Honestly, one of the most impressive Geekerwan videos in a while. They go in depth comparing nodes from TSMC N3(P) and Intel 18A cell libraries, gate pitch, transistor density you name it. They also compare TSMC's PoP packaging with Samsung's new HPB packaging.

Kurnal has a site listing logic transistor density if you want to see: https://kurnal-insights.com/calculators/logic/

Onto the Exynos2600. While some of this is repeated from this post before https://www.reddit.com/r/hardware/comments/1sxgkc8/chipwise_reptalicant_annotated_die_shot_of/ I'll just reiterate some of Geekerwan's analysis.

With an area of 141.5mm2, it's one of the largest SoC compared to the latest and that isn't even including the modem (which is external, measuring 34.5mm2 on SF4) which would make it \~160mm2 as large as an Apple M2. The NPU and GPU take most of the die's area, where the RDNA4 Xclipse 960 takes up 23% of the die, 46% larger than 8EG5's GPU.

Onto Spec2017, the E2600's C1-Ultra isn't that far behind the D9500 C1U and 8EG5 L core. In INT fairly same performance as the 9500 and in FP just a bit behind it. L core beats both but not by a wide margin. Geekerwan states SF2 is a useable node, just as decent as the rest.

As for GB6 (which was discussed here: https://www.reddit.com/r/hardware/comments/1t2cs6q/samsung_exynos_2600_geekbench_6_multi_core_power/ ) , with E2600's core layout, the perf/power curve is just as good if not just beats 8EG5 while easily leading D9500.

Onto the GPU which hasn't really discussed before, this one is really interesting as for what Samsung has cut out from the RDNA4 IP. No WMAA units, can't run BF16/FP8/INT8 ops, seemingly would not be able to effectively run an FP8 FSR4 upscaler. FP32 ALU's are said to be larger than Strix Point's 890M iGPU with the same 16CUs. This is because each CU has 224 FP32 ALUs which is in contrast to modern RDNA IPs which have 64 (starting from RDNA2 dual issued VOPD instructions so 128 if taken advantage of).

With all the compute they do argue against the large GPU since the bandwidth is only as large as other mobile SoCs and smaller than laptops as well as the small of amount of L2 it has (similar to the 890M) which hits a wall at higher freq. In 3DMark SNL bench the D9500 GPU leads while also being \~26% smaller. Comparing it to AMD's own mobile SoCs it easily beats the 780M and is not too far behind the 890M.

As for the remaining bits, with real gaming, with all of it's cores, scheduling can be really rough. It expectedly throttles over 10W and settles down, seems like HPB does some good in this case. In their comprehensive 5G battery test, 13% worse than the 8EG5, but they believe the external Samsung modem is partly to blame.

Overall Geekerwan found the E2600 to be rather great despite it's faults. SF2 is good on par with the current rest, HPB is an interesting alternative to PoP, the CPU core layout is designed rather well. Though, in practice scheduling is horrid, the GPU (and I suppose NPU) is designed too large as well as the mediocre modem performance.