Tegra K1 (Denver) and Nexus 9 benchmarked
- D
- AnonD-294462
- x7v
- 27 Dec 2014
Returned my Nexus9 two days ago. Reason? Lousy performance.
Me and many tech reviewers have been wondering: "why do some trivial tasks like app launching/switching take so damn long on the Nexus9?"
We all initially thought that it's due to the new Lollipop that still needs some performance tweaks, and this issue will be gone with some minor updates.
Sadly no, it is an inherent problem related to the Denver design's nature: Unlike Apple's A8, Denver isn't a native ARM design, but it has its own instruction sets, and REQUIRES some sort of binary translation at runtime.
Remember Transmeta's Crusoe and Efficeon that plagued the UMPC landscape with lackluster performance some time ago? nVidia acquired Transmeta and Denver is the result.
The problem? It only excels at benchmarks where the same routines are executed over and over. In real world applications, the translation overhead causes huge hiccups when largely different code chunks are to be executed, app launching/switching, scene changing within a game etc.
Oh man that sucks. You should avoid Denver at all costs.
And nVidia should abandon Denver ASAP, unless they want to follow Transmeta into Oblivion.
http://en.wikipedia.org/wiki/Project_Denver
- D
- AnonD-153078
- mpc
- 09 Nov 2014
Anonymous, 17 Oct 2014Why is the nVIDIA Shield faster in the multi-core benchmark? Bigger heatsink probably.
- D
- AnonD-153078
- mpc
- 09 Nov 2014
Anonymous, 17 Oct 2014Why is the nVIDIA Shield faster in the multi-core benchmark? Bigger heatsink.
- D
- AnonD-153078
- mpc
- 09 Nov 2014
AnonD-153078, 18 Oct 2014Cost of course is another important consideration. The cost... moreThe big thing to remember about the 64 bit Tegra K1 Denver is that it built on 28nm planar process.
The Apple A8 ARM processor is which is built on 20nm FinFET process - planar to FinFET is one generational jump, and 28nm to 20nm is another, but Apple's A8 chip is still slower than the Tegra K1, and Apple had to add an extra core to put the transistor count up to three billion as a go "faster stripe" in order to earn benchmark bragging rights over the Tegra K1 Denver. The problem is that while adding more cores may increase the multicore benchmark scores, it doesn't improve UI responsiveness of speed on tablets, smartphones or desktops very much. That is the reason Apple made the A8 dual core instead of quad core like other ARM SoC manufacturers. The Tegra K1 Denver beats the A8 dual core and A8X three core processors on single and dual core performance.
The Intel Haswell/Baytrail is 22nm FinFET dual core processor, which is at the same process level as the Apple A8 and two generational steps ahead of the Tegra K1 in terms of process, but is slower than both.
The Intel core M is three generational steps ahead of the Tegra K1 Denver 28nm -> 22nm -> 14nm and planar -> FinFET, but it isn't going to be able to compete with either the A8/A8X or Tegra K1 Denver if it only offers a small increase in performance at low power, because it is going to be a whole lot more expensive. It is going to be a Windows laptop/high end hybrid processor.
Tegra manages to gain its impressive performance through clever Architectural design. Basically it uses a much cheaper 28nm planar fabrication process which does not require double patterning or 3D transistors, and achieves competitive low power and high performance by reducing the number of transistors and replacing the out of order execution hardware required for good single core performance with on the fly instruction recoding hardware like Transmeta tried, but the re-coded instructions are stored in a cache for re-use rather than used once only as Transmeta did.
What does this mean in real life? It means that the 28nm Tegra K1 Denver matches the 20nm Apple A8X and beats the Intel Baytrail processor, but will be a lot cheaper than both (even taking into account Intel's massive subsidies on its Bay Trail processors in order to gain a foothold in the mobile market). nVidia is like other ARM manufacturers are skipping 20nm to go straight 14nm. It also means the Tegra K1 will trounce the A8X when both move to 14nm and the processes are on par, and will trounce the Intel Core M and Atom based Silver Trail (also on 14nm) on low power performance, when SoCs become available in volume in Q2 2015. The 28nm Tegra K1 Denver will remain competitive in low cost devices even after the 14nm SoCs are in production because 28nm requires only single patterning and 2D transistors and cannot be beaten on performance/price by the 20nm or 14nm processors.
