Nvidia Jetson TX2 packs Pascal in a pint-size package

The Raspberry Pi Foundation makes some fun and well-supported single-board computers, but sometimes a small embedded system needs more graphical horsepower than the four 300 MHz stream processors in the Raspberry Pi 3 can provide. Nvidia's Jetson TX2 single-board computer module is ready to tackle that situation with its 256 Pascal SPs, six CPU cores, and 8GB of LPDDR4 memory. The machine also packs multiple Pi-pounding connectivity options with built-in USB 3.0, Gigabit Ethernet, Bluetooth, and 802.11ac Wi-Fi.

Anyone who's been paying attention to Nvidia's press conferences of late probably knows that the company and its CEO Jen-Hsun Huang are obsessed with cars. To better fit automobiles, the Jetson TX2 sports a CAN bus interface for communicating with automotive embedded computer systems. The TX2 still packs same the UART, SPI, I2C, and I2S interfaces from its Jetson TX1 forebear. There are also onboard connectors for up to six cameras. The integrated SoC has a pair of Nvidia's Denver 2 ARM CPU cores along with four Cortex A57 cores. The entire Jetson TX2 module measures a compact 2" x 3.5" (5 cm x 9 cm).

The first shipping Jetson TX2 systems will be the development kits. Nvidia didn't offer exact dimensions, but the TX2 development board appears to be about the same size as the TX1, at 6.7" x 6.7" (17 cm x 17 cm). The development boards have all sorts of I/O ports, active cooling solutions, and PCIe x4 slots.

This burly hardware and robust connectivity comes at a cost. The Nvidia Jetson TX2 development board is available for pre-order now for $600. The Pi-sized Jetson TX2 module is scheduled to ship in Q2 2017, and will cost $400 when purchased in quantities of 1000 or more. Nvidia says the existing Jetson TX1 development kit's price will come down to $400 when the TX2 starts shipping. Multiple outlets, including Anandtech, report that Nvidia will offer the TX2 development board to the educational market for $300.

Comments closed
    • willmore
    • 3 years ago

    Always on the lookout for the next SoC for a good tablet. Not this one. *sigh*

    • Shobai
    • 3 years ago

    [quote<]The TX2 still packs same the UART, ...[/quote<] Looks like the order of 'same the' should be reversed.

      • ronch
      • 3 years ago

      OoO execution reorder fail.

    • Meadows
    • 3 years ago

    Impressive, but not a real Jetson until it can fit a flying car in a briefcase.

    • Flying Fox
    • 3 years ago

    If actual implementations in real cars still need that fan, it will be $$ to service it. 🙁

      • Voldenuit
      • 3 years ago

      Hook it into the existing water cooling loop on cars. /jk

    • deruberhanyok
    • 3 years ago

    The price is way out of league for anyone looking into Raspberry pi, obviously, but the potential of all that horsepower is tantalizing!

    • NoOne ButMe
    • 3 years ago

    Interesting fact from Ryan Smith on twitter, it uses an A9 core for audio.

    [url<]https://twitter.com/RyanSmithAT/status/839308894132301825[/url<]

      • chuckula
      • 3 years ago

      Why exactly would they use a general purpose CPU core for audio instead of a DSP like everybody else?

        • NoOne ButMe
        • 3 years ago

        I agree. Just an interesting oddity.

        • tipoo
        • 3 years ago

        Another case of rushed to market/too late to implement?

        • deruberhanyok
        • 3 years ago

        My guess would be so anyone interested in working on audio processing can write whatever code they want and get some decent performance out of it, instead of having to limit to whatever a DSP was capable of doing and then tossing the rest on the main CPU.

          • NoOne ButMe
          • 3 years ago

          Seems they would use an A7 instead, smaller, lower power, presumably cheaper.
          I think Tipoo is probably correct.

            • willmore
            • 3 years ago

            The A9 is higher performing than the A7 by a good amount. The A7 suplanted the A8 as the low power/efficient core. The A9 was supposed to be the higher performance/hotter core.

            • NoOne ButMe
            • 3 years ago

            thanks for the info.
            I think the a7 is enough for audio, and less area helps.

      • ImSpartacus
      • 3 years ago

      That IS pretty fascinating. I always love finding out about extra cores like that.

    • tipoo
    • 3 years ago

    Why A57s on 16nm? There’s not a single area where A72s are worse, die area, power use, performance. Probably just licencing cost?

      • NTMBK
      • 3 years ago

      Because this design has been ready and “on ice” for years, after 16nm got delayed.

      The Tegra roadmap that NVidia showed off in 2013 went straight from 28nm Logan to 16nm Parker, with Parker meant to ship in 2015. Obviously that didn’t happen and we got a rushed 20nm part with no Denver (and non-functional A53 cluster). And now we get a 2015-vintage design released in 2017, hence the A57 cores.

        • tipoo
        • 3 years ago

        Interesting, thanks. Nvidia never seems to have luck launching up to date SoCs with their contemporaries. Granted the A72s wouldn’t be a huge deal with the A57s on the same node.

          • NoOne ButMe
          • 3 years ago

          usage case is automotive anyways, quicker tape out and probably less validation (much A57 already doable on x1?) needed for a market where the extra .5-1.5W per CPU is probably a win.

        • willmore
        • 3 years ago

        There is some good reason to believe this. In early 2014, TSMC and ARM showed a wafer of 16nm A57 cores. Then 16nm crashed and burnt and we got two more years of 28nm. The A57 was too hot at 28nm. Anyone remember the Snapdragon 810?

        By the time 16nm was ready to ship, the *native to 16nm* A72 was out and the A57 was dumped in the bin.

        So, that suggests that this really is a 2014 or 2015 vintage design.

        When did we find out that 16nm wasn’t going to be on schedule? I would suggest the design was from *before* then.

          • ImSpartacus
          • 3 years ago

          Wait, did you mean 20nm?

          I thought 20nm was the tsmc process that kinda shit the bed for anything not mobile and tsmc 16nm is just 20nm with finfets and some additional optimization.

          I just don’t think tsmc 16nm was ready in early 2014.

            • willmore
            • 3 years ago

            16nm Cortex-A57 in early 2014:
            [url<]http://www.theinquirer.net/inquirer/news/2258650/tsmc-tapes-out-arm-cortex-a57-using-16nm-finfet-transistors[/url<] [url<]https://semiaccurate.com/2014/03/03/6-core-16nm-finfet-arm-cortex-a57-chips-spotted-wild/[/url<]

    • chuckula
    • 3 years ago

    Can somebody finagle a way to get a full PCIe 16 card in there?

    I want to run a Vega GPU with my Ngreedia CPU.

      • Goty
      • 3 years ago

      [url<]http://www.jewelrytools.com/images/products/plr-781.00.jpg[/url<] The only question is whether you're up for mutilating the card or the slot!

      • frenchy2k1
      • 3 years ago

      [url<]https://www.newegg.com/Product/Product.aspx?Item=9SIA4SR1R86033[/url<] This would probably work, but you'd need the right drivers... Search for "pcie x4 x16 riser" for other options.

    • NTMBK
    • 3 years ago

    According to Anandtech reporting, the Denver cores are only enabled in 15W mode. Not a good look when your custom CPU core is less efficient than an outdated A57.

      • Goty
      • 3 years ago

      I think it’s a bit murkier than that.

      The Denver cores are only available in the “Max-P” mode, but the TDP is 15W in this mode running EITHER the A57s OR the Denvers at 2GHz or both together at 1.4 GHz. It seems like they’re likely pretty similar at these speeds (at least in TDP), but they may not clock down as well as the A57s. Given that I don’t think the Denver core was ever meant to ship in small mobile devices (i.e. smaller than tablets), I don’t think that’s very surprising.

        • NoOne ButMe
        • 3 years ago

        Interesting, as it seems Denver cores use about 2x the power of A57 cores at ISO clock.

        Wonder how high up it clocks. Denver1 shipped at 2.5Ghz peak. These should be 3W per CPU there. If there is ever a usage requiring high ST performance on chromebook or WART, these will be amazing. Should hit 3Ghz on 16nm

        [url<]http://www.anandtech.com/show/9330/exynos-7420-deep-dive/5[/url<]

Pin It on Pinterest

Share This