Best performance per watt

[Picuino]

Now that low processor power consumption is so fashionable, but computing power is still highly valued, I was wondering which processor will have the best ratio of these two variables.
Of course, the main problem with this question is to define what is computing power. It could be measured in MIPS and MFLOPS but they are very erratic measures.
Does anyone know if there is any serious study or ranking on this point and which is the best processor?

Regards.

[Picuino]

* Performance per watt (Wikipedia)
* New RISC-V CPU claims recordbreaking performance per watt

Only for desktop computers and laptops CPUs:
 * Power Performance (PassMark)
 * Which CPU gives the most performance per watt? (Technoscad)
 * Performance per Watt Is the New Moore’s Law (ARM)

[golden_labels]

No, but I think that neither synthetic benchmarks nor per-TDP values are a good measure.

Benchmarks suffer the same issues as they do in general peformance measurement. They don’t use realistic or expected workloads. And there is always the risk of results being volkswagened. Performing tasks, which one is likely to actually use the CPU for (video encoding or running a server handling realistic requests), is more accurate.

TDP indicates thermal requirements. Higher energy consumption and heavier loads are obviously correlated. But not neccessarily maximum performance implies operating at maximum power. Actual energy consumption should be measured.

But it’s hard to attract big audience by publishing tests irrelevant to them. And it’s more expensive to measure power than just check specs on manufacturer’s website.

[Picuino]

Nordic Semiconductor proves world-leading processing efficiency with the revolutionary nRF54H20 SoC
* https://www.nordicsemi.com/Nordic-news/2023/09/nordic-semiconductor-proves-world-leading-processing-efficiency-with-the-revolutionary-nrf54h20-soc
* https://www.eembc.org/ulpmark/ulp-cm/scores.php

[coppice]

I couldn't find any real details about the nRF54H20, but the description of it having some ARM M33 cores and some RISC-V cores sounds like a hotch-potch for a new generation device.

[brucehoult]

Now that low processor power consumption is so fashionable, but computing power is still highly valued, I was wondering which processor will have the best ratio of these two variables.

Absolutely meaningless question without specifying the workload.

- do you need to run an OS, or bare metal?

- how much address space is needed?

- what is the element size of the important data?

- is the data integer? fixed point? floating point? Is overflow likely and does it need special handling e.g. saturation?

- does the algorithm use a small amount of code and a lot of data, or complex algorithms on small amounts of data?

- is the control flow predictable?

- is the data flow predictable?

- does the algorithm benefit from SIMD/SIMT?

- can the algorithm be partitioned into dozens / hundreds / thousands of independent cores? With how much communication?


There will be workloads most efficiently done by an 8 bit AVR-like ISA (whether or not Atmel's actual chips are power efficient). There will be workloads most efficient done by an FPGA. There will be workloads most efficiently done by a GPU. There will be workloads most efficiently done with a very wide SIMD/Vector e.g. 16k bits or more.

Ditto for Google's TPUs, Esperanto's ET-SoC-1 chip with 1088 simple in-order ET-Minion cores, each with big FP and int vector units, Microsoft's Azure Cobalt 100 with 128 Arm N2 cores ...

Pretty much every different or weird thing out there is aiming to get maximum performance per Watt or per dollar on something.

[SiliconWizard]

Indeed. You may get a very impressive performance/energy ratio with some recent ultra-low power MCUs, for instance, but the architecture to get there wouldn't scale up to anything useful for, say, desktop computing.
So that can't be considered out of context.

[David Hess]

Too often the benchmarks leave out the power drawn by necessary auxiliary circuits, like PCIe and RAM, so power benchmarks of entire systems are more realistic.

[Marco]

Meteor lake seems to be finally delivering on the low power promises of Alder Lake, able to scale far lower.

[Picuino]

"The ULPMark-CM score is unitless, but is derived from iterations per milli-Joule, divided by 1000."
https://www.eembc.org/ulpmark/ulp-cm/scores.php

In this sense the "iterations" can be simple 8-bit instructions, powerful 64-bit instructions or floating point calculations.
For the case at hand, I would define iteration as 32-bit RISC instruction, which is almost the current standard from microcontrollers to smartphone microprocessors.

You could even refine it further and say that it refers to ARM or similar cores (such as MIPS).

[coppice]

"The ULPMark-CM score is unitless, but is derived from iterations per milli-Joule, divided by 1000."
https://www.eembc.org/ulpmark/ulp-cm/scores.php

In this sense the "iterations" can be simple 8-bit instructions, powerful 64-bit instructions or floating point calculations.
For the case at hand, I would define iteration as 32-bit RISC instruction, which is almost the current standard from microcontrollers to smartphone microprocessors.

You could even refine it further and say that it refers to ARM or similar cores (such as MIPS).

You can define iteration however you want, but the ULPMark tests are based on iterations of the ULPMark test code.

[radiogeek381]

The biggest problems with the OP, as almost all the replies have pointed out, are the words “Best” and “performance.” 

“Per Watt,” as others have also pointed out, is subject to perspective. All of the perspectives are right, somewhere.

Some datacenter operators look at lots of factors including wall power (kVA out of the socket),  cooling power,  power/rack-unit, and even the cabinet exhaust air temperature and flow. Socket power is irrelevant for them.

As for “performance” and “best” this is generally unanswerable even when we eliminate the “power” part of the question.

There have been attempts at benchmark suites in the HPC part of the industry. The Green500 suite is the one I know of. https://www.top500.org/lists/green500/ This was derived by grownups, but its reliance on the Linpack benchmark.

The Green500 ignores any “special” aspect of the hardware that doesn’t directly contribute to their specific approach to a solution of a linear system. It also focuses on double-precision floating point performance. This leaves out a huge chunk of the HPC industry at this point. On the other hand, it is almost reasonably objective when comparing x86 based systems. The huge investment in Linpack by the intel side of the industry, however, puts most other architectures at a bit of a disadvantage.


Just sayin'

[nightfire]

The energy consumption of a CPU is somewhat meaningless without knowledge about the task it shall fulfill, and which other parts in a computer are needed to do so.

In my last company, we did webhosting, and my colleagues did some benchmarks to evaluate new server hardware regarding performance and operating costs- basically the metric was something along the lines of: Number of delivered webpagaes /page impressions per hour and per watt to compare with other models of hardware.

Everything else like synthetic benchmarks will give you some rough idea how some CPU might compare against each other, but the whole ecosystem has to be taken into account.