nvidia tegra K1

[jdubs03]

That 60fps interests me, why would it be 12fps higher than the Lenovo? Especially with the potential for greater cooling capacity. I'm surprised that 3dmark score wasn't as high either, for 2.5x the amount of GPU cores and a new architecture, only a 40% increase. Strange results, and the Snapdragon 805 will be competitive.

The Antutu result is also disappointing, for comparison that K1 was clocked at 2.0 ghz vs the tegra 4 clocked at 1.9 ghz, yet it has a worse score? TBH I was expecting an Antutu score of around 50,000. So, that was a let down.

 

[ams23]

The 3dmark unlimited score is almost certainly bandwidth-limited. Kepler.M cannot stretch it's legs in a bandwidth-limited scenario.

As for Antutu, the score appears to be heavily dependent on native panel resolution. Considering that the Lenovo panel has a 4K resolution (4x greater than 1080p), that is not even close to an apples to apples comparison (and explains why Tegra Note 7 with T4 has a higher Antutu GPU score in comparison).

 

[alawadhi3000]

I'm impressed, all is left is the power efficiency which nVidia always been subpar on that front.

 

[Blitzvogel]

I wonder if Nvidia possibly gave the Denver version GDDR5 compatibility for applications that may warrant it. Otherwise yes, it would just be bottlenecked for all eternity.

 

[ams23]

I'm impressed, all is left is the power efficiency which nVidia always been subpar on that front.

Sure, but I think improved power efficiency is pretty much a done deal due to process, experiential, and most importantly architectural improvements in TK1. The R3 Cortex A15 has superior power efficiency (2x better SPECInt perf. per watt in some instances) compared to the prior gen. Cortex A15, and appears to have significantly lower peak power consumption too even at max frequencies. And at the max GPU perf. levels of A7 and S800, TK1 appears to have at least 1.5x better GPU power efficiency in comparison (and that is not the maximal point for TK1's GPU power efficiency either). Hope to find out much more soon.

 

[tviceman]

More benchies this time from the note 7 inch reference tablet.

Note the tablet scores 60 vs the monitors 48.

http://blog.gsmarena.com/tegra-k1-gpu-benchmarked-beats-intel-hd-4400/

Wow, if these benchmarks hold up and TK1 armV15 can be bought reasonably within the first half of this year, color me impressed.

 

[Exophase]

Sure, but I think improved power efficiency is pretty much a done deal due to process, experiential, and most importantly architectural improvements in TK1. The R3 Cortex A15 has superior power efficiency (2x better SPECInt perf. per watt in some instances) compared to the prior gen. Cortex A15, and appears to have significantly lower peak power consumption too even at max frequencies.

Where is any of this coming from?

 

[North01]

Wow, if these benchmarks hold up and TK1 armV15 can be bought reasonably within the first half of this year, color me impressed.

I don't think the benchmark does justice for many of the devices on that list.

GFXBench 2.7 is OpenGL ES 2.0 based.

 

[alawadhi3000]

Sure, but I think improved power efficiency is pretty much a done deal due to process, experiential, and most importantly architectural improvements in TK1. The R3 Cortex A15 has superior power efficiency (2x better SPECInt perf. per watt in some instances) compared to the prior gen. Cortex A15, and appears to have significantly lower peak power consumption too even at max frequencies. And at the max GPU perf. levels of A7 and S800, TK1 appears to have at least 1.5x better GPU power efficiency in comparison (and that is not the maximal point for TK1's GPU power efficiency either). Hope to find out much more soon.

Are you sure that the power efficiency improvements will come because of the newer Cortex A15 revision rather than the HPm 28nm process which reduces the voltage at higher clockspeeds vs 28nm HPL therefore improving efficiency and performance??

 

[ams23]

Where is any of this coming from?

This is coming from a chart posted in Anand's TK1 preview: http://images.anandtech.com/doci/7622/Screen Shot 2014-01-06 at 6.18.12 AM.png

Note that for a SPECInt score of ~ 950 per core (which is close to what most T4-powered devices would achieve as a maximum per core), the SPECInt perf. per watt in TK1 is at least 2x greater in comparison (due to the combination of three factors I listed above). This massive perf. per watt advantage is maintained at most points along the curve.