- D
- AnonD-153078
- mpc
- 18 Oct 2014
Cost of course is another important consideration. The cost of this will be a fraction of the Intel Core M processor - with similar performance or slightly lower power in low power configurations.
- D
- AnonD-259899
- y0B
- 17 Oct 2014
Anonymous, 16 Oct 2014The new exynos (galaxy alpha) is faster than either in mult... moreOutright speed matters less it's how much you deliver per watt. Perf/watt Qualcomm rules because it scales from a small phone to a larger form factor like a tablet where the wattage can be as much as double than what a smartphone can handle in terms of dissipation effectively.
Plus Qualcomm has the superior LTE solution on-die hands down. This is the reason for their over all success. It's the whole package SW/HW that forms an SoC which intern forms the MDP for ISVs.
- D
- AnonD-259899
- y0B
- 17 Oct 2014
Dave Granger, 17 Oct 2014Meh. We'll see how it compares to the upcoming Core M CPUs.Not at all, the TDPs claims are fishy at best. Another spin at SDP. Plus at Intel's pricing they cant sustain subsidizing the chips to Chinese white box tablets indefinitely.
Bottom line Intel has nothing on ARM/Qualcomm/Mediatek/AMD or any other arm based ISA players. Not even by a long shot.
- D
- Dave Granger
- mpx
- 17 Oct 2014
Meh. We'll see how it compares to the upcoming Core M CPUs.
- S
- Sync
- KZ8
- 17 Oct 2014
Kettzy, 17 Oct 2014True! But you can't move it around with you, can you? :)It just a cpu
- S
- Sync
- KZ8
- 17 Oct 2014
Kettzy, 17 Oct 2014True! But you can't move it around with you, can you? :)Yes.i always can carry my i7 4970k
- D
- AnonD-320698
- mUn
- 17 Oct 2014
Too bad the Mipad released this few months ago.
- ?
- Anonymous
- 2mC
- 17 Oct 2014
Anonymous, 16 Oct 2014Actually, the S805 is more powerful than the A8. Have you s... moreThe Nexus9 kicked samsung eynos in the A.....ss.
Tegra K1 really power beast.
- K
- Kettzy
- bER
- 17 Oct 2014
Sync, 17 Oct 2014Guess what?my intel i7 4970k is far better than this oneTrue! But you can't move it around with you, can you? :)
- u
- user
- dQT
- 17 Oct 2014
Anonymous, 17 Oct 2014Why is the nVIDIA Shield faster in the multi-core benchmark? becouse shield has 4 cores and nexus has 2 cores. Single from nexus is better but quad is more powerful
- ?
- Anonymous
- uEM
- 17 Oct 2014
Why is the nVIDIA Shield faster in the multi-core benchmark?
- D
- AnonD-143076
- KFP
- 17 Oct 2014
HTC u shouldnt forget ur old customers and u cant just raise ur hands up by saying that nvidia isn't supporting you guys if it was so then how could u guys make nexus 9 with a tegraa processor you need tooo do something for Htc one x as well as Htc one x+ u should force them to support for hox+ & hox. Guys I tell you that Htc and nvidia have problem updating their old line up which I have mentioned above so I recommend don't buy this device I own a hox+
- S
- Sync
- KZ8
- 17 Oct 2014
Guess what?my intel i7 4970k is far better than this one
- ?
- Anonymous
- IHb
- 17 Oct 2014
AnonD-171579, 17 Oct 2014Is this benchmark from the Nexus 9 running Lollapalooza?Indeed, but the app isn't optimized for lollapalooza (lollipop) yet, so this can be seen more as a reference. The single-threaded score should be in the 2,000 mark once optimized.
- D
- AnonD-171579
- 0Ja
- 17 Oct 2014
Is this benchmark from the Nexus 9 running Lollapalooza?
- ?
- Anonymous
- IHb
- 17 Oct 2014
AnonD-257092, 17 Oct 2014its already at 64 bitHe's referring to the app being optimized for 64 bit, more than likely.