Tweaktown : Nvidia Jetson (Tegra K1)

[Khato]

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.

I would too if he said that. The twitter account you linked never said "whole board" however. It effectively said SoC + memory. Are you going to next claim that the 11W figure included power supply losses too? Edit: Though now that I translate the Japanese tweets I see where you're coming from, not with respect to 'whole board' but with subtracting 3W to arrive at SoC power consumption. However such also gives a pretty good reason to doubt the accuracy of his report since he's saying that memory accounts for 2-3W when 14.9GB/s of LPDDR3 should be pretty much exactly 1W - http://pc.watch.impress.co.jp/video/pcw/docs/457/163/p6.pdf

And nVidia is one of the few companies who doesn't hestitate with power numbers. Their 5W for Tegra 4 was spot on.

According to? As well, when precisely did they state 5 watt power consumption for the Tegra 4 SoC? A quick search is failing to yield results.

 

[ams23]

According to one of the GTC 2014 presenters from NVIDIA, the Tegra K1 SoC [326 GFLOPS SKU] consumes ~5w with very compute heavy tasks. This is remarkably good energy efficiency.

Presumably the <11w number is artificially derived. The CPU power consumption can reach as high as ~5w (in this SKU) with all four CPU cores active. The GPU power consumption can reach as high as ~5w (in this SKU) with all 192 CUDA cores active. So combined that would be ~10w, but in reality the Tegra K1 SoC will consume a maximum of ~5w that needs to be allocated and balanced between CPU and GPU depending on workload.

There will be Tegra K1 SKU's with higher performance (likely for actively cooled devices like a micro game console that has ~365 GFLOPS GPU throughput) and there will be SKU's with lower performance. I expect smartphone SKU's to have ~200 GFLOPS throughput while tablet SKU's may have closer to 250-300 GFLOPS throughput. We'll see how it goes.

 

[ams23]

nt

 

[TreVader]

Looks like K1 is just gonna be another Tegra failure. Barely S801 performance and probably won't be released on anything but a home brew tablet nobody will buy.







Nvidia needs to leave mobile to apple and samsung. They produce crap SoCs.

 

[ams23]

LOL, right, at ~ 1.5x more power efficient than Snapdragon 800, the Tegra K1 GPU should be significantly faster in comparison to Snapdragon 801's GPU, even in a smartphone form factor (and with much higher GPU headroom and a much more forward-looking GPU feature set to boot).

Tegra K1 is helping to kick start the Tegra automotive business. There are currently ~ 4.5 million cars in total on the road with Tegra inside, and within the next few years that will grow to ~ 30 million cars in total on the road with Tegra inside, which is exponential growth in that space! The ASP's should also be significantly higher in the automotive space than the consumer space. That said, I see no logical reason why Tegra K1 would not see some success in the consumer (and embedded) space, given the exceptional performance and feature set.

 

[VirtualLarry]

LOL, right, at ~ 1.5x more power efficient than Snapdragon 800, the Tegra K1 GPU should be significantly faster in comparison to Snapdragon 801's GPU, even in a smartphone form factor (and with much higher GPU headroom and a much more forward-looking GPU feature set to boot).

Tegra K1 is helping to kick start the Tegra automotive business. There are currently ~ 4.5 million cars in total on the road with Tegra inside, and within the next few years that will grow to ~ 30 million cars in total on the road with Tegra inside, which is exponential growth in that space! The ASP's should also be significantly higher in the automotive space than the consumer space. That said, I see no logical reason why Tegra K1 would not see some success in the consumer (and embedded) space, given the exceptional performance and feature set.

GPS units, with "Tegra Inside" stickers? LOL.

 

[teejee]

LOL, right, at ~ 1.5x more power efficient than Snapdragon 800, the Tegra K1 GPU should be significantly faster in comparison to Snapdragon 801's GPU, even in a smartphone form factor (and with much higher GPU headroom and a much more forward-looking GPU feature set to boot).

