Nvidia’s Tegra K1 SoC has Denver CPU cores, Kepler graphics

In early 2011, during a CES press event, Nvidia revealed its Project Denver CPU initiative. On Sunday evening, at another CES press conference, the company provided a glimpse of the first Denver-based processor: the Tegra K1. This next-generation SoC features dual Denver CPU cores clocked at up to 2.5GHz. The cores were designed by Nvidia, and they’re compatible with the 64-bit ARMv8 instruction set. They have a seven-way superscalar pipeline and a hefty 128KB + 64KB of L1 cache. The L1 instruction cache appears to be the larger of the two.

The Tegra K1 comes in two pin-compatible configurations. In addition to the Denver-based duallie, there’s a quad-core version based on ARM Cortex-A15 cores similar to those used in the Tegra 4. That second chip is limited to 32-bit applications, and its CPU cores will only scale up to 2.3GHz.

Source: Nvidia

Based on Nvidia’s die shots, two Denver cores appear to have the same silicon footprint as four Cortex-A15s. It will be interesting to see how the two implementations compare.

At the very least, both chips should have plenty of graphics horsepower. The Tegra K1’s DirectX 11-class integrated graphics has 192 shader processors based on the Kepler microarchitecture. In Nvidia’s desktop and notebook GPUs, each Kepler SMX unit also has 192 shader processors. The low-end GeForce GT 640 includes two SMXes, while the high-end GeForce GTX 780 Ti has 15 of them.

The Tegra K1 die (Cortex-A15 version). Source: Nvidia

Nvidia stopped short of revealing the graphics clock frequency, but it did show the Tegra K1 running the next-generation Unreal Engine 4. Scott was on the scene, and he describes the demo as "freaking amazing." I don’t believe that demo was running on the Denver-based variant. Nvidia did a live demo of its home-grown duallie on stage, though. The chip has only been back from the fab for a few days, according to the company.

Like its Tegra 4 predecessor, the Tegra K1 is built using 28-nm fabrication technology. Well, the Cortex-A15 version is, anyway. The official product page doesn’t list the Denver incarnation yet.

Comments closed
    • ronch
    • 6 years ago

    Is Denver Nvidia’s first-ever CPU core designed entirely in-house?

      • chuckula
      • 6 years ago

      Yes. All the other Tegra parts have been using stock ARM cores.

    • ronch
    • 6 years ago

    Apart from Jerry Sanders and Steve Jobs, Jen Hsun Huang is also one of the Silicon Valley big guys I admire. These three people practically started from scratch with only their vision and not much else, yet they tirelessly built their respective companies and led them through countless challenges to become some of the industry’s most successful companies. Ok, maybe AMD isn’t so successful right now but they are by all means an amazing company which started its slow decline only after Jerry left. Let’s hope Cook doesn’t mess Apple up and Jen Hsun finds a worthy successor when he eventually rides off into the sunset.

    • ronch
    • 6 years ago

    So the first Denver-based SOC only has 2 CPU cores? If only Nvidia can make use of Intel’s smaller fab processes….

    Edit – Really don’t know why I was downthumbed. IT HURTS!!!! XD

      • tipoo
      • 6 years ago

      If they have higher enough single threaded performance than the quad core Cortex A15, it will still be better imo. The A7 shows a strong dual core is usually better than a quad with weaker cores. And most testing also shows those quads can’t sustain high clocks on all cores very long, and typical workloads are still only hitting the first two cores while the others are offline or in low power mode, the third core sometimes gets hit mildly, the fourth almost never.

        • ronch
        • 6 years ago

        I know that. But today most companies would like to cram as many cores in there so they can hype it up as much as possible. I’m sure Nvidia wants to put 4 cores in there but they’re hampered by 28nm, which limits cost as well as power and therefore, also how much they can push these cores.

        Given the choice between two fast cores vs. four slow cores, I’d take the former, of course. (Although I got the FX-8350 instead of the 3570K… but it’s mainly due to my fascination with the Bulldozer architecture.)

          • smilingcrow
          • 6 years ago

          “Given the choice between two fast cores vs. four slow cores, I’d take the former, of course. (Although I got the FX-8350 instead of the 3570K… but it’s mainly due to my fascination with the Bulldozer architecture.”

