Lilbits (5-15-2014): What to expect from the NVIDIA Tegra K1

The first tablet with an NVIDIA Tegra K1 processor is on the way, and there’s a chance we could see more soon. So what kind of performance can we expect from the platform? If early benchmark results are anything to go by, it really is one of the fastest ARM-based chips to date.

Phoronix reports that the processor is comparable to an Intel Celeron J1900 Bay Trail chip when it comes to Linux-based CPU performance tests. And Chinese site Evolife reports that the Tegra K1 outperforms the Tegra 4 and Qualcomm Snapdragon 801 in Android tests.

tegra k1 bench

Here’s a roundup of tech news from around the web.

You can keep up on the latest news by following Liliputing on Facebook, Google+ and Twitter.

  • leafar

    Tegra 4 ,nvidia shield marvelous, K1 simple amazing, no comment

  • ChicagoBob

    Compare k1 to Atom cpus

    • CyberGusa

      Well, keep in mind that for Nvidia the early benchmarks are not power limited like they would be in final tablets… So it’s the kind of performance you’d only see in a desktop like system… But Phoronix compared the results they’re seeing to the performance of a Bay Trail D Celeron J1900 ATOM, which is pretty good for a upper limit but we’ve yet to see how much performance it’ll actually have in a mobile configuration…

      • HasithaP

        thanks, your comments are always useful. keep it up.

      • Abram Carroll

        No they are still doing 5 watt TDP in those tests and they have benched it on batteries.
        Board quality(needs to handle spikes to 11 watts) and heat dissipation could effect final product, just like with their competitors.
        Id want a battery close to 7000 mah though as it has a 5 watt TDP.

      • CyberGusa

        Sorry but in actual mobile tablets they can’t let it even spike to 11W… Mobile tablets have to remain fan-less and that means keeping the SoC operating at around 5W or less at least 99.9% of the time!

        Really, nearly everyone in the mobile tablet market can provide more performance than they do now but have to throttle performance to keep their products fan-less… not to mention keep the battery life from getting too short either…

        This is why they develop power efficiency methods like balancing power usage between the CPU cores and the GPU… So either the CPU Cores or the GPU can use the full power limit but not both at the same time and thus never exceeds the limit… Thus performance varies depending on whether it relies fully on one or the other or needs to share lower performance with both running…

        Besides, mobile devices also need smaller components and don’t get options like fans and larger heat sinks…

        All while needing to work in very confining space that also shares space with the battery, which can’t get too hot, and screen, which also generates heat…

        http://media.bestofmicro.com/nvidia-gtc-2014,O-N-428567-3.jpg

        The actual performance tests for the Jetson are coming with it either with a fan, as that image shows, or a large heat sink that nearly covers the whole top of the board… While it never has to deal with being next to a screen or having its power internalized…

        The up to 11W spike is also for only GPU + CPU (+ memory and IO) but in final device you have to account for the screen and a more limiting environment!

        Remember also that the present K1 isn’t the final version but is basically just an updated Tegra 4 as it’s still using the same CPU Cores and mainly just updated the GPU…

        So you can expect a big graphical boost but don’t expect much of an improvement in CPU performance with this present version… At least until the 64bit Dual Core version of the K1 comes out later and then we can see a CPU performance increase like we saw with Apple’s 64bit A7… Maybe more, since Nvidia won’t limit RAM and the K1 platform supports up to 8GB max… and it looks like the mobile market is slowly set to increase from 2GB max average to 4GB over the next year as these new products start coming out to market and we start seeing more and more 64bit products released…

      • Abram Carroll

        Great post CyberGusa, TY.

        5 watt TDP is quite standard for tablets and their max draw can hit 15 watts total for the device. Assuming no improvements in build quality and large inefficient screens this could hit 19 watt spikes for the device Even though NV Shield is actively cooled I’d point to it in talking about heat and build quality. Shield doesn’t throttle when pushed to the limits using 3.6 to 8.4 watts for the device and this includes the fan and screen. I get about 6 hour battery life pushing it hard, 10 hours streaming and close to 12 with light browsing. It’s very much on the low end for power draw considering 14 watt max draw products and high in terms of battery life. I’d say that overall build quality and components is a big factor in devices as are screens. The ability to hold heat from spikes is important, as is how much heat other components produce. Low build quality can dramatically change this. Active dissipates more heat leaving headroom or a larger heatsink(weight issue) holds more. The TDP is simply how much it needs to dissipate, and doesn’t account for how much it needs to hold as TDP assumes average output as a constant. When in reality it’s variable and you actually need to handle more. Although I don’t have a stack of tablets, throttling is a common and widespread complaint I see for every device I look at.

