In the past few years, we’ve seen NVIDIA shift their mobile strategy dramatically with time. With Tegra 2 and 3, we saw multiple design wins in the smartphone space, along with the tablet space. These SoCs often had dedicated GPUs that were quite unlike what we saw in NVIDIA’s desktop and laptop GPUs, with a reduced feature set and unique architecture. However, with Tegra K1 we saw a distinct shift in NVIDIA’s SoC strategy, as the Tegra K1 was the first mobile SoC to achieve parity in GPU architecture with desktop GPUs. In the case of the Tegra K1, this meant a single Kepler SMX which made for truly incredible GPU performance. However, in the time since we’ve seen companies like Apple release new SoCs such as the A8X, which managed to bring largely similar performance with less power.

NVIDIA of course is never content to stay idle for too long and let themselves be surpassed, which has led to the company developing their Tegra SoCs on a roughly yearly cadence. In NVIDIA’s development timeline, 2014 brought us the Tegra K1-32, the company’s first SoC to integrate a modern high-performance GPU architecture (Kepler), and later on the TK1-64, a K1 with NVIDIA’s own in-house developed ARMv8 Denver CPU in place of the 32bit ARM A15 CPU.

NVIDIA's GTC 2014 Tegra Roadmap

2014 also brought a revision to NVIDIA’s public Tegra roadmap: for 2015 NVIDIA would bump the previously announced Parker SoC and release a new in its place, Erista. This schedule change ended up being quite a surprise due to the fact that NVIDIA had up until this time released their roadmaps roughly 2 years out, whereas Erista was put on the map less than a year before it was to be announced. More unusual, NVIDIA offered up minimal details of Erista; it would be powered by a Maxwell GPU with no additional information on the CPU or manufacturing process. Parker on the other hand – NVIDIA’s planned Denver + Maxwell + 16nm FinFet part – fell off the schedule, with NVIDIA not officially canceling it but also saying little else about it.

Now in 2015 and with the launch of the Tegra X1, we can finally begin putting the picture together. Erista as it turns out is something of a rapid release product for NVIDIA; what had been plans to produce a 16nm FF part in 2015 became plans to produce a 20nm part, with Erista to be that part. To pull together Erista NVIDIA would go for a quick time-to-market approach in SoC design, pairing up a Maxwell GPU with ARM Cortex A57 & A53 GPUs, to be produced on TSMC’s 20nm SoC process.


We’ll kick off our look at Erista then with a look at Erista’s CPUs, where one of the bigger surprises in Erista is the CPU configuration. While Tegra K1 had a version with NVIDIA’s custom Denver core for the CPU, NVIDIA has elected to use ARM’s Cortex A57 and A53 in the Tegra X1. The A57 CPU cluster has 2MB of L2 cache shared across the four cores, with 48KB/32KB L1s (I$+D$) per core. The A53 cluster has 512KB of L2 cache shared by all four cores and 32KB/32KB L1s (I$+D$) per core. NVIDIA representatives stated that this was done for time to market reasons.

However, rather than a somewhat standard big.LITTLE configuration as one might expect, NVIDIA continues to use their own unique system. This includes a custom interconnect rather than ARM’s CCI-400, and cluster migration rather than global task scheduling which exposes all eight cores to userspace applications. It’s important to note that NVIDIA’s solution is cache coherent, so this system won't suffer from the power/performance penalties that one might expect given experience with previous SoCs that use cluster migration.

Although cluster migration is usually a detriment to power and performance (especially without a cache coherent interconnect), NVIDIA claims that Tegra X1 significantly outperforms Samsung System LSI’s Exynos 5433 in performance per watt with 1.4x more performance at the same amount of power or half the power for the same performance on the CPU. It's likely that this difference comes from optimizing the implemented logic and differences between Samsung and TSMC 20nm processes rather than differences in thread management. NVIDIA continues to use System EDP (Electrical Design Point) management to control throttling and turbo rather than ARM's IPA (Intelligent Power Allocation) drivers, and in general it seems that while the CPU architecture is ARM's reference design it's quite clear that everything else is distinctly NVIDIA in design.

All of this of course raises the question of where’s Denver, and what is to happen to it? The use of the ARM Cortex A57 and A53, as NVIDIA tells it, was based on a time-to-market decision, and that NVIDIA could bring an off-the-shelf Cortex-based SoC to the market sooner than they could another Denver SoC. On the surface there is at least some evidence to agree with this, as NVIDIA would be able to buy a 20nm design from ARM versus spending the time porting Denver over from TSMC’s 28nm process to their 20nm process. At the same time however integrating an ARM CPU core into an SoC is not an easy task – logic LEGO this is not – so it’s hard to say whether this actually saved NVIDIA much time.  

In any case, much like the still in the dark Parker design, NVIDIA is holding their cards close to their chest and hinting that Denver will still be in future generation products. If Parker is still part of NVIDIA’s plans – and they are without a doubt working on some kind of 16nm FF Tegra SoC – then we may still see Denver as part of the original Parker plan. Otherwise at this point while there’s no reason not to believe NVIDIA about the fate of future Denver SoCs, it’s anyone’s guess when we might see it again.


Outside of the CPU and GPU, NVIDIA has also dramatically improved the rest of Tegra X1 in comparison with Tegra K1. We see a move from 64-bit wide LPDDR3 to 64-bit wide LPDDR4 on the memory interface, which improves peak memory bandwidth from 14.9 GB/s to 25.6 GB/s and improves power efficiency by around 40%. In