Brilliantdeve
Newcomer
NVIDIA Tegra Architecture
- Thread starter french toast
- Start date
-
- Tags
- nvidia
The numbers are what they are. All we have seen so far is the GPU performance per watt for Logan. It would be pretty illogical to suggest that the GPU performance per watt for Logan is going to go magically down between now and early next year (and equally illogical to suggest that the number of CUDA "cores" is going to go down)! Anyway, in a power constrained environment, performance per watt = performance provided that there is performance headroom to scale. Obviously Kepler has lots of performance headroom to scale.
Think logically about what you are saying. Adreno 330 in S800 has ~ 50% more GFLOPS throughput than the ULP GPU in Tegra 4, in addition to a unified shader architecture, and yet it only has ~ 20% faster graphics performance (comparing tablets to tablets). That actually hurts your case because the Adreno GPU architectural efficiency is worse than Tegra 4.
AMD was unable to make the same investments in ultra low powered processors as NVIDIA, Intel, Qualcomm, etc have done. As a result, they will not be a major player in the ultra low power mobile space.
And just FYI, it is "Kepler", not "Keplar". I have seen so many people across forums misspell that. "Kepler" is named after a scientist. "Keplar" doesn't exist.
Brilliantdeve said:
Clearly we will see 20nm products by second half of 2014, but the process node maturity will not be spectacularly high. So approaching or exceeding the peak performance (ie. close to 400 GFLOPS throughput) of ultra mobile Kepler will be no easy task at that time.
Last edited by a moderator:
Brilliantdeve
Newcomer
What I meant is that the ULP Tegra 4 GPU has higher FPS per GFLOP than the Adreno 330 GPU. Even at a 50% disadvantage in terms of GFLOP throughput, it is "only" at a 20% disadvantage in terms of fps (comparing shipping tablets).
Edit: Redundant... ams already answered, although it's not a useful comparison of architectural efficiency without knowing power/thermal efficiency as well.
ams was considering how much performance was being delivered (as measured by a bench like Gfxbench 2.7) compared to how many GFLOPs are supposedly available in the design and then comparing an Adreno 330 product like the Galaxy S4 to the nVidia Shield based upon those ratios.
ams was considering how much performance was being delivered (as measured by a bench like Gfxbench 2.7) compared to how many GFLOPs are supposedly available in the design and then comparing an Adreno 330 product like the Galaxy S4 to the nVidia Shield based upon those ratios.
Are you talking about before or after the 2nd run of 3dmark when T4's performance plummets due to throttling?
The important statistic is perf/W, and it's quite obvious to anyone (with the sole exception of you no doubt) who the clear leader is. What's more, they have actual shipping products.
Whether Shield or a Tegra tablet was used for the calculation, it doesn't change the outcome of which architecture is supposedly delivering more benchmark performance per GFLOP (assuming the GFLOPs figures being purported are actually accurate in the first place), and it still doesn't meaningfully account for power/thermal efficiency to truly allow for a comparison of architectural efficiency.
Brilliantdeve
Newcomer
You mean after 10 looped runs of 3dmark where the CPU was throttling more heavily than the GPU?
What do you think would happen to most high performance quad-core ultra mobile SoC's when subjected to the same testing procedure? Anyway, we have no data on the performance per watt of the Tegra 4 ULP GPU nor the Adreno 330 GPU, so there is nothing definitive to say about that.No all we've seen is Nvidia's usual nonsense with a cherry picked benchmark.
Yes, that's why T4 can't get any traction in tablets while Qualcomm has all the big design wins.
You mean AMD realised the stupidity of trying to compete with Qualcomm and didn't feel like making the same losses as Nvidia. How much has Tegra cost Nvidia now? It must be closing in on half a billion dollars - probably closer to a billion by the years end.
Anything that is 6" or larger is clearly a phablet or tablet. There will be lots of these devices with S800. Tegra 4 will have some too (like HP Slate 8).
LOL. Investing in ultra low power processors was a very smart move by Intel, Nvidia, etc, and you will start to see why starting later this year or early next year. As for Tegra, IIRC, the incremental investment cost is ~ $400 million for ths Fiscal Year 2014, so Tegra as a whole will actually be profitable this year (because Tegra revenues as a whole are projected to be ~ $500 million).
Brilliantdeve
Newcomer
Adreno 300 is way better than Tegra 4 theoretically .. but where Tegra 4 might shine is CPU because 4x Cortex-A15 !
it'll be interesting to compare next coming product ..
it'll be interesting to compare next coming product ..
Brilliantdeve
Newcomer
Differences in theoretical GFLOP throughput tend not to correlate very well with differences in gaming performance (ie. FPS) when comparing different GPU architectures. Cortex A15 is clearly quite a bit faster than Krait 400 on a clock-for-clock basis, but in a power-constrained environment there is usually not a very big difference. If anything, the power hungry nature of Cortex A15 at peak performance is in some ways it's achilles heel, so ARM and it's licensees will really have to work on that moving forward.
You see, you completely lack foresight. To be a significant player in the ultra mobile space requires an ultra low power GPU, an ultra low power CPU, and a 4G LTE baseband modem that can be integrated on die when needed. That is no easy nor quick task, no matter how much you trivialize it, and that is why it has taken companies such as NVIDIA and Intel years to get there.
Similar threads