        The form factor is a big challenge, but it can be overcome. Cost however is a big factor. You could for example use a 1.25 amp battery, however that is going to cook the tablet. K1 was shown with a 5 amp battery and that will stay very cool. Also SOC heat.

        How is that going to get out? Through the battery, screen or case(conductive metal or a plastic blanket)? The TDP doesn’t matter if the heat is trapped and it’s hot boxed. It will throttle regardless and TDP will simply effect the time it takes to do that. The boards power system is also important, as less robust system will get very hot. Screens seem to be a bigger factor than the SOC. Power draw is all over the place and the larger the screen and resolution the worse it gets.

        K1 clearly requires a more expensive tablet to hit it’s numbers.

        On paper Apple tablets are turds in everything but single thread, storage, memory bandwidth, and pixel fill rate. Yet they have a good build quality and in reality can hold their own against actual products way better than the number would suggest.

        Seeing Tegra K1 stamped on a product doesn’t tell you about the device build quality. Look for testing and in depth reviews. I’d expect benchmarks to be all over the place without QC from NV.

        Nvidia’s reference tablet hits it’s numbers, but they don’t cut corners.

        http://i1.ytimg.com/vi/ZCCaqQyEIPI/maxresdefault.jpg

        Yes the first K1 versions are Tegra 4 CPU’s with a full Kepler desktop SMX. I’m also quite interested to see NV’s ARM v8(Denver) vs. Apple’s. Single thread preference matters, as multi-threading is usually poorly done. That is single thread code broken up with locks that lead to thread blocking and stutters. Some code simply can’t be broken up and lacks multi-tread alternatives.

      • CyberGusa

        Uh, sorry but you seem to be not understanding how severely limiting a fan-less design really is…

        Sure, build quality can help but the simple fact of the matter is passive cooling is way too inefficient and even a small slow fan can more than triple the thermal limit!

        Thus there’s no way a mobile device can be allowed to spike anywhere near 19W… Really, even a high powered screen would add maybe 2.5W and total system may peak at around 8W but that’s it for a fan-less design…

        The Nvidia Shield’s fan, btw, only uses up to 0.44A under max load and that’s a very low power draw and doesn’t add much at all to the total power usage…

        While anything else would either require active cooling or a really big heat sink and mobile devices are too thin for a big heat sink…

        Really, even ~12W TDP systems like the last gen of Cedar Trail netbooks (6.5W for a N2600, 2.5 for the rest of the system, and additional 2.5W for the screen) still needed either a fan or a large heat sink that would make it too bulky for the thin and light tablets common for mobile devices…

        If it wasn’t for the constant push for higher performance ARM SoC’s would all still be well below 5W for the entire device but most are pushing those limits and most of the development over the last couple of years was the industry figuring out multiple ways to do more with less power…

        Even the modern Intel ATOMs are pushing for ever improving power efficiency and operate similarly to keep everything operating within the strict limits allowable by mobile designs…

        Connected Standby, power gating, power load balancing, etc. are all ways to maximize idling and minimum power usages…

        Quad core Bay Trail uses up to about 2.2W max under load for all 4 CPU cores for example, and usually stays around 2.2W for SDP or average operating range and would rarely ever exceed 4W max and thus the entire system, screen and all, stays below 8W and usually less than 5W…

        And the reason why they’re still pushing dual cores is because running multiple cores also multiplies how much power is used and thus heat generated… So some limit design to dual cores to focus instead on per core performance and limit maximum loads…

        Limiting the Tegra 4 Note 7 to just 1GB also was not just for keeping cost low but also to keep the system operating within those same limits I’ve been stating… as more memory uses more power and lets the system work harder and is one of the reasons why there isn’t a 4GB option widely available yet…

        And the performance of the Note is below what the Nvidia Shield offers… So they obviously had to re-optimize it and limit it for a fan-less design for the Note…