Tegra K1 is helping to kick start the Tegra automotive business. There are currently ~ 4.5 million cars in total on the road with Tegra inside, and within the next few years that will grow to ~ 30 million cars in total on the road with Tegra inside, which is exponential growth in that space! The ASP's should also be significantly higher in the automotive space than the consumer space. That said, I see no logical reason why Tegra K1 would not see some success in the consumer (and embedded) space, given the exceptional performance and feature set.

The price requirements in automotive industry is very very tough (I know...). Usually the suppliers are choosen by the purchase departement and they will always choose the cheapest alternative that is good enough. And soon there will be cheap SOC's fulfilling automotive requirements that are good enough even for advanced infotainment systems, so this is not a future high profitability area.

 

[TrulyUncouth]

LOL, right, at ~ 1.5x more power efficient than Snapdragon 800, the Tegra K1 GPU should be significantly faster in comparison to Snapdragon 801's GPU, even in a smartphone form factor (and with much higher GPU headroom and a much more forward-looking GPU feature set to boot).

Tegra K1 is helping to kick start the Tegra automotive business. There are currently ~ 4.5 million cars in total on the road with Tegra inside, and within the next few years that will grow to ~ 30 million cars in total on the road with Tegra inside, which is exponential growth in that space! The ASP's should also be significantly higher in the automotive space than the consumer space. That said, I see no logical reason why Tegra K1 would not see some success in the consumer (and embedded) space, given the exceptional performance and feature set.

As much as I would love for your optimistic view of tegra performance to be true, I am doubtful until we see real products in hand. I was HUGELY optimistic on Tegra 2, and it was behind in performance. I got re-hyped for Tegra 3 but it did not perform up to par and only saw real traction in the Nexus 7, which I bought and was very disappointed with. Tegra 4 was a trainwreck of epic proportions basically only seeing use in products Nvidia themselves produced- and some of those requiring a damn fan to work.

I hope more than anything they pull this off, especially if it makes it easier for companies to port from PC to Nvidia processors- but I have been burned by Tegra too many times to take Nvidia at their word on mobile.

 

[monstercameron]

As much as I would love for your optimistic view of tegra performance to be true, I am doubtful until we see real products in hand. I was HUGELY optimistic on Tegra 2, and it was behind in performance. I got re-hyped for Tegra 3 but it did not perform up to par and only saw real traction in the Nexus 7, which I bought and was very disappointed with. Tegra 4 was a trainwreck of epic proportions basically only seeing use in products Nvidia themselves produced- and some of those requiring a damn fan to work.

I hope more than anything they pull this off, especially if it makes it easier for companies to port from PC to Nvidia processors- but I have been burned by Tegra too many times to take Nvidia at their word on mobile.

oh come on tegra 3 was the fastest when it first came out [literally the first quad core] it did get overtaken quickly but that is another story. Tegra 4 also was a very good product and even though nvidia used a fan in shield [not a necessity and tablets dont need it] it still sips power [ http://www.notebookcheck.com/Test-NVIDIA-Shield-Spielkonsole.113587.0.html - nvidia shield review ] as the device only uses ~8W max [as sontin always points out] so maybe ~4-5W for the soc.

I am optimistic on tk1 and tk1v2, and nvidia partners porting games like mount and blade [ https://www.youtube.com/watch?v=zazHnQRg-1M ]

 

[TrulyUncouth]

oh come on tegra 3 was the fastest when it first came out [literally the first quad core] it did get overtaken quickly but that is another story. Tegra 4 also was a very good product and even though nvidia used a fan in shield [not a necessity and tablets dont need it] it still sips power [ http://www.notebookcheck.com/Test-NVIDIA-Shield-Spielkonsole.113587.0.html - nvidia shield review ] as the device only uses ~8W max [as sontin always points out] so maybe ~4-5W for the soc.