          I’ve heard stories about motorway/freeway traffic being slow due to drivers slowing down to look at accidents on the carriage going the other way. Sound familiar?
          I recall this video at the time of AMD’s crash – [url<]https://www.youtube.com/watch?v=SArxcnpXStE[/url<]

          • tipoo
          • 6 years ago

          I’m not interested in what’s most marketable. Two larger faster cores with good idle characteristics is the sane way to go.

            • ronch
            • 6 years ago

            I’m not talking about the consumer’s perspective, I’m talking about the chipmaker’s perspective.

            • tipoo
            • 6 years ago

            I dunno. Maybe like Intel giving up the GHz race long ago, Nvidia is giving up the core race in favor of sane solutions. They did market themselves into a hole, but so did Intel back then, and core performance going forward will speak for itself to those who care for such things.

      • shank15217
      • 6 years ago

      Probably multi-threaded.

    • ronch
    • 6 years ago

    I’m looking at the die shots and am trying to locate the CPU cores. With most dies, the CPU cores are easy to spot because you just have to look for symmetric, repeating structures that also have memory/cache structures within them. Not so with this SOC. I guess, just like most of today’s ‘die shots’, they’re just spewed out by the design tools instead of being actual die shots.

    Edit – Ok, got it. I was looking at the big die shot with the four A15 cores.

      • Beelzebubba9
      • 6 years ago

      These aren’t even die shots – they’re just artists renditions.

        • ronch
        • 6 years ago

        As I’ve mentioned, these aren’t real die shots. I wish they would put out real die shots again, but this looks like the new fad among chipmakers these days.

          • djgandy
          • 6 years ago

          Chipmakers? Don’t put everyone else in the Nvidia basket. Technically Nvidia aren’t really a chipmaker.

    • vargis14
    • 6 years ago

    I wonder how it will compare to the SOC that is in the SHIELD.

    I think it would be pretty cool if these new denver chips make it into some 1080p screened phone’s or tablets that will let you stream just like the SHIELD.

    I wonder if this is the chip for the ShIELD 2 that will have to be Full HD

      • tipoo
      • 6 years ago

      The Cortex A15 model should be pretty similar in CPU performance at least. I wonder how well Nvidias first custom CPU cores will do. The GPU, anyones guess as the last one wasn’t a comparable architecture.

    • sweatshopking
    • 6 years ago

    My favorite part is where they say the ps3 was dx9.

      • ferdinandh
      • 6 years ago

      Yep very similar to a GeForce 7800gs which is a DirectX9 card. Though you probably wanted to sound smart by flaunting your ‘knowledge’ that ps3 uses a special version of opengl.

        • sweatshopking
        • 6 years ago

        OH SNAP CALLED OUT ON MY PLAN. GJ YOU CAUGHT ME!!! NOW WHAT DO I DO?!

      • tipoo
      • 6 years ago

      Sure the console doesn’t use DirectX, but the chips feature level was up to DirectX9.

    • tipoo
    • 6 years ago

    Haven’t they claimed Xbox 360 performance for a while now? Heck, they claimed Core 2 Duo performance with the Tegra 3 which was a bonkers claim, even the Apple A7 is more like a first gen ULV Core Duo.
    This time I do believe them more though, from the Gflops as a ballpark, but sustained bandwidth and throttling and all that jazz will still be the telling factor.

    • tipoo
    • 6 years ago

    The “192 cores” marketing is already confusing and falsely exciting a lot of less technical places on the internet. Sure, and my GPU has 2000 “cores”.

      • derFunkenstein
      • 6 years ago

      It confused supposedly technical sites, too. Droid Life led with the “192 core” headline. Derp.

      • ronch
      • 6 years ago

      Heck, even Creative has jumped on this sort of misleading insanity. Their latest SoundCore3D chip is a quad core! It’ll blow your ears off, man.

        • willmore
        • 6 years ago

        Idiotic headline deserves idiotic response: “But, I only have two years, why do I need four cores?”

          • ronch
          • 6 years ago

          Well, it means each of your ears gets two cores. And two cores are better than one, or half of a core, right? Trust me, you need a quad core sound card. /sarcasm

            • Meadows
            • 6 years ago

            640 cores will be enough for anyone.

            • ronch
            • 6 years ago

            Depends what sort of cores you’re talking about. It seems different companies have given ‘core’ their own meanings.

            • djgandy
            • 6 years ago

            Think you missed the joke…

    • blastdoor
    • 6 years ago

    Apple must *love* the fact that the component suppliers to its competitors provide all of this info regarding future products.