Now, most TK1-powered devices will have a SPECInt score of ~ 1300 per core maximum. So, based on the graph, the peak power consumption per CPU core will be ~ 50% less for TK1 than for T4 (even though peak performance is nearly 40% higher in comparison)!

A device like Shield v2 will have a SPECInt score of ~ 1400 per core maximum (ie. the full 2.3GHz CPU clock operating frequency). Shield v2 will also have ~ 365 GFLOPS GPU throughput (ie. the full 951MHz GPU clock operating frequency).

 

[ams23]

Are you sure that the power efficiency improvements will come because of the newer Cortex A15 revision rather than the HPm 28nm process which reduces the voltage at higher clockspeeds vs 28nm HPL therefore improving efficiency and performance??

As mentioned above, the power efficiency improvement of R3 Cortex A15 in TK1 relative to Cortex A15 in T4 is due to three things: process, experiential, and architectural improvements.

 

[Roland00Address]

That 60fps interests me, why would it be 12fps higher than the Lenovo? Especially with the potential for greater cooling capacity. I'm surprised that 3dmark score wasn't as high either, for 2.5x the amount of GPU cores and a new architecture, only a 40% increase. Strange results, and the Snapdragon 805 will be competitive.

Most likely answer for the lenovo is
1) The chip are running different clock speeds
2) The lenovo is using early drivers.

We know the Denver version of K1 is only a few days old, we have no clue how old the A15 version of K1 is. The lenovo monitor is after all a prototype with beta silicon it is very possible that it has an older version of the chip while the inhouse tablet has more up to date internals, clock speeds, and drivers.

For example the lenovo monitor that toms hardware played with has a max cpu speed of 2 ghz when the keynote stated the max cpu of the A15 version will be 2.3 ghz. Maybe lenovo has a lower bin, maybe lenovo ran it at 87% cpu clock due to thermals (unlikely due to the larger surface area of a 28" monitor vs a 7" tablet), or maybe that lenovo is showing off a beta product and they don't want it to crash on the ces floor. Things like clockspeeds can always be changed before final production. Now this is a cpu clock speed we have no idea on gpu clockspeed, we know that A15 version of K1 reduces the clock speed to 1.4 ghz when you are focused on gpu and using all 4 cores, but we have no clue what the gpu clocks are.

-----------------------------------

I have the same question on the wide difference between 3D Mark vs GFX Bench. Is it memory bandwidith limited, rops bound, or some other reason?

 

[mikk]

That 60fps interests me, why would it be 12fps higher than the Lenovo? Especially with the potential for greater cooling capacity. I'm surprised that 3dmark score wasn't as high either, for 2.5x the amount of GPU cores and a new architecture, only a 40% increase. Strange results, and the Snapdragon 805 will be competitive.

OEMs devices are often slower compared to reference smartphones/tablets, nothing new. It could be also that Nvidia showed a Denver K1 Soc which might clock higher than the initially K1 devices from Lenovo.

 

[ams23]

No, everything showed to the public on the floor was a TK1 variant with R3 Cortex A15.

 

[mikk]

This:

OEMs devices are often slower compared to reference smartphones/tablets, nothing new.

 

[ams23]

That is true, but to the best of my knowledge, no TK1 "reference" device has actually been formally benchmarked yet.

As you can see from Tegra 4-powered devices, the perf. can vary up to ~ 15% depending on implementation. The CPU and GPU clock speeds used in the Lenovo prototype are at the low end of what I would expect to see in TK1-powered devices when they ship.

FWIW, the Lenovo 4K monitor is not being marketed as an Android all-in-one or Android tablet. It is being marketed as an ultra high res computer monitor that can also run Android apps (and will allow users to browse the internet, check email, watch videos, play casual games, and social network without turning the main computer on).

