3 ms·

A tangential question: why haven't chip designers decided to use the extra transistors for more complex operations? Call it Hyper-CISC, if you will. We've seen

A tangential question: why haven't chip designers decided to use the extra transistors for more complex operations? Call it Hyper-CISC, if you will. We've seen it to a certain extent with MMX extensions and the like, but how much could we speed up operations if we could take pieces of the chip, introduce redundancies for more elaborate instructions, something like the old VAXes?

kjhgfvbnjmk 16y ago
They do - most of the transistors in your PC are in very specific HW on the graphics card.
- Andys 16y ago
  Today's GPU is a stream processor which is arguably less complex than a general purpose CPU. I think this is in the opposite direction of the spectrum to what the parent is suggesting.
borisk 16y ago
CPU companies don't want to pay the very high costs of designing complex extensions.
Andys 16y ago
It is happening slowly and incrementally. These changes seem architecturally dirty, but don't rock the boat too much and so are feasible to deliver on time and on budget. Check out the population count and string compare/search instructions in Intel SSE4.2, and the upcoming Intel AVX which make SSE more useful by allowing more operands, more operator combinations, and wider vectors.
kenjackson 16y ago
Probably at least a couple of reasons: 1) You have to codegen for it. It can be more difficult to codegen for a more complex operation (presumably more work getting data to the right place, and then pulled back to the right places on output), and so is less likely to be implemented. 2) I suspect Intel has already looked at this and seen little if no perf benefit. I bet if you ask 100 companies for a great complex operation that would benefit their code you might get 100 different answers. Of course, there has been talk of using reconfigurable logic, since a given process may have a good idea of what would be useful for it.

kjhgfvbnjmk 16y ago

They do - most of the transistors in your PC are in very specific HW on the graphics card.

Andys 16y ago
Today's GPU is a stream processor which is arguably less complex than a general purpose CPU. I think this is in the opposite direction of the spectrum to what the parent is suggesting.

borisk 16y ago

CPU companies don't want to pay the very high costs of designing complex extensions.

Andys 16y ago

It is happening slowly and incrementally. These changes seem architecturally dirty, but don't rock the boat too much and so are feasible to deliver on time and on budget. Check out the population count and string compare/search instructions in Intel SSE4.2, and the upcoming Intel AVX which make SSE more useful by allowing more operands, more operator combinations, and wider vectors.

kenjackson 16y ago

Probably at least a couple of reasons: 1) You have to codegen for it. It can be more difficult to codegen for a more complex operation (presumably more work getting data to the right place, and then pulled back to the right places on output), and so is less likely to be implemented. 2) I suspect Intel has already looked at this and seen little if no perf benefit. I bet if you ask 100 companies for a great complex operation that would benefit their code you might get 100 different answers. Of course, there has been talk of using reconfigurable logic, since a given process may have a good idea of what would be useful for it.