      • HisDivineOrder
      • 6 years ago

      So far, it doesn’t seem to help them all that much. Not sure Apple runs its business responding to other companies as if they are, they aren’t really succeeding at it.

      Now if Apple runs its business like it’s in charge and leads the way, telling everyone else what’s important and expecting both its customers and its competition to follow in its wake, well then they’re doing a damn fine job of doing that.

      A smart company would do a mix of the two, but Apple isn’t the smartest of companies. They have smart people, but they do some pretty stupid things.

        • blastdoor
        • 6 years ago

        Apple knew that nobody else would have a 64 bit ARM SOC in 2013 because everybody else announces their plans a year or two in advance.

        Nobody knew that Apple was going to trump everyone and introduce a 64 bit ARM SOC in 2013 because Apple can keep its plans secret.

          • TheEldest
          • 6 years ago

          Except that 64-bit SOCs are useless right now. The only advantage is addressable memory and devices are still using less than 4GB.

            • derFunkenstein
            • 6 years ago

            That’s actually not the only advantage. Anand’s dissection of the Apple A7 SoC would be helpful for you.

            • MathMan
            • 6 years ago

            There’s an article out there (can’t find it) that shows how 64 bit pointers can (and are) abused by iOS to store additional metadata that makes memory allocation and deallocation for objects up to 15 times faster. It stands to reason that Android should be able to achieve similar speed up for these kind of common operations.

            • WillBach
            • 6 years ago

            Nitpick: it’s not abuse if it’s part of the ISA 😉

            • MathMan
            • 6 years ago

            They were using the lower 40bits of a pointer as address and the upper bits as reference counter and some other object attributes. This really reduces the amount of times they had to go to external memory.

            Maybe not abuse, but definitely a hack! And something the ARM designers had envisioned.

            • WillBach
            • 6 years ago

            If “hack” means quick fix that has negative implications for future revisions or something that just happens to work then I wouldn’t even call it a hack. The fact that there’s built-in support for masking those non-address bits in the ISA is baising me towards calling it a feature. You could do the same thing on most AMD or Intel chips which use either 40 or 48 bits of a 64-bit pointer (though they have standards for “canonical” addresses) – that would definitely be a hack in my opinion.

            • MathMan
            • 6 years ago

            The support in the ISA for this feature is called the ‘AND’ instruction. So, yes, no canonical support, a real hack. Which is not an insult in any way from an engineering point of view if it does what it is supposed to do.

            And it doesn’t result in future backward compatibility, since it is the OS that deals with memory management and that assigns addresses etc.

            I found the article!
            [url<]https://www.mikeash.com/pyblog/friday-qa-2013-09-27-arm64-and-you.html[/url<]

            • WillBach
            • 6 years ago

            It’s a good article and I understand that you didn’t mean “hack” as any kind of insult but I still think support goes beyond “AND”. An in-person presentation on ARMv8 that I saw mentioned explicit support for an architecture defined (but implementation dependent) number of non-address bits in pointers and explicit support for masking those during certain operations. Maybe I misunderstood but at the time it seemed pretty clear. I’ll try and look it up.

            • MathMan
            • 6 years ago

            If you saw an in-person presentation about this, I’m sure you’re more qualified to talk about the subject that I am!

            • WillBach
            • 6 years ago

            I should try to be more helpful (than I have been so far) and pull up a web-accessible version of the same stuff 🙂

            • Farting Bob
            • 6 years ago

            The biggest advantage for apple is it can market its phone/tablets as “the worlds first 64bit phone”. For a company that places big emphasis on being revolutionary compared to the rest it will certainly help maintain that image, even if its nearly impossible to tell a difference outside of benchmarks.

            • blastdoor
            • 6 years ago

            There is a confound here between 64 bits and ARMv8 — ARMv8 is apparently incredibly useful right now, at least in Apple’s implementation.

            Apple knew this would be the case, and Apple knew (more or less) when everyone else was going to introduce ARMv8 SOCs. But nobody knew when Apple would do it. This is clearly a huge advantage to Apple.

            And looking forward, nobody knows what Apple is cooking up for the A8 and A9, but Apple knows a fair bit about what everyone else is doing.

            • sherb
            • 6 years ago

            Why is armv8 such a huge improvement?

            My 2 year old laptop running win 7 64bit(all the specs are better too) performs pretty similarly to my 5 year old desktop on winxp :/ The start up time on laptop is actually much worse.

