The fastest Tegra 3 GPU has an operating frequency of 520MHz, while the fastest Tegra 2 GPU has an operating frequency of 400MHz.
It was 333MHz last time I checked. If there's a 400MHz T2 I stand corrected for the PS floating point difference.
So the difference in peak pixel shader performance will be 2*(520/400) = 2.60x difference. Kishonti's GLBench Fill Rate data referenced above gives 2.58x difference. So there does appear to be a reasonably good correlation. Not perfect, but reasonably good as an approximation. You are right though that the benchmark is measuring MTexels/s and not MPixels/s.
I said it before it would be high time that Kishonti considers registering GPU frequencies just as CPU frequencies wherever possible for its database for all SoCs.
Why would anyone care about theoretical GFLOP throughput for a GPU used in a handheld (smartphone, tablet, clamshell) device? These products are not geared towards High Performance Computing. And besides, as we have discussed before, differences in GFLOP throughput typically does not correlate well with differences in gaming performance, particularly when comparing totally different GPU architectures
Because the majority is counting in GFLOPs these days otherwise the claimed performance increases will never materialize. Have a second look at NV's own future Tegra roadmap and tell me where the huge (granted per SoC) increases are supposed to come from.
It still stands though that you can't that easily guess on fillrates without knowing or have a reasonable guess on the possible architectural layout.
As a current case example the Mali400MP4 in the 32nm Exynos4 runs in current products at 440MHz, which gives it (pixel and vertex shader ALUs combined) ~15.8 GFLOPs with a texel fillrate of 1.76 GTexels/s. Upcoming next generation GPU T604MP4@500MHz is at 72 GFLOPs with a texel fillrate of 2.0 GTexels. ARM itself claims on its own website a 5x times performance increase for the T604 compared to the former generation. I'm sure you wouldn't suggest that that increase is supposed to mean texel fillrates do you?
IMG claims >20x times the performance increase, but the tricky part of creative marketing here is that they most likely compare a DX10.1 SGX545 with 4 Vec2 ALUs against a DX10.1 four cluster GC6400 for a comparable die area. Considering a SGX545@640MHz is at barely 10.24 GFLOPs and the GC6400 at >210 GFLOPs; remains to be seen if die are ends up being comparable under 28nm for the latter, which is a tough cookie for me to swallow.
Back to Adreno320 I believe and would like to stand corrected that it consists of 4*SIMD16 (or 64SPs in desktop marketing parlance) where each cluster has a single TMU, giving 4 TMUs for a 4 cluster 320. Here compared to Adreno225 (ignoring architectural differences and based only on sterile unit counts) its 8 Vec4 ALUs account for 32SPs accompanied by 2 TMUs. Both should be clocked at 400MHz based on paper specs at least there's nothing more to claim than a factor 2x. Higher performance comes IMHO from the architectural advancements of the Adreno3xx family.
The only other problem is that NV manages to stay highly competitive with Adreno225 GPUs (especially for the lower complexity shader stuff) with barely 12SPs@520MHz vs. 32SPs@400MHz. NV doesn't have any magic wand nor isn't designing its hw with pixie dust, it just shows how damn efficient their driver/compiler really is.
In hindsight I'm not surprised to read ever repeating idiotic rumors that Qualcomm supposedly has licensed GPU IP from IMG or that they are interested in buying AMD's graphics department or whatever else, since different internet rumor mongerers try to find a feasable "solution" for Qualcomm's GPU roadmap for >DX9L3 and stuff like windowsRT drivers. I wouldn't suggest that Qualcomm needs as radical solutions as proposed left and right. IMHO if Qualcomm should be truly interested f.e. in AMD's GDP it would be rather because they have far more ambitious future plans than just small form factor SoCs and not to solve some silly driver/compiler issue which is (always IMHO) a resource problem money can buy. In that regard Qualcomm is anything but stingy when it comes to constantly hiring engineering talent of all kinds; au contraire.
