From Anandtech:
There's also
this post from OlegSH. Though if we can summon
@AlexV , he'll probably take away all the doubts.
We've had that FP16/PSNR rubbish- throw everything into one pot again in similar discussions here in the forum. For the record the benefits of having FP16 are bigger in terms of power consumption and less in terms of performance in these cases IMHO.
Allwinner A80, G6230@533MHz (64 FP32 SPs, 96 FP16 SPs)
TRex offscreen: 20,60 fps
Manhattan 3.0: 8,60 fps
Manhattan 3.1: 3,90 fps
Mediatek HelioX10T, G6200@700MHz (64 FP32 SPs)
TRex offscreen: 27,10 fps
Manhattan 3.0: 10,20 fps
Manhattan 3.1: 4,90 fps
Care to show me what I am missing and the 6230 isn't at least on par with the 6200 despite the >30% higher frequency?
I'm not sure what you're disagreeing with. Perhaps you misread my post?
I wasn't comparing the X1 to K1's GPU. I simply stated that
a single ALU in X1 can either do one FP32 operation or 2*FP16 operations, as long as it's the same operation:
Here's your exact quote:
Tegra X1's Maxwell 2.5 does twice the FP16 throughput than it does with FP32
I thought you said "2.5x".
"Way behind" as in the A9X is definitely closer to 50% the sustained performance of a PC with a HD7770 or Xbone than it is from 100%.
According to this picture, IMG themselves would consider ipad pro's 12-cluster Series 7XT to have the equivalent performance of a Geforce GT 730M:
GT 730M uses a GK208 with 2*SMX. 32 TMUs, 8 ROPs at 700MHz for ~550 GFLOP/s.
Both the 1GHz HD7770 and the Xbone top 1.3 FLOP/s with over twice the fillrate of a GT 730M.
I'd call that
way behind.
Hold it; here's the problem: I'm having the XBox360 in mind and you mean with "XBone" the XBoxOne. If you'd follow my crap through the years you'd know that I avoid comparisons like that unless I insert a footnote that it's a DX10 vs. a DX11 unit. Look above I mentioned Xenos/C1 and that in terms of compliance/functionality is the closest console GPU you can get to the A9X GPU.
Ignore the marketing rubbish of that diagram; I think but am not sure that the GT7900 comes with DX11.2. Despite it's 16 clusters and let's say clocked at 600MHz it could see cases where it embarasses the XBoxOne GPU, but even there I wouldn't believe in a true all around winner even if I'd see it. For that you'd rather need IMHO an unannounced 8XT 16 cluster config @DX11.2 and again a frequency not lower than 600MHz. DX11.x needs significantly more die area compared to vanilla DX10.0, which sounds impossible that it will not impact temperatures, power consumption, throttling etc.
----------------------------------------------------------------------------------------------------------------------------------------------------------------------------
Outside the above: who said that the A9X doesn't throttle at all? I find it impossible that even Apple's GPUs don't throttle in the worst case up to 20% under constant heavy usage. If anyone should look at long term performance results in Gfxbench, please look again since it's running TRex onscreen meaning vsynced at 60Hz. When offscreen scores are in TRex as high as on the A9X it's fairly impossible to detect or define any possible throttling or lack thereof. Kishonti seems to have now a Manhattan 3.1 long term performance test which should solve the former headache for some time.