            • Beelzebubba9
            • 6 years ago

            What you’re referring to is perceived performance, which has more to do with the platform and your installed software and system configuration. Due to how OS X and modern Macs are optimized, my 2013 rMBP ‘feels’ faster than my gaming desktop, but any heavy work (or gaming) quickly demonstrates the actual difference in power between them.

            The deltabetween the performance of the ARM v7 and the ARM v8 ISA are pretty technical and unlikely to be noticed by the average user except in some extreme edge cases like AES encryption where ARM v8 shows something like an 800% speed improvement. You can read more about the details here:

            [url<]http://anandtech.com/show/7335/the-iphone-5s-review/4[/url<]

    • chuckula
    • 6 years ago

    Here’s the Nvidia theme song for Denver: [url<]http://www.youtube.com/watch?v=tWeuXsQMCGY[/url<]

      • Yeats
      • 6 years ago

      [url<]https://www.youtube.com/watch?v=A3yt2aH42JE[/url<]

    • Pax-UX
    • 6 years ago

    It’s faster then the 360 and PS3 not bad… Makes you wonder if nVidia will go after high performance in the mobile sector to undermine next gen consoles?

    Can’t wait till I can buy two mobile phones and do SLI mobile gaming 😛

      • tipoo
      • 6 years ago

      [quote<]Makes you wonder if nVidia will go after high performance in the mobile sector to undermine next gen consoles? [/quote<] Duh? 😛 It's already known Android is their vision of the future of gaming. They've expressly said so themselves.

        • bittermann
        • 6 years ago

        Only because that is the last direction they have left…no other choice.

          • tipoo
          • 6 years ago

          Which makes it no less true though. Their focus for the future is on Android regardless of why.

            • bittermann
            • 6 years ago

            Or no less false…

            • tipoo
            • 6 years ago

            Or no less counterantitrue.

            • Diplomacy42
            • 6 years ago

            Or no less sexy…

            • willmore
            • 6 years ago

            Lie to me, baby! Lie to me!!!

        • WaltC
        • 6 years ago

        Ah, Android. The Win3.1/95 of the future! What’s not to like?

        I mean, why should nVidia care if 98% of its current profits and revenues come from x86 hardware markets? Psssshaw. Who needs it? Right, JHH?

          • ferdinandh
          • 6 years ago

          I have no idea what your point is.
          1 Agreed. Android is the Win95 of the future which means the next 10 years it will have a 95% monopoly and all important software will be written for it.
          2 Are you sure those Tesla super computers are x86?
          3 Why does x86 matter? A GPU works on either x86 or arm. And you need a cpu if you make a soc. Most computers sold today have a soc.

            • cygnus1
            • 6 years ago

            No, Win95 didn’t really last 10 years, I think that he was implying it’s destined to be replaced by something that looks similar but is really drastically better. Win32 has stuck around, but not the old 3.1/9x code base. Most people shifted off of that to NT with XP and some earlier with Win2k.

            Also, what’s a tesla supercomputer? Do you mean the tesla gpu compute cards (http://www.nvidia.com/object/tesla-workstations.html) that about 99.999% of are installed in x86 based systems? Including workstations and supercomputers?

      • Pettytheft
      • 6 years ago

      They are just now getting to PS2/Xbox level graphics. I’ll believe it when I see it in hand.

      • maxxcool
      • 6 years ago

      until i can be verified the demo was run on Denver… not real in my book. Even the author has doubts to the claim from, the company that showed us a pcb cut in half and glued on bits from 2 years ago calling it a gpu.

    • renz496
    • 6 years ago

    so no low power companion cores for the denver based cpu?

      • MadManOriginal
      • 6 years ago

      The die shots are a lie.

        • renz496
        • 6 years ago

        i haven’t see nvidia mention 2+1 either unlike the quad core K1.

      • tipoo
      • 6 years ago

      Maybe with their custom cores they’re able to get idle/low power down low enough without it, a la Apple A7.

    • Alexko
    • 6 years ago

    “Based on Nvidia’s die shots”

    Be careful with those, they’re actually Photoshop jobs (and have been since Tegra 4). A quick look at a GK104 die shot will confirm that a Kepler SMX looks absolutely nothing like what NVIDIA is showing here. I guess they’re trying to convey the idea that the so-called “192 Cuda cores” are real independent cores, like on a CPU.