 

[alawadhi3000]

As mentioned above, the power efficiency improvement of R3 Cortex A15 in TK1 relative to Cortex A15 in T4 is due to three things: process, experiential, and architectural improvements.

I feel its mostly because of the move to HPm, as we seen from the performance difference between Snapdragon 600 and 800.

 

[Arkadrel]

That is true, but to the best of my knowledge, no TK1 "reference" device has actually been formally benchmarked yet.

The device the benchmarks came from probably looked like this:

from: http://semiaccurate.com

Ironically even in this large device the K1 is thermally constrained, is running nowhere near full clocks.

-S/A

That is not a 5W heatsink, think more than 2x that number
The heatsink on the K1s that Nvidia was showing in private was about 2 x 2 x .4 inches, a bit large for a 5W part don’t you think? In fact it is a bit large for a 28″ AIW device at least as far as Z/height goes, Lenovo can’t cool the 2.3GHz variant in the depth allowed. That should give you a very good idea about how much power a 2.3GHz K1 takes under load.

- S/A


Basically charlie is saying that those benchmarks where not done at like ~3 watts a smart phone might use, but probably around 10-15 ish.


If hes right doubt we ll see alot of these K1 in phones.

 

[NTMBK]

Jesus, that heatsink. That's the same size as what you needed for a fanless E-450:

And that's an 18W part.

 

[sontin]

Aha.

I bet using the same logic a 7" tablet with Tegra 4 would be never faster than a 21" AIO desktop with Tegra 4... Oh wait.

nVidia shown more than 10 reference tablets with Tegra K1 using the same Tegra Note plattform.

And this is the heatsink on the dev plattform:

/edit:

Basically charlie is saying that those benchmarks where not done at like ~3 watts a smart phone might use, but probably around 10-15 ish.

Nope. He said it in context of Tegra 4:

Take a look at how this was done with the Tegra 4 launch at MWC 2013. Anandtech got an exclusive on it and had some impressive numbers to share. Nvidia would not allow Anand to measure the power used to get those peak numbers, nor was he told what they were. The conclusion was hopeful but not unusually optimistic. Only much later was it revealed that the demo was pulling over 10W, beating the competition by large margins by using a large multiple of their power is not a difficult task.

He has no clue. Shield uses 5W and gets the same result. He has no clue, no information.

 

[Grooveriding]

The device the benchmarks came from probably looked like this:

from: http://semiaccurate.com

Look at that thing. That's the size of an HTPC...

Well it's no surprise. Nvidia always hypes and misrepresents Tegra until it falls apart when true power consumption is revealed and it winds up in hardly any devices except ones they make themselves.

Guess K1 will see some use in an updated Shield and get slaughtered in device wins by Qualcomm's and Apple's next gen SoCs in actual huge volume phone and tablet devices.

Typical Tegra BS marketing hype.

 

[jpiniero]

The question I suppose is that whether it's the A15 or the GPU.

 

[ams23]

The device the benchmarks came from probably looked like this

Well no, that makes no sense at all. All of the TK1 reference tablets shown at CES 2014 used a Tegra Note 7 inch tablet chassis with no fan/heatsink assembly. It doesn't take a rocket scientist to realize that adding a heatsink/fan assembly will allow higher sustained clock operating frequencies (just like any other SoC on the planet), but the difference is smaller than you think in a portable device (for instance, Shield is "only" ~ 15% faster in graphics benchmarks compared to Tegra Note 7).

Kepler.M GPU appears to be at least 1.5x more power efficient than the best ultra mobile GPU's today, while also having tremendous ability to scale to higher frequencies. That means that Kepler.M will be significantly faster with or without power constraints (irrespective of whether or not it is used in a smartphone, tablet, or actively cooled device like Shield).

On a side note, as mentioned earlier, the newly implemented Cortex A15 on 28nm HPM appears to have much better power efficiency and much lower peak power consumption compared to the A15 variant used in Tegra 4.

 

[Unoid]

Bump.. Do we have any new news on the Denver cores? I'm really interested in Nvidia's first real custom ARM CPU.

 

[witeken]

A quick Google search says no.