        The only benchmark that didn’t take a noticeable hit, compared to the Shield, for the Note was the GPU benchmark and that’s because they’re prioritizing GPU performance but it remains they lowered performance overall for a fan-less design…

        They are making some improvements, which lets them push the limits a little further this time but the limits are still pretty much the same and we’ll have to wait for significant FAB and architectural improvements before it really changes much and those changes are still about a year away…

        So again, keep in mind that the present Tegra K1 is still using the same CPU cores as the Tegra 4 and is even still being made on the same 28nm FAB… There’s some performance tweaks but it’s not really realistic to expect they’ll offer much more than much better GPU performance right now for actual mobile devices designed around the present K1…

        4GB RAM options will also likely not be pushed until the 64bit dual core version of the K1 is released, further limiting what can be done with a present 32Bit K1 in a mobile device…

        While 8GB options, which the K1 is suppose to support, will only likely ever be seen in laptop or desktop implementations….

        None of this is to say the K1 isn’t a good improvement over the Tegra 4… the graphical boost alone justifies it… but we have to temper our expectations of what they could deliver for actual mobile devices…

        Though, being capable of higher performance does mean they could find applications in systems such as servers and actual desktops that don’t have to worry about mobile limitations…

      • Abram Carroll

        “Really, even a high powered screen would add maybe 2.5W and total system may peak at around 8W but that’s it for a fan-less design”

        This one hits 14.5 watts
        http://www.notebookcheck.net/Review-Nokia-Lumia-2520-Tablet.116320.0.html

        This one hits 16.6 watts
        http://www.notebookcheck.net/Review-Asus-Transformer-Pad-TF701T-Tablet.115868.0.html

        I’m just saying a poor design with poor build quality would be an issue with spikes to 11 watts. A lot of underclocking would need to be done to lower that.

        “So again, keep in mind that the present Tegra K1 is still using the same CPU cores as the Tegra 4 and is even still being made on the same 28nm FAB”

        ” but we have to temper our expectations of what they could deliver for actual mobile devices”

        I think the general consensus is that it’s going to stutter and frame dip it’s way over the line for most games. The line being games that don’t look like they are from the 90′s. Lots of optimization will be required in terms of draw call batching to not get dips and well you know…

        “4GB RAM options will also likely not be pushed until the 64bit dual core version of the K1 is released”

        First ARMv8 Tegra K1?
        NVIDIA Tegra K1 processor packed with 4GB of RAM, 128GB of storage
        https://www.google.com/atap/projecttango/#devices

        Also Nvidia just doubled their performance per watt on the same 28nm node with Maxwell. If you haven’t checked out the 750 ti.

        They should be able to go 256 cores and with a lower TDP or faster clock and 128 cores for phones. Without any power saving features 128 cores are pulling less than 5 watts on the desktop @ >1 Ghz

      • CyberGusa

        Yes, a poor design would make it worse but I was talking about what to realistically expect from the average tablet design and in the first link you should note…

        The rest of the competition consumes significantly less energy; all of the devices use less than 11 Watts. In fact, the Sony Xperia Z2 Tablet only needs 7.4 Watts.

        So it’s an obvious exception… The Sony Xperia Z2 is even considered a good build tablet and as indicated it operates below 8W!

        While in the second link, note things like the reason Asus typically markets the battery life with the use of the Keyboard Dock that uses a secondary battery because the battery life isn’t good enough with the battery in just the tablet by itself, despite putting a much larger than normal capacity with a large 47WHr (most are 30WHr or less), because it’s using too much power and in the mobile market it has to be at least close to what an iPad offers for run time or it’s considered poor!

