Tweaktown : Nvidia Jetson (Tegra K1)

[altairauditore]

This article:

http://www.tweaktown.com/articles/6226/a-tour-of-the-nvidia-hq-in-santa-clara-california/index.html

The last 2 pics show Jetson.. with a heatsink and a small fan.

The size of that heatsink.. K1 is probably running at least 20-30W+

I guess Tegra K1 when clocked conservatively will fit into a tablet, but in order to do all this automotive stuff, It is being clocked higher..

Nvidia states 326 GFLOP's, but if I remember well, the GTC Audi guy said around 500 GFLOP's were required for vehicle and pedestrian detection.

 

[ViRGE]

The device on the left looks to be Jetson Pro, not Jetson TK1. Jetson Pro a Tegra 3 board specifically designed for use with a discrete GPU and intended for automotive markets.

http://www.nvidia.com/object/jetson-automotive-development-platform.html

This is Jetson TK1: http://images.anandtech.com/doci/7905/Jetson_TK1.jpg

 

[NTMBK]

This is Jetson TK1: http://images.anandtech.com/doci/7905/Jetson_TK1.jpg

Actually, that's more of an artistic mockup which leaves out the cooler. The Jetson K1 looks like this:

 

[ViRGE]

Actually, that's more of an artistic mockup which leaves out the cooler. The Jetson K1 looks like this:

Well yeah. My point is that the real TK1 is a complete board in and of itself. It's not a card like in the original picture.

 

[NTMBK]

Well yeah. My point is that the real TK1 is a complete board in and of itself. It's not a card like in the original picture.

The card was a G-Sync board, not Tegra And the dev-platform in that article was actually the same one used to demo the K1 at CES: http://semiaccurate.com/2014/01/20/well-nvidias-tegra-k1-really-perform/

Perhaps the difference in cooler size is down to the difference between K1-with-A15s and K1-with-Denver? (Either way, the board in the article isn't the same as the Jetson K1 they just launched. It may just be different because it's a prototype.)

 

[ViRGE]

The card was a G-Sync board, not Tegra And the dev-platform in that article was actually the same one used to demo the K1 at CES: http://semiaccurate.com/2014/01/20/well-nvidias-tegra-k1-really-perform/

Perhaps the difference in cooler size is down to the difference between K1-with-A15s and K1-with-Denver? (Either way, the board in the article isn't the same as the Jetson K1 they just launched. It may just be different because it's a prototype.)

I think we're talking past each other here.

The device in the TT article is clearly Jetson Pro. Jetson Pro is the huge device with a ton of I/O (which a K1 SoC doesn't have). And the card next to it isn't the CPU (my bad), rather that's an MXM video card. Meanwhile the heatsink the OP is discussing is the heatsink for the MXM, not the SoC.

MXM card example: http://www.notebookcheck.net/fileadmin/_migrated/pics/module_example_01.jpg (compare the edge connectors)

(Note that it's not a G-Sync board, which is this)

On a side note, I should point out that Jetson Pro uses a Tegra 3, not Tegra K1. This specifically because T3 is the fastest Tegra to support PCI-Express, and hence the fastest Tegra to support a discrete video card.

Charlie's picture is also Jetson Pro. Note how it's identical to the Jetson Pro picture on NVIDIA's site. I have no idea why Charlie thinks it's a K1 demo (as far as anyone knows, K1 doesn't have PCIe). Maybe NVIDIA was running Jetson Pro at one point to showcase Kepler graphics?

The only K1 systems that were shown as of GTC were the Jetson TK1 systems, such as the one you linked to earlier.

 

[dorion]

Thanks for that additional pic of the real K1 ViRGE, I was sick of the entire internet using only the two press pics with the bad camera angle to hide the top right corner, but nano-itx hopes are now supremely crushed.

 

[NTMBK]

Thanks for that additional pic of the real K1 ViRGE, I was sick of the entire internet using only the two press pics with the bad camera angle to hide the top right corner, but nano-itx hopes are now supremely crushed.

In case the mass of pin-outs didn't give it away, this is a dev-board. The whole point is to cram on as much I/O as possible, so that developers can prototype any design they want to using it. It's not meant to sit in a HTPC case.

 

[ams23]

Nvidia states 326 GFLOP's, but if I remember well, the GTC Audi guy said around 500 GFLOP's were required for vehicle and pedestrian detection.

Nope, objection recognition requires ~ 180 GFLOPS throughput @ 30Hz refresh rate, while 3D scene interpretation requires ~ 280 GFLOPS throughput @ 30Hz refresh rate: http://cdn.pcper.com/files/imagecac...2014-03-25/Computer Vision On NVIDIA CUDA.jpg

FWIW, according to Rys @ B3D, the Tegra K1 Jetson kit will be available with a passively cooled variant. Even at north of 300 GFLOPS throughput, the TK1 GPU peak power consumption should be close to ~ 5w. A TK1 variant in a phone will likely have ~ 200 GFLOPS throughput while consuming less than 3w of power, while a TK1 variant in a tablet will likely have somewhere between ~ 200-300 GFLOPS throughput while consuming less than 5w of power.

 

[sontin]

Max power consumption of Jetson K1 is less than 11W under full load. This includes everything from CPU, GPU, RAM, I/O, HDMI...

https://twitter.com/ProfMatsuoka

 

[witeken]

No, it says 11W for GPU + CPU (+ memory). So it is indeed as power hungry as expected.

