nvidia tegra K1

[Homeles]

Last I checked Tegra was still running at a net-negative for profit. That was back during Tegra 3 days, though, so I'm not sure how it's changed.

AMD can't do well in that area. They're not an analog company, and don't own the IP, nor do they have the engineering talent. AMD's not interested in obtaining it either. The best they can do is offer SoCs and pair them with 3rd party baseband if necessary.

Nvidia, on the other hand, does have the resources.

 

[Arkadrel]

Fun fact - Mullins (AMD's next x86 tablet chip) scores 570 @ 3DMark11 according to AMD, thats a slightly lower score than Kabini A4-5000 (~590 pts). If Tegra K1 really does 60 FPS @ GFXBench 2.7 then it will outclass AMD's solution big time in the graphics department, Mullins should score around 35-36 FPS based on AnandTech's Kabini results.

http://www.notebookcheck.net/NVIDIA-Tegra-K1-SoC.108310.0.html

vidia claims the ability to render 60 fps in the GFXBench 2.7.5 T-Rex offscreen test.

Anyone know if the 2.7.5 differnce matters much? any perf. differnces?

That is intresting though.

This Mullins (with the 570 score) is rated as ~2W SDP (4.5w TDP)

Im not sure how K1 's 5watts are measured but itd be intresting to see how they match up.

 

[mikk]

570 in 3dmark11 is probably way too slow for K1 and Cherry Trail.


I couldn't notice differences in 2.7.5. Should be the same. This is only a minor version difference, unlikely that they changed something performance related.

 

[sushiwarrior]

i always felt AMD missed the boat not getting into this sooner like nvidia did. With the pc market continuing its shrink, I wonder how AMD will keep afloat, since they don't have many other established avenues of income like nvidia.

AMD has been planning to make ARM processors for several years now - essentially eons compared to Nvidia, who were able to quickly convert Denver to be compatible with ARMv8 using software conversion.

I think both AMD and Nvidia have been planning to succeed with this market for a long time now.

 

[sontin]

Im not sure how K1 's 5watts are measured but itd be intresting to see how they match up.

5W is TDP for the whole SoC.
2W is the average power consumption of the Kepler.M gpu core.

 

[ams23]

AMD Mullins only narrowly fits inside a thin fanless tablet (it will only be used in 11.6" or larger tablets). Tegra K1 will be used in much smaller tablets and high end smartphones (among other things such as cars, tv displays, micro-consoles, etc.).

 

[caswow]

AMD Mullins only narrowly fits inside a thin fanless tablet (it will only be used in 11.6" or larger tablets). Tegra K1 will be used in much smaller tablets and high end smartphones (among other things such as cars, tv displays, micro-consoles, etc.).

yea just like last time? and last time? and last time? i dont think so:whiste:

 

[Khato]

5W is TDP for the whole SoC.
2W is the average power consumption of the Kepler.M gpu core.

According to? I know that Anandtech provided some speculation as to how the K1 GPU could get down to a 2-3W range, but I'm not certain that matches up very well with the PR materials that NVIDIA provided. At least not when running at peak performance.

In addition to NVIDIA's general 5W power figure for the comparison against previous generation consoles (both of which were given a 100W power figure, so who knows what power consumption they're actually measuring) they also provided AP+DRAM power figures while running GFXBench 3.0 at the same performance level of the Apple A7. If you assume that the memory portion is consuming somewhere around 0.5W during this load then the AP is roughly 1.25W for Tegra K1 and 2.05W for A7.

However according to another of NVIDIA's slides peak performance of the Tegra K1 is 2.5x that of the A7. Which means that at that roughly 1.25W level the GPU only needs to run at roughly 380 MHz. Increase that up to the 950 MHz necessary to reach their 365 GFLOPS claim and thanks to both the frequency and voltage increase I'd be surprised if it wasn't at least in the 4-5W range.

All that said, it is quite the impressive mobile GPU, especially since it catapults the feature set up to parity with what we're used to on PC gaming. (And unlike Baytrail, actually has the performance to make some use of it.) Not to mention that it'll be quite the efficient solution when the workload doesn't require it to run at full frequency. (Sure it'll be pretty good at full frequency given the performance it'll offer, but I'd expect it's quite a bit away from the 'sweet spot' of efficiency.)

