AMD Still Backing FMA4 Secretly
Last year, AMD released their new Zen architecture. In almost every aspect, it was superior to the previous Bulldozer based designs. Due to the improved IPC, and efficiency, the Red team is now competitive with Intel. Along with good pricing, they have managed to seize market share from Intel. In one respect, however, there has been a downgrade, with FMA4 instructions no longer support. However, it has come out that the chips still have built-in hidden support.
The FMA instructions or Fusions Multiply Add, allow for greater efficiency in processing instructions. Along with the efficiency gain, latency is down as well. Instead of having to do B x C and then (B x C) + A separately, they can be done together simply as A + B x C. The problem is that there are 2 versions of FMA, FMA3 and FMA4. The main difference is the number of operands that it can handle at once. FMA3 is simpler and faster but FMA4 is more flexible.
FMA4 May Make Comeback Eventually
FMA4 was the first to get into hardware with AMD’s Bulldozer. However, when Intel decided to support FMA3 only, AMD added FMA3 support as well. With Zen, the initial indication was that AMD got rid of FMA4, keeping just FMA3. Due to the massive market share Intel holds, most developers would only choose to use FMA3, rendering the extra work for FMA4 support moot. However, AMD has kept FMA4 around, but just hidden.
Despite the superiority of FMA4, AMD may have chosen to depreciate it for now due to bugginess. With less use by developers, the FMA4 path is less tested and buggier. Due to the importance of Zen, AMD may have decided to sacrifice FMA4 to make sure Zen put its best foot forward. With a stronger position and more time, perhaps we will see the instruction return in the future versions of Zen. Hopefully, Intel will follow suit as well.
