The performance per watt of very slow CPUs is always very good, so that alone is not impressive unless much more data is provided to prove that is really better than what is typical.
On the other hand, the performance per area of the "traditional" CPU that is used for comparison, which I assume that it is some kind of RISC-V, is very poor, so that the performance per area of the new CPU is likely to be also poor, even if much better than that of the strawman.
Their "traditional" CPU has an area that is slightly bigger than that of Cortex-A720 made with a CMOS process almost twice as dense.
So if they were made with the same process, the "traditional" CPU and the new CPU would have an area somewhat bigger than a half of a Cortex-A720.
But Cortex-A720 is an out-of-order 4-wide superscalar core with advanced SIMD and scalable vector extensions, with a good branch predictor and with many other enhancements, for which the area ratio in the same process is expected to be much greater than 2 in comparison with the CPUs evaluated in TFA, while its performance must be several times greater than that of a 2-wide in-order CPU core.
So both the "traditional" CPU and the Frostbite CPU appear to have a much bigger area than expected for their performance.
Moreover, the performance of in-order CPUs is very sensitive to the benchmark selection and to the quality of the compiler optimizations.
So the performance obtained for the benchmarks selected by them, like FFT or LU decomposition, is expected to be much better than for benchmarks that would execute an application not dominated by such regular loops, but having lots of hard to predict branches, e.g. a compiler benchmark.
They say that the ISA will be published. At that time it will be easier to appreciate whether this CPU architecture is really innovative. For now, they have done a lot of work, by fully implementing the CPU core, but the data presented is not sufficiently convincing to demonstrate that it is indeed an architecture that will be able to compete with SOTA CPU cores, e.g. with those from the Arm company.
I am of the opinion that even if this novel architecture doesn’t yield dramatic improvements in peak performance across broad classes of workloads, it may still be a welcome improvement in the legibility of software performance down the stack.
By this I mean: currently, transforming a program written in a typical high-level language to optimized ISA code involves what can only be described as black magic inside the compiler. If you don’t know the implementation details of modern superscalar OOO branch-predicting microarchitectures then you may have no clue why two very similar-looking functions could have dramatically different running times. Cue Matt Godbolt tech talk…
What I see as the more promising side of this type of development is a tighter coherence, and thus more transparency in expected runtime performance, between the software above the ISA and the hardware below it. Essentially we now have a linear machine with a linear source language, and with the popularity of Rust nowadays we actually have a critical mass of developers who are already used to applying ownership discipline in program design.
> The performance per watt of very slow CPUs is always very good
Generally this has not been my experience. Typically both very slow CPUs like the ATMega328 and high-performance CPUs like an Intel i7 are in the neighborhood of 1 nanojoule per instruction. (For operations more than 8 bits wide, the AVR is significantly worse.) Low-power CPUs like the STM32L line are commonly around 0.2nJ/insn. Ambiq's subthreshold logic gets them closer to 0.05nJ/insn.
I agree with the rest of your comment.