque design, it enabled them to increase their clock rates substantially) This meant that AMD was able to perform more integer operations per second than Nvidia, which matters for hash cracking.
But the real reason AMD has had the edge over Nvidia for hash cracking is that AMD has had instructions for BFI_INT and BIT_ALIGN_INT, which are heavily used in hashing algorithms. This means that we can reduce the instruction count for most algorithms on AMD vs Nvidia, since Nvidia was lacking these instructions in their ISA. Maxwell changes this because it introduces LOP3.LUT, which is a single instruction that can do everything BFI_INT and BIT_ALIGN_INT can do, plus more.
This is why Maxwell with 1664 cores @ 1250 Mhz can now match the hash cracking performance of GCN with 2816 cores at 1000 Mhz. And they can do this while drawing *half* the power that GCN draws.