'Cores' are really overstated in GPUs. CUDA cores are really SIMD lanes and if you counted it the same way as a CPU does, you'd get somewhere in the dozens of cores range even for modern GPUs.

That seems backwards to me. Sure, a GPU core is less general, but in terms of concurrent execution, memory bandwidth, and FLOPS I would expect hundreds to thousands of cores for all new GPU offerings. Apple’s double-digit GPU core counts for instance sound extremely understated.

It's not. The best comparison is the SM count for Nvidia hardware, or the wavefront count for AMD hardware. So a 4070 has 46 cores as you'd count them on a CPU.

If you have a truck that can carry 64 boxes, then according to Nvidia marketing each box would be considered its own core even though the boxes all go the same way.

Meanwhile if you had 64 passengers cars, then each of them would be able to drive a different route at the same time and if you counted the seats you would actually end up with 320 "Nvidia cores".

A proper method is counting ALUs instead of vague "cores".

If you want to count "cores" in a GPU, the best equivalency is streaming multiprocessors.

While it's true that they're in the dozens, strictly speaking, we've got GPUs with well over 100 now (the RTX 4090, for instance) and 64+ is a pretty common configuration.

From the abstract: "The eGPU architecture is a streaming multiprocessor (SM) machine with 512 threads. Each SM contains 16 scalar processors (SP)."