 

[tviceman]

Until then though x86 is just too strong and if Intel continues to improve or better yet uses kepler IP I don't see android going anywhere from a gaming perspective in this form factor but I don't know anything.

Would be fantastic if Intel switched to 100% Nvidia IP for integrated GPU, but highly doubtful.

 

[Enigmoid]

5W is TDP for the whole SoC.
2W is the average power consumption of the Kepler.M gpu core.

Tegra 4 has something like a TDP of 10W under max load. Doubt K1 is going to cut that in half.

 

[NTMBK]

Tegra 4 has something like a TDP of 10W under max load. Doubt K1 is going to cut that in half.

To be fair the K1 finally has unified shaders in the GPU, so we should see a decent efficiency boost.

 

[ams23]

Tegra 4 was rated at 5w TDP just like Tegra K1. Peak power can probably go higher. Most revewers are only able to measure system power (ie. CPU + GPU + mem + LCD screen etc.).

T4 appears to work well in Tegra Note (7" tablet) and Xiaomi Mi3 (5" smartphone), among others, so I don't expect anything less with TK1.

I think you will hear a lot more about TK1 at Mobile World Congress 2014 (Feb 24-27). That is the proper venue to show real-world mobile products.

Maxwell info should come at GTC 2014 (Mar 24-27).

 

[Ajay]

To be fair the K1 finally has unified shaders in the GPU, so we should see a decent efficiency boost.

Much better efficiency and built on HPM, but more than twice the number of shaders - so I'd imagine that the power usage would wind up being similar for much better GPU performance. I guess we'll find out. I think allot of the numbers Anand was tossing about were hypothetical - though they may be from based on some information from NV, which they are not releasing yet.

 

[Khato]

I think allot of the numbers Anand was tossing about were hypothetical

That's definitely my impression. More an explanation of how NVIDIA could be ending up at the kind of power figures necessary for the market than a good estimate. For example, he jumps from 5W at 1GHz to the 2-3W range at 900 MHz (despite the fact that the 365 GFLOPS number requires 950 MHz). Even dropping from 5W to 3W for running at 90% the frequency would require running at 80% the voltage. Not to mention the earlier step of simply subtracting 6W worth of leakage which he does state isn't exactly the proper way to do such calculations - the real point is just that NVIDIA can tame a single SMX down to the kind of power consumption necessary for both smartphone and tablet operation.

 

[Enigmoid]

That's definitely my impression. More an explanation of how NVIDIA could be ending up at the kind of power figures necessary for the market than a good estimate. For example, he jumps from 5W at 1GHz to the 2-3W range at 900 MHz (despite the fact that the 365 GFLOPS number requires 950 MHz). Even dropping from 5W to 3W for running at 90% the frequency would require running at 80% the voltage. Not to mention the earlier step of simply subtracting 6W worth of leakage which he does state isn't exactly the proper way to do such calculations - the real point is just that NVIDIA can tame a single SMX down to the kind of power consumption necessary for both smartphone and tablet operation.

950 Mhz is pushing the sweet spot on Kepler (going by GK 110 speeds).

Nvidia is also omitting the fact that the 740m with DDR3 is an extremely bandwidth limited part, generally weaker than the 650m + GDDR5 at 835 mhz and a lot of that GPU is sitting unused.

That said Mobile Kepler is an extremely OC-able part.

AMD was able to fit 128 GCN cores into Kabini but had to greatly reduce speeds and at a much higher power envelope.

 

[witeken]

i think intel and qcomm will be worried if nvidia executes on time. denver CPUs should put in the leagues of silvermont or better. and I dont think even cherrytrail GPU would be big enough to beat this. qcomms 805 looks like a re-spin of last years chip

We don't know anything about the performance/watt. High Denver CPU performance doesn't mean anything if it consumes tons of energy like Apple's A7. 14nm 3D Tri-Gate Airmont/Goldmont with gen8/9 gpu should be very competitive with everything (CPU + GPU) at 28 and 20nm. It would also surprise me if Denver beats Silvermont at performance/watt.

 

[jdubs03]

Denver is a massive upgrade, I think it will easily be competitive with Airmont/Goldmont CPU wise. It could be competitive with i3-U and Broadwell-i5/i7-Y. Denver will likely have higher multi-thread peformance than K1v1 (A15r3).

