Comment by delusional

E-cores can typically execute ~4-instructions per clock tick in highly optimized code.

P-cores go up to like... 6-instructions. Better yes, but not dramatically better. The real issue is that P-cores have far more resources than E-cores: deeper reorder buffers to perform more out-of-order execution. Deeper branch prediction, more register files, larger caches, etc. etc.

So P-cores should be hitting the max of 6-instructions per clock tick on more kinds of code. E-cores have much smaller caches (and other resources) so they'll run out and start stalling out to memory-limitations, which is like 0.1 instructions per clock tick or slower.

----------

But guess what? If a fancy P-core is memory-bound (like a lot of Blender code, due to the large-scale dozens+ GBs nature of modern 3d scenes), then those fancy P-cores run out of resources and are 0.1 IPC as well.

If both P-cores and E-cores are stalled out waiting on memory, you'd rather have 32x E-Cores all executing at 0.1 IPC, rather than only 8x P-cores executing at 0.1 IPC.

Its going to be a complex world moving forward: modern CPUs are growing far more complex and its not clear what the tradeoffs will be. But this reality of E-core and P-cores stalling out and waiting on memory is just how modern code works in too many cases.

And remember, its 4x E-cores are equivalent in area/costs to 1x P-core. So there's no contest in terms of overall instructions-per-second for E-core vs P-cores. The E-cores simply are better, even if the individual threads run slower.

Indeed a single thread is most simple to reason about, but if you have a single task that can already use 2 cores uniformly, going to 8 cores (assuming enough workload) should be a pretty clean 4x speedup (as long as you don't run into memory bandwidth limits, but that'd cap the single-threaded code too).

But the performance difference between E-core and P-core perf is way less than 4x; the OP article shows a 1.6x/1.7x difference in SPEC for skymont vs lion cove, and 1.3x/1.7x for crestmont vs redwood code; and some searching around for past generations gives numbers around 1.4x.

Increasing core counts being a much more area- and energy-efficient way for hardware to provide more total performance than making the individual cores faster is a pretty fundamental thing.

For stuff like path tracing you have to work very hard not to trash the caches, so you're often just waiting for memory.

That's why such workloads gets a near linear scaling when using hyper-threads, unlike workloads like LLMs which are memory bandwidth bound.

Obviously it is easier to write any program as a single sequential thread, because you do not need to think about the dependencies between program statements. When you append a statement, you assume that all previous statements have been already executed, so the new statement can access without worries any data it needs.

The problem is that the speed of a single thread is limited and there exists no chance to increase it by significant amounts.

As long as we will continue to use silicon, there will be negligible increases in clock frequency. Switching to other semiconductors might bring us a double clock frequency in 10 years from now, but there will never be again a decade like that from 1993 to 2003, when the clock frequencies have increased 50 times.

The slow yearly increase in instructions per clock cycle is obtained by making the hardware do more and more of the work that has not been done by the programmer or the compiler, i.e. by extracting from the sequential program the separate chains of dependent instructions that should have been written as distinct threads, in order to execute them concurrently.

This division of a single instruction sequence into separate threads is extremely inefficient when done at runtime by hardware. Because of this the CPU cores with very high IPC have lower and lower performance per area and per power with the increase of the IPC. Low performance per area and per power means low multithreaded performance.

So the CPU cores with very good single-threaded performance, like Intel Lion Cove or Arm Cortex-X925 have very poor multi-threaded performance and using many of them in a CPU would be futile, because in the same limited area one could put many more small CPU cores, achieving a much higher total performance.

This is why such big CPU cores that are good for single-threaded applications must be paired with smaller CPU cores, like Intel Skymont or Arm Cortex-X4, in order to obtain a good multi-threaded performance.

Writing the program as a single thread is easy and of course it should always be done so when the achieved performance is good enough on the current big superscalar CPU cores.

On the other hand, whenever the performance is not sufficient, there is no way to increase it a lot otherwise than by decomposing the work that must be done into multiple concurrent activities.

The easy case is that of iterations, which frequently provide large amounts of work that can be done concurrently. Moreover, with iterations there are many tools that can create concurrent threads automatically, like OpenMP or NVIDIA CUDA.

Where there are no iterations, one may need to do much more work to identify the dependencies between activities, in order to determine which may be executed concurrently, because they do not have functional dependencies between them.

However, when an entire program consists of a single chain of dependent instructions, which may happen e.g. when computing certain kinds of hash functions over a file, you are doomed. There is no way to increase the performance of that program.

Nevertheless even in such cases one can question whether the specification of the program is truly what the end user needs. For instance, when computing a hash over a file, the actual goal is normally not the computation of the hash, but to verify whether the file is the same as another (where the other file may be a past version of the same file, to detect modification, or an apparently distinct file coming from another source, when deduplication is desired). In such cases, it does not really matter which hash function is used, so it may be acceptable to replace the hash algorithm with another that allows concurrent computation, solving the performance problem.

Similar reformulations of the problem that must be solved may help in other cases where initially it appears that it is not possible to decompose the workload into concurrent tasks.