I am optimistic on tk1 and tk1v2, and nvidia partners porting games like mount and blade [ https://www.youtube.com/watch?v=zazHnQRg-1M ]

It came out in a very limited set of product with reduced frequency and by the time it was widely available it was nothing special performance wise. I came within a hairs' breadth of getting the transformer prime since I am in love with that form factor, but it had numerous problems and by the time they fixed those and came out with the plastic-backed transformer the complaints about performance were all over the internet. If I remember correctly, even the exynos quads that were in phones 3-6 months later had same/higher performance in a phone envelope.

I ended up getting the nexus 7 and it was a lag-fest in the big games of the time, especially dead trigger. It would stutter constantly, and that game was supposedly optimized for tegra.

As for Tegra 4, I can't say much since it was in such a limited set of products. I would have bought a shield in a heartbeat but I happened to go AMD this last graphics update so it wouldn't have been too useful to me. My understanding was that it was equal on performance but used more power than its contemporaries, is that wrong?

I can't stress enough how much I want Nvidia to succeed in this space- but they have been chronically a day late and a dollar short so I don't trust their pre-release speak anymore.

 

[Sweepr]

Dont know if already mentioned but when can we expect the first K1-based devices? They need to capitalize on its strenghs before Cherry Trail and Qualcomm 810 chips arrive late 2014/Q1 2015.

 

[ams23]

The price requirements in automotive industry is very very tough (I know...). Usually the suppliers are choosen by the purchase departement and they will always choose the cheapest alternative that is good enough. And soon there will be cheap SOC's fulfilling automotive requirements that are good enough even for advanced infotainment systems, so this is not a future high profitability area.

Due to the insatiable need for higher and higher graphics and compute processing power in a car, and due to the extremely robust hardware and software stack required, automotive SoC&#8217;s are not so easily commoditizeable. Being a high volume low cost chipset provider is not enough to fill the needs in this space, especially for the luxury car and sports car European brands (in addition to some Japanese brands) that NVIDIA has courted who are really pushing electronics technology forward.

Keep in mind that Tegra K1 is much more versatile than simply powering infotainment systems. TK1 will be used to power high res digital dash systems (the 2016 Audi TT will have TWO TK1's powering a 12.3 inch TFT display). TK1 will also be used to power advanced driver assistance systems (ADAS) that require high compute performance and require high programmability. In fact, a Tegra K1 automotive visual computing module (VCM) will actually save automakers quite a bit of money compared to what they would have needed to use in the past.

 

[ams23]

As much as I would love for your optimistic view of tegra performance to be true, I am doubtful until we see real products in hand.

I'm sure we will start to see some of that over the next 1-6 months (depending on TK1 SKU). Keep in mind that it generally takes ~ 9-12 months to go from initial product sampling to commercial availability, and as far as I can tell, TK1 first started sampling no earlier than July of last year.

TK1 is simply a much more versatile Tegra product than anything else NVIDIA has ever created in the past. It will be used in not just some high res tablets, clamshells and high end smartphones but also in: high res computer monitors, 4K televisions, micro game consoles, in-car infotainment systems, digital car dash systems, advanced driver assistance systems, healthcare systems, robotics, military unmanned drones, etc.

With TK1, NVIDIA finally has a unified graphics architecture from supercomputers to PC's to notebooks to ultra mobile, which means leveraged hardware design and software stack. The importance of this step cannot be underemphasized. TK1 will also serve as a bridge to a fully custom 64-bit CPU core, which is another milestone for the company.

To be considered a success, TK1 needs to be in high enough demand to accommodate what NVIDIA can supply to partners in the marketplace, so obviously we wait and see.

 

[PingviN]

4 generations of Tegra has gone and Nvidia has yet to deliver. I wouldn't get my hopes up that the 5:th time is a charm.

 

[ams23]

4 generations of Tegra has gone and Nvidia has yet to deliver. I wouldn't get my hopes up that the 5:th time is a charm.