    That said, the CPU portion of the “die shots” could be real.

      • jjj
      • 6 years ago

      ” the CPU portion of the “die shots” could be real.”

      And with what kind of magic did they made that 5th A15 core so tiny?

        • Alexko
        • 6 years ago

        ARM typically offers different implementations of their cores ( they call it POP or processor optimization pack: [url<]http://arm.com/products/physical-ip/processor-optimization-pack/index.php[/url<] ). Some are optimized for high clocks and tend to be somewhat large, others are optimized for low power at lower clocks, and tend to be denser. So the companion core is almost certainly smaller than the other ones. The disparity shown here does seem larger than what I'd expect, but I can't say for sure.

        • Hattig
        • 6 years ago

        It’s probably now using an A7 in a weird big.LITTLE setup with 4 big and 1 little, rather than NVIDIA’s own 4+1 mechanism which uses different characteristics for the +1 core (low power transistors, etc) to make it use less power. Or it’s just how NVIDIA have photoshopped the die image.

        Clearly the die image is photoshopped to bits, we’ll have to wait for chipworks to take a real die photo, maybe late this year when there are actual products using it available to buy.

          • Alexko
          • 6 years ago

          NVIDIA says it’s still 4+1 Cortex A15.

          [url<]http://www.nvidia.com/object/tegra-k1-processor.html[/url<]

    • internetsandman
    • 6 years ago

    I’m glad more SoC’s are moving to 64-bit architectures, but now we gotta see all of the software catch up before the benefits are really felt (and on a mobile platform those benefits will be harder to sell than on a desktop, I’m sure)

      • Voldenuit
      • 6 years ago

      Samsung Galaxy Note 3 already has 3 GB of RAM, so giving more addressable RAM to the OS is already a good thing, even if individual apps are still mostly 32-bit.

        • internetsandman
        • 6 years ago

        There’s more benefits than that though, aren’t there? And with the optimizations in android 4.4 as well as the fact that most apps don’t eat too much memory space, there are bound to be other, more tangible benefits down the line, right?

          • Helmore
          • 6 years ago

          On ARM the main benefit comes from the fact that the AArch64 ISA gives a bit better performance. I don’t know the exact details for why that is, but the ISA is more streamlined in many ways.

            • Beelzebubba9
            • 6 years ago

            AArch64 gives a pretty big speedup for many operations over AArch32:

            [url<]http://anandtech.com/show/7335/the-iphone-5s-review/4[/url<] Considering how impressive Apple's A7 is at ~1.3Ghz, I expect good things from Project Denver at 2.5Ghz.

            • smilingcrow
            • 6 years ago

            That’s a custom core so it’s hard to say how much of the gains are down to Apple’s own skill.

            • Beelzebubba9
            • 6 years ago

            True, but I assume other implementations of ARM v8 will be similarly robust. It’s not like Apple has a long and storied history as a chip designer that they leaned on to make the A7. 🙂

        • smilingcrow
        • 6 years ago

        “Samsung Galaxy Note 3 already has 3 GB of RAM.”

        In the same way that there are many people with 16GB of RAM who probably never use more than 6 to 8GB; they probably upgraded when DDR3 was really cheap.

        I wonder when more than a tiny minority of smart-phone users will actually use more than 4GB of RAM? 2016, 2017!
        It’s better to get the hardware ready in advance but let’s not get carried away with the 64 bit hype trail.

      • MathMan
      • 6 years ago

      In the case of Android, it should be a matter of simply adapting Dalvik?

    • jjj
    • 6 years ago

    “Tegra K1 is built using 28-nm fabrication technology.”

    Have you actually confirmed that the Denver version is on 28nm or just assuming?

      • Dissonance
      • 6 years ago

      Looks like just the Cortex-A15 version so far: [url<]http://www.nvidia.com/object/tegra-k1-processor.html[/url<]

      • Hattig
      • 6 years ago

      For large volume shipments in H1, 28nm is pretty much the only option, 20nm is too new, and only just entering production at TSMC.

      Maybe the 64-bit version in H2 will be using 20nm, as it might be a more niche product that would be a good candidate to test on 20nm. But I think NVIDIA need to perform well with Tegra this year, and they don’t want any risks.

      🙁

      • jdaven
      • 6 years ago

      Availability of Denver is slated for H2 2014 so it is possible that the 20 nm process is used.

Pin It on Pinterest

Share This