 

[SlowSpyder]

Nvida HQ looks like an awesome place to work! I'd totally bring my dog, she's familiar with technology.

 

[Insomniator]

Cool stuff... are they implying jetson is the kind of chip that will be used for self driving cars?

I want to work for Nvidia

 

[Nothingness]

Max power consumption of Jetson K1 is less than 11W under full load. This includes everything from CPU, GPU, RAM, I/O, HDMI...

Witeken already fixed your assertion: the 11W are only for CPU and GPU.

OTOH perf per Watt on SGEMM looks good: 26 SP GFLOP/W is three times more efficient than the 8.54 SP GFLOP/W of Xeon Phi 5110P (of course the comparison is silly, the Phi can do much more such as DP and is running SGEMM at 1796 GFLOPS vs TK-1 290 GFLOPS, but still ). Do some people have other SGEMM figures for other CPU?

 

[sontin]

No, it says 11W for GPU + CPU (+ memory). So it is indeed as power hungry as expected.

Nope:

Steve Oberlin NVIDIA Jetson (28nm) power efficiency. Approx. 26GF/W SFP including GPU/CPU, 2GB memory, other I/O.

From the same Twitter account.

 

[erunion]

Max power consumption of Jetson K1 is less than 11W under full load. This includes everything from CPU, GPU, RAM, I/O, HDMI...

https://twitter.com/ProfMatsuoka

So the K40 offers 4.2 GFLOPS, huh?

 

[Khato]

No, it says 11W for GPU + CPU (+ memory). So it is indeed as power hungry as expected.

Nope:

Steve Oberlin NVIDIA Jetson (28nm) power efficiency. Approx. 26GF/W SFP including GPU/CPU, 2GB memory, other I/O.

From the same Twitter account.

Perhaps you could clarify what exactly is the point of disagreement? That Witekin didn't include the ambiguous 'other I/O' which is part of the SoC regardless in the figure? Or that he rightly calls 11W for SoC + memory as being rather power hungry? After all, the memory side of it is likely less than half a watt, so you're talking at least 10W for the SoC.

 

[sontin]

Perhaps you could clarify what exactly is the point of disagreement? That Witekin didn't include the ambiguous 'other I/O' which is part of the SoC regardless in the figure? Or that he rightly calls 11W for SoC + memory as being rather power hungry? After all, the memory side of it is likely less than half a watt, so you're talking at least 10W for the SoC.

11W is the number for the whole board including RAM, I/O operations and output.

That brings the power consumption of the Chip down to ~8W under full load.

 

[strata8]

At least it's obvious that the K1 won't come close to the 360 GFLOP figure in tablets and phones. Not when it's consuming 8W+ to achieve that while current chips consume 3-4W or even less.

It's more efficient than the Adreno 330 for sure but it remains to be seen how it will compete with Adreno 420.

 

[sontin]

At least it's obvious that the K1 won't come close to the 360 GFLOP figure in tablets and phones. Not when it's consuming 8W+ to achieve that while current chips consume 3-4W or even less.

You can't compare the Jetson K1 number to tablets or even phones. Chips in these products throttle down a lot.

Tegra 4 in Shield is 5W for the CPU or GPU part.

 

[Nothingness]

So the K40 offers 4.2 GFLOPS, huh?

The K40 figures should be TFLOPS not GFLOPS (which then agrees with the GFLOP/W figure ).

 

[Khato]

11W is the number for the whole board including RAM, I/O operations and output.

That brings the power consumption of the Chip down to ~8W under full load.

It does? Because none of the sources that you provide state such. The slide simply states "CPU + GPU" while the twitter post states "GPU/CPU, 2GB memory, other I/O" - where do you see "whole board"? If such is an assumption on your part then that's all it is. Sorry, but given how cagey NVIDIA has been when it comes to providing power consumption numbers for, well, every single Tegra SoC thus far we have good reason to expect that they're measuring power for nothing more than they specifically state.

 

[witeken]

Indeed, if TK1 consumes something like 8W, they could have claimed a much bigger performance/watt advantage, which marketing wouldn't hesitate to do.

 

[Vesku]

Wonder if the retail K1 dev board will come with a cooler. Assuming it is using Northbridge mounting dimensions there aren't that many options available nowadays due to SoC direction of PCs.

http://www.newegg.com/Product/Produc...82E16813190005

 

[sontin]

nVidia showed this slide at the GTC:

40GFLOPs SGEMM / Watt is only for the chip and this is around ~8W (7.4W) for 296GFLOPs.

It does? Because none of the sources that you provide state such. The slide simply states "CPU + GPU" while the twitter post states "GPU/CPU, 2GB memory, other I/O" - where do you see "whole board"? If such is an assumption on your part then that's all it is. Sorry, but given how cagey NVIDIA has been when it comes to providing power consumption numbers for, well, every single Tegra SoC thus far we have good reason to expect that they're measuring power for nothing more than they specifically state.

The guy from the twitter account attended the presentation. If he says that the 26GFLOPs SGEMM/Watt is for the whole board then i believe him.
And nVidia is one of the few companies who doesn't hestitate with power numbers. Their 5W for Tegra 4 was spot on.