I don't agree with that assessment. The original Nexus 7 (with Tegra 3) helped to kickstart Android tablets, and NVIDIA played a critical role in engineering a cost-effective SoC platform for that tablet. The Xiaomi Mi3 (with Tegra 4) for China Mobile has also been well received (among various other products). Obviously TK1 is such a big leap in so many areas that it doesn't take a rocket scientist to figure out that it will be very competitive compared to other high end SoC's on the market.

 

[NTMBK]

I don't agree with that assessment. The original Nexus 7 (with Tegra 3) helped to kickstart Android tablets, and NVIDIA played a critical role in engineering a cost-effective SoC platform for that tablet. The Xiaomi Mi3 (with Tegra 4) for China Mobile has also been well received (among various other products). Obviously TK1 is such a big leap in so many areas that it doesn't take a rocket scientist to figure out that it will be very competitive compared to other high end SoC's on the market.

The Tegra 3 was intended to be a high end, premium SoC, and yet NVidia ended up dumping it in one of the cheapest tablets on the market, the Nexus 7. That's not a resounding success, that's aggressive price slashing to get the chips out of the door. And then Google/Asus dumped them the following year.

 

[ams23]

The Tegra 3 was intended to be a high end, premium SoC, and yet NVidia ended up dumping it in one of the cheapest tablets on the market, the Nexus 7.

No, you don't get it. The Nexus 7 became a reality because the PLATFORM cost was sufficiently reduced by NVIDIA (based on Project Kai). The SoC cost was not magically reduced for the hardware maker Asus, nor was the SoC dumped on the market. And the Nexus 7 is certainly not close to being one of the cheapest tablets to produce on the market. Google makes virtually no profit on the hardware sales.

At the end of the day, the original Nexus 7 was purchased by many millions of customers and really boosted Tegra revenues, so it was an unqualified success for NVIDIA no matter how you spin it.

 

[PingviN]

No, you don't get it. The Nexus 7 became a reality because the PLATFORM cost was sufficiently reduced by NVIDIA (based on Project Kai). The SoC cost was not magically reduced for the hardware maker Asus, nor was the SoC dumped on the market. And the Nexus 7 is certainly not close to being one of the cheapest tablets to produce on the market. Google makes virtually no profit on the hardware sales.

At the end of the day, the original Nexus 7 was purchased by many millions of customers and really boosted Tegra revenues, so it was an unqualified success for NVIDIA no matter how you spin it.

Google said, this is what we'll pay. Nvidia took the deal to sell Tegra 3 - which wasn't happening in other devices. Qualcomm could've cut it's pricing, but didn't, because it had other sales. I'm not sure Qualcomm had the capacity to deliver enough SoCs to fill Google's need without cutting another customer off. Second generation Nexus 7 and Nvidia was out the door.

Nvidia can boost Tegra revenue all year long, but needs to be able to turn revenue into profits. That's not happening. A product that isn't profitable is not successful in the long run. Nvidia is approaching gen. 5 Tegra and is still losing money off it.

 

[ams23]

Profitability is a function of both revenues and costs. Tegra K1 is the first ultra mobile architecture that can truly leverage hardware and software design work done on their existing modern-day GPU architecture. This means mostly shared developmental costs for Tegra and all other lines of business moving forward. In fact, there is no great way to precisely determine Tegra profitability net of all expenses moving forward because so much of the R&D work will be shared and leveraged.

 

[NTMBK]

Profitability is a function of both revenues and costs. Tegra K1 is the first ultra mobile architecture that can truly leverage hardware and software design work done on their existing modern-day GPU architecture. This means mostly shared developmental costs for Tegra and all other lines of business moving forward.

The GPU in the Tegra is heavily customised- it's not like they just cut and pasted a Kepler SMX and called it a day. There is still plenty of R&D involved in that work. And while the GPU is moving to a (mostly) common design, the CPU is moving to a custom design in Denver.