Tegra 4 had about 3000 multi-thread on geekbench 3. Their whitepaper has K1v1 at 1.4x performance. So, thats 4200+ multi thread. Denver has 2 cores, so we're already at least 2100 points for single-thread (I said near 2,000 in my last post, and assuming better multi-thread scaling).

i5-4200U is ~2500, 4300U is about ~2700.

And if that one GFXbench result is accurate to true GPU performance, then we're seeing HD4400 level performance with 3x less TDP.

All this on 28nm is pretty incredible, imagine 16nm FinFET (though 20nm BEOL) will give an additional 35% perf (according to TSMC), that is 4500U perf. territory.

A lot of ifs, but Denver is a massive jump in terms of CPU and GPU compared to past iterations.

 

[ams23]

We don't know anything about the performance/watt. High Denver CPU performance doesn't mean anything if it consumes tons of energy like Apple's A7.

What makes you say that A7 [Cyclone] CPU consumes "tons of energy"? A7 seems to work fine in the relatively small and thin iphone 5s (albeit with some throttling compared to the variant in ipad Air).

 

[sontin]

Tegra 4 has something like a TDP of 10W under max load. Doubt K1 is going to cut that in half.

Nonsense. TDP for Tegra 4 is 5W.

 

[ams23]

And if that one GFXbench result is accurate to true GPU performance, then we're seeing HD4400 level performance with 3x less TDP.

All this on 28nm is pretty incredible, imagine 16nm FinFET (though 20nm BEOL) will give an additional 35% perf (according to TSMC), that is 4500U perf. territory.

A lot of ifs, but Denver is a massive jump in terms of CPU and GPU compared to past iterations.

True, but I think that ultra mobile technology will get really interesting when we combine custom CPU/GPU with features like unified virtual memory (to improve application processor performance and efficiency) and stacked DRAM (to improve memory bandwidth). This should happen within the next 2-3 years.

 

[sontin]

However according to another of NVIDIA's slides peak performance of the Tegra K1 is 2.5x that of the A7. Which means that at that roughly 1.25W level the GPU only needs to run at roughly 380 MHz. Increase that up to the 950 MHz necessary to reach their 365 GFLOPS claim and thanks to both the frequency and voltage increase I'd be surprised if it wasn't at least in the 4-5W range.

Running a GPU centric benchmarks allows the SoC to give most of the power budget to the GPU.

 

[Khato]

Running a GPU centric benchmarks allows the SoC to give most of the power budget to the GPU.

Correct, which results in the GPU power being that much higher than the 2W figure you stated. Since according to NVIDIA's own numbers the AP+DRAM consumes 1.75W when running at 40% of its peak performance. Even if you say that only 1W of that is from the GPU then you're still looking at it jumping up to at least 4W once running at full speed. And that's on the low end considering that it'd go from 1W to 2.5W if it could run at full speed at the same voltage... but just going from 0.8V to 1V would bump that up to ~3.9W. (For comparison, desktop Kepler goes from something like 0.875V to 1.125V I believe it is?)

Really, I wouldn't be surprised if the actual numbers look worse than that though as I suspect that at the lower performance levels everything other than the GPU in that power figure is using less than half a watt. (One figure I found for LPDDR3 1600 is that it's around 67mW/GBps, so if you assume that full performance saturates the bandwidth while 40% of peak only uses that percentage then the DRAM would only be around 345mW.)

 

[sontin]

The 2W number comes from an "average of popular Android games" (see the whitepaper) and is for the Kepler.M gpu core.
The 1.5x more efficient of Tegra K1 comes with the same performance - but this number is for the whole SoC.

 

[Khato]

The 2W number comes from an "average of popular Android games" (see the whitepaper) and is for the Kepler.M gpu core.

Thanks for providing a source - http://www.nvidia.com/content/PDF/tegra_white_papers/Tegra-K1-whitepaper-v1.0.pdf for those who don't want to search for it, first paragraph of page 10. And that definitely explains the discrepancy seeing as how I'd be surprised if an 'average of populate Android games' even gets the GPU up to a 50% load. Especially if that's run on an equivalent to the Tegra Note 7 at only 1280x800 resolution with a 60 fps cap.

 

[Enigmoid]

Nonsense. TDP for Tegra 4 is 5W.

But at max clocks under full load 5W is not going to cut it. That is easily seen in a lot of devices.

4x A15 @1.8 ghz + GPU at full load is over 5W.