You don't seem to understand just how GPU dependent the benchmark really is. Most games on ARM devices are generated at 720p (or less) and then upscaled. 720p is a HUGE resolution for GPUs that weak (for example, the E-350 has problems with most PC games over 800x600).
As proof, let's go back to the Anandtech spec you posted above. Let's compare them with anands other z2760 and E-350 benchmarks.
You can see a more extensive comparison at notebookcheck
In all the different CPU tests (anand's and notebookcheck's), the E-350 is NEVER even 2x faster. If ice storm were only dependent on a CPU, then we would expect to see similarly close scores, but instead we see a difference of more than 10x. For the atom to be that much slower, it would need to be about the same speed as a 1997 233MHz pentium II (E-350 GFLOPS source Pentium II-233 GFLOPS source). I think you would agree that this is not the case. Since the Z2760 and E-350 devices were running windows and are x86 and are using directX and the relative performance is established, the only remaining factor is relative GPU power. The z2760 has an Imagination SGX545 GPU clocked at 533MHz (Intel.com). According to Anandtech, the SGX545 has 4.8GFLOPS at 300MHZ (source) which gives the intel chip 8.528GFLOPS of power (4.8GFLOPS / 300MHz to get GFLOPS per MHz and then multiply by 533MHz for final answer). The E-350's GPU (HD6310) offers 80GFLOPS of power (source). Since both are VLIW architectures (source), we can conclude that real processing power should be similar for the same situation (read about the VLIW real power problem here). It's interesting that the available GPU power differs by almost exactly 10x. This seems to indicate where the performance difference could come from.
Similarly, the Exynos 5250 CPU (an A15 clocked at 1.7GHz) is known to be faster than the Snapdragon 600 (see anand benches here) which is clocked at 1.9GHz, but is significantly slower per clock in many floating point operations (which are most important to physics -- the biggest features of the past couple of bullet physics updates were adding NEON and ARM FPU support). It is also known that the Mali T604 in the 5250 is a little slower than the Adreno 320 (see below) in non-theoretical benches (probably linked to bandwidth constraints in the exynos). How can you account for these known factors (using many other tests both synthetic and real)?
This seems to come down to "the company said so", but that is a non-answer.
Edited by hajile - 7/9/13 at 6:07am