 

[sontin]

Tegra 3 was the dominant player in the Android tablet market.
Tegra 3 is used in several cars from different manufactures - Tesla, VW, Audi, BMW etc.

The GPU in the Tegra is heavily customised- it's not like they just cut and pasted a Kepler SMX and called it a day. There is still plenty of R&D involved in that work. And while the GPU is moving to a (mostly) common design, the CPU is moving to a custom design in Denver.

Tegra K1 uses the same base Kepler architecture. It's not "heavily customised".

 

[ams23]

The GPU in the Tegra is heavily customised- it's not like they just cut and pasted a Kepler SMX and called it a day. There is still plenty of R&D involved in that work. And while the GPU is moving to a (mostly) common design, the CPU is moving to a custom design in Denver.

The core hardware components of a Kepler GPU [SMX] are all in place for the TK1 GPU. Same with the API feature set. Same with the software stack. All of which are leverageable. Even the Denver CPU core designs will be leverageable as they make their way up through other products (that may take some time though, and probably not before next year). The entire GPU architecture will be a mobile first design moving forward starting with Maxwell. So like it or not, Tegra is intertwined with other lines of business. Obviously incremental investment costs for Tegra will still exist (to the tune of $200-300 million per year IIRC?), but certainly nothing outrageous or extreme given the revenue streams.

 

[sham63]

It came out in a very limited set of product with reduced frequency and by the time it was widely available it was nothing special performance wise. I came within a hairs' breadth of getting the transformer prime since I am in love with that form factor, but it had numerous problems and by the time they fixed those and came out with the plastic-backed transformer the complaints about performance were all over the internet. If I remember correctly, even the exynos quads that were in phones 3-6 months later had same/higher performance in a phone envelope.

I ended up getting the nexus 7 and it was a lag-fest in the big games of the time, especially dead trigger. It would stutter constantly, and that game was supposedly optimized for tegra.

As for Tegra 4, I can't say much since it was in such a limited set of products. I would have bought a shield in a heartbeat but I happened to go AMD this last graphics update so it wouldn't have been too useful to me. My understanding was that it was equal on performance but used more power than its contemporaries, is that wrong?

I can't stress enough how much I want Nvidia to succeed in this space- but they have been chronically a day late and a dollar short so I don't trust their pre-release speak anymore.

I have the original N7 and the only game I recall not playing well was NOVA 3. Dead Trigger played fine, even Dead Trigger 2 played ok. My N7 is not rooted and stock.

 

[TrulyUncouth]

I have the original N7 and the only game I recall not playing well was NOVA 3. Dead Trigger played fine, even Dead Trigger 2 played ok. My N7 is not rooted and stock.

I guess there is always the chance mine was faulty, you never know. Mine was also stock and unrooted with only a few games installed. I wonder if you were playing at a later date where either the game or OS itself were updated- thats the only logical explanation I can think of. For me, Dead trigger would lag every 30-40 seconds with pretty serious hangs every few minutes. The performance in browsers was more than good enough for a 7 inch tablet but I clearly remember performance in Dead trigger and the OS to be lacking.

I was super disappointed since at the time I was super pro-android and even pushed friends and family to convert from Apple.

Are you sure you didn't have the same performance and maybe just wrote it off as acceptable for Android still in its early tablet days? I tried to justify the poor performance to myself before eventually giving up and selling the tablet.

I got the Note 2 soon afterwards and it didn't suffer from the issues in Dead Trigger or the OS making me think it was the Nexus and not problems with the game.

 

[sham63]

I guess there is always the chance mine was faulty, you never know. Mine was also stock and unrooted with only a few games installed. I wonder if you were playing at a later date where either the game or OS itself were updated- thats the only logical explanation I can think of. For me, Dead trigger would lag every 30-40 seconds with pretty serious hangs every few minutes.

I played it when it first came out. If it performed the way yours did I would have uninstalled it pretty quick.