        Most tablets don’t make use of any such docks and thus have to rely on what the internal battery can provide and smaller tablets can’t fit as large batteries as larger tablet… To show another factor you may not be considering…

        The batteries also add significantly to the weight and the keyboard dock and tablet combination was compared to something like the Apple MacBook Air, which may be light for a laptop, but tablets generally have to be as light as possible because unlike a laptop they need to be able to be held for prolonged periods of time and thus why anything over a pound is still considered heavy for tablets and this model is still heavier than a pound, unless they’re also very large but that’s one of the reasons why there is a differentiation between mobile tablets and tablets you only use at home or work, like you would a laptop…

        Those models that used above average power also made good use of metal with near all metal designs, meaning the casing was essentially a heat sink and thus a fairly large heat sink but most tablets are made of plastic and similar materials that aren’t very conductive of heat and it’ll take more than just build quality to compensate for that difference… and it gets only harder for smaller tablets that go down to 7″, like the Nvidia Note 7…

        Really, they aren’t developing things like docks with fans and scalable performance that increases with those docks if cooling wasn’t a big factor…

        The Core M tablet MS demonstrated recently for example was shown with a similar docking station that blows air in front and behind the tablet when docked and this allows performance to be increased by 30% while docked… and that’s without any internal active cooling… just boosting the normal convection cooling from the exterior…

        While some of the tricks designers use is for things like max clocks to only apply to a single core… allowing the rest to be turned off and thus act as heat sinks to better allow that one Core to operate at max performance longer… Sacrificing multiple thread performance in favor of single thread performance…

        Intel’s Turbo Boost system, just using it as an easily referable example, for example works by budgeting the max clock speeds under the thermal limit of the system… So only allow the max clock speeds up to the point the system starts over heating and then scales everything down until it cools down again…

        Similarly, most mobile SoCs have thermal limits set that when exceeded induces additional performance throttling to give the system a chance to cool down…

        The Nexus 5 is an example of a device that runs into this fairly often, being a phone makes the limit more extreme than on a larger tablet as well and thus a more noticeable limitation…

        http://www.anandtech.com/show/7517/google-nexus-5-review/4

        The reality is that no mobile device right now is exempt from the
        realities of thermal throttling or the thermal constraints of a small mobile devices, even tablets.

        So, even in the rare cases you do have a device that pushes above the normal limits it doesn’t mean it isn’t sacrificing something in order to make that possible…

        Yeah, they’re pushing the limits higher every generation but the rate of performance increase is growing faster than the rate they can improve power efficiency and improve thermal limits…

        Thus the need for docks that provide active cooling… which helps tap the full performance when really needed without sacrificing the mobile, fan-less, usage where performance is less needed…

        Keep in mind that unless they make a large tablet that a smaller tablet can’t push the limits as far and is thus more thermally constrained… The Nvidia Note 7 is only a small 7″ tablet but the Nokia Lumia 2520 and the other examples you gave were all just 10″ or more in size…

        It may not seem like a big difference but heat dissipation is effected by available surface area and even that small difference can significantly reduce the available surface area and thus further constrain the thermal limits the device will have to work with…

        And no, a development kit is not an example of what a final device to market will be configured…

        The K1 platform can support 4-8GB of RAM but for all the reasons I gave and more it’s not likely to be offered yet in final products…

        Really, many already support 4GB RAM and have for over a year… but yet do you see any 4GB RAM devices?

        So increasing the max support to 8GB is really only future proofing the platform but the market realities have to catch up before such capacities can be realized…

        Present RAM choices are limited by cost, availability, and performance/power efficiency concerns…

        Since 4GB isn’t a common option it means it’s not mass produced anywhere near as much as 2GB and thus cost significantly more… while increasing RAM generally means increasing power usage and that effects battery life that for mobile devices have to remain “good enough” and most would unfairly compare the run time even if it gave significant performance gains, which is the market reality system designers have to work with…

        Most will wait for the transition to LP-DDR4… which offers better power efficiency and ultimately lower costs onces it fully takes over as it can offer higher capacities in the same amount of space as the present LP-DDR3 does for lower capacities… and yes, space is another concern as keeping things thin and light means every tiny spec of space matters too…

        While I’m aware of GPU’s performance per watt… which is why you can expect a good GPU performance boost, as I already stated, but the rest of the SoC doesn’t have anywhere near that kind of improvement yet…

        One of the reasons they’re also switching to dual cores for the actual 64bit version of the K1 is because those limits means it’s easier to just reduce the number of cores and thus allow each core to perform better and not have to share power with so many other other cores… So like it or not there are still significant limits they have to work with and if they put the K1 into a similar 7″ tablet design then those limits will still be enough to significantly lower performance from what they are showing with a actively cooled model… or even what is possible with a larger tablet with less constraining thermals…