01-15-2021, 06:37 AM
Hello,
Great question. Our entire research is based on developing systems that can do more with less power/energy.
First of all, it is important to differentiate between power and energy. Power is all about how much current your circuit draws in an instant. But it does not really tell you how much work your circuit does. The story is a bit more complicated, but basically if you run a circuit at half the speed (half the MHz), the power will also halve. Obviously your circuit will also take twice as long to finish the same job. This is where energy comes in. Energy is Power x Time. So in the above example (grossly simplified), power would be half, but the energy would be the same.
They are both important, (again much simplified) but if you are looking to optimize the life time of a battery operated device, it is energy that is more of a concern (as the battery only stores a certain amount of energy), but when you talk about solar panels (energy harvesting), there is a certain power that you can obtain from a panel, and you have to make sure that you stay below that level.
The tricky part is, you can optimize circuits for low power or more energy efficient operation, but the optimizations are not always the same. We identify two different sources of power consumption. One is dynamic (when the circuit is doing something useful), and the other one is static (when the circuit is powered, there is some unwanted current flow). Normally the static power consumption is really small, and you ignore it, but as you save more and more power, the static power consumption starts to dominate, so some of the tricks that worked well for you, stop working after a while. Reducing the operating voltage is one of them. Once you reach near threshold (or sub threshold) you basically reach a point where dynamic and static power consumption are roughly equal.
So what can you do for low power / higher energy efficiency
* First of all it is a technology question. Newer technologies (7nm, 22nm) are smaller and faster transistors than older (180nm).
* Since static power has become so important, newer technologies do not have ONE transistor, there are usually 10 or more variations of transistors that change may have 100x difference in their static power. Take a look at https://www.synopsys.com/designware-ip/t...-16ff.html for a longer description of different flavors in ONE technology
* Smaller area means less transistors, everything else being equal, less static power. So if you are going for POWER, (again simplified) smaller is better. Our smaller micro/zeroRISCY (now Ibex) are optimized for this reason
* Energy efficiency means that you can finish a given program quicker, parallelization and making sure that all units are kept busy and none is waiting doing nothing (and wasting power) are key tricks to increase the efficiency. RI5CY (now CV32E40P) was adapted for this. It is much bigger (more kGE) but it is more ENERGY efficient than its smaller brothers.
* Whereas Ariane (or CVA6) was built just to run Linux, it is not particularly optimized for one or the other. Since it is a 64 bit core, it is much larger (about 4x RI5CY), it is not really low power
* Whatever the core, the memory needed is MUCH MORE important actually. roughly speaking 5-10 kGEs is the size of 1 kilobyte memory. Basically 8kByte memory is one RI5CY core. (or 2kByte memory is one Ibex core).
All these are gross simplifications and approximations, but our entire research is based on developing architectures that are better.
Practically all our presentations are online under : https://pulp-platform.org/conferences.html These contain all the information in the papers that you may not have access to.
I hope this helps,
Great question. Our entire research is based on developing systems that can do more with less power/energy.
First of all, it is important to differentiate between power and energy. Power is all about how much current your circuit draws in an instant. But it does not really tell you how much work your circuit does. The story is a bit more complicated, but basically if you run a circuit at half the speed (half the MHz), the power will also halve. Obviously your circuit will also take twice as long to finish the same job. This is where energy comes in. Energy is Power x Time. So in the above example (grossly simplified), power would be half, but the energy would be the same.
They are both important, (again much simplified) but if you are looking to optimize the life time of a battery operated device, it is energy that is more of a concern (as the battery only stores a certain amount of energy), but when you talk about solar panels (energy harvesting), there is a certain power that you can obtain from a panel, and you have to make sure that you stay below that level.
The tricky part is, you can optimize circuits for low power or more energy efficient operation, but the optimizations are not always the same. We identify two different sources of power consumption. One is dynamic (when the circuit is doing something useful), and the other one is static (when the circuit is powered, there is some unwanted current flow). Normally the static power consumption is really small, and you ignore it, but as you save more and more power, the static power consumption starts to dominate, so some of the tricks that worked well for you, stop working after a while. Reducing the operating voltage is one of them. Once you reach near threshold (or sub threshold) you basically reach a point where dynamic and static power consumption are roughly equal.
So what can you do for low power / higher energy efficiency
* First of all it is a technology question. Newer technologies (7nm, 22nm) are smaller and faster transistors than older (180nm).
* Since static power has become so important, newer technologies do not have ONE transistor, there are usually 10 or more variations of transistors that change may have 100x difference in their static power. Take a look at https://www.synopsys.com/designware-ip/t...-16ff.html for a longer description of different flavors in ONE technology
* Smaller area means less transistors, everything else being equal, less static power. So if you are going for POWER, (again simplified) smaller is better. Our smaller micro/zeroRISCY (now Ibex) are optimized for this reason
* Energy efficiency means that you can finish a given program quicker, parallelization and making sure that all units are kept busy and none is waiting doing nothing (and wasting power) are key tricks to increase the efficiency. RI5CY (now CV32E40P) was adapted for this. It is much bigger (more kGE) but it is more ENERGY efficient than its smaller brothers.
* Whereas Ariane (or CVA6) was built just to run Linux, it is not particularly optimized for one or the other. Since it is a 64 bit core, it is much larger (about 4x RI5CY), it is not really low power
* Whatever the core, the memory needed is MUCH MORE important actually. roughly speaking 5-10 kGEs is the size of 1 kilobyte memory. Basically 8kByte memory is one RI5CY core. (or 2kByte memory is one Ibex core).
All these are gross simplifications and approximations, but our entire research is based on developing architectures that are better.
Practically all our presentations are online under : https://pulp-platform.org/conferences.html These contain all the information in the papers that you may not have access to.
I hope this helps,
Visit pulp-platform.org and follow us on twitter @pulp_platform