Yes if for instance tegra 4 had advanced api support..like open gl es 3.0...then that would give it a massive advantage over every other soc....games really would look from a different class.
Nothing against that; however the indication of a synthetic benchmark which intends to "foresee" up to some degree the future isn't entirely worthless either.
It didn't strike you that you're comparing above vector vs. "scalar" ALUs? In my comparision above all compared GPUs have in fact vector ALUs, irrelevant if they have USC ALUs or not. The majority if not all IHVs involved in GPU development for the small form factor market will have IMHO "scalar" ALUs for the coming Halti generation of GPUs. And to get back to the initial point: if the highest variant of Wayne has "only" 64 GFLOPs maximum, while in comparison IMG with it's GPU IP can range at the moment from 100GFLOPs up to 1 TFLOP there's no efficiency magic wand possible to close that gap and that's exactly the reason why 64 GFLOPs don't sound like all that much. And no the whole enchilada doesn't come with just sterile GFLOPs comparisons; it'll get even more hairy when you start comparing achievable real time triangle rates, texel/pixel/Z fillrates and what not especially for the latter on a TBDR.
Vsync ahoi.
Yes of course would it be more fruitful to see real games benchmarks, but I don't think you'd find all that many mobile games out there where the MP4 in the iPad3 wouldn't trounce by a significant margin any ULP GF in T3.
Vsync explains part of the difference, but memory bandwith explains part of the difference too. You do realize that the improvement in memory bandwith in iphone 4s (with SGX543MP2) vs. iphone 4 (with SGX535) is 2x, even though GFLOP throughput increases by 8x?No it mostly doesn't.Ahoi no
Not all configurations of Rogue are necessarily above 100 GFLOPS. It's a scalable architecture
While I obviously agree that real world gaming benchmarks are a very good thing and GFLOPS are only one of many important metrics, it's still an important number. The difference between GFLOPS and 'cores' is that the former does require a certain minimum amount of silicon to implement (even if you need more silicon to be able to use it efficiently, e.g. single issue vs co-issue) while the latter has no real definition and therefore no real cost.
Well, what else is he going to say? I'm also pretty sure that the performance of SGX devices in GLBenchmark 2.5 disproves his claim that Tegra is better at higher levels of geometric complexity...
Not all configurations of Rogue are necessarily above 100 GFLOPS. It's a scalable architecture
Depends on the workload I guess, it's likely that an AAA dev like Epic would optimise very heavily towards the actual platform so in the case of the SGX535 that might mean as few ALU operations as possible, and very importantly using LowP (10-bit integer) for most pixel shaders. LowP isn't actually any faster on SGX543 compared to the original SGX, so SGX543MP2 would only have 4x as much ALU performance rather than 8x.
I can't speak for TKS, but he probably only meant that the first variant would exceed 100 GFLOPS.
Sure, I'd much rather buy a Tegra 3 tablet than most alternatives in today's market, it's a solid overall platform and the exclusive games are a very nice bonus if you're personally interested in a specific one.
It does highlight that running the same content n times faster isn't the point, though. Who wants to run Doom at a million fps?
Same average real-world target power, yes, but the power at peak or under heavy load is significantly higher.
Disagreed on the first point, agreed on the second. They could have undervolted the GPU and achieved even lower power consumption. There is a fundamental trade-off between area and power for parallel processors (see one of my favourite presentations: http://www2.lirmm.fr/arith18/slides/ARITH18_keynote-Knowles.pdf) and I think it's clear that NVIDIA was more cost conscious than some competitors with both Tegra 2 and Tegra 3. That has relatively little to do with the kind of end-device (e.g. $199 Nexus 7 vs $499 HTC One X) and more to do with strategic and financial imperatives IMO. Also remember that a faster GPU would have required a more expensive 64-bit memory bus given that Tegra is a classic IMR without framebuffer compression.
Do past launches also include the T3 related delays? Neither NVIDIA or anyone else is immun from any possible delays. By the way "we will see it" stands for a SoC announcement or final device availability?
Do past launches also include the T3 related delays? Neither NVIDIA or anyone else is immun from any possible delays. By the way "we will see it" stands for a SoC announcement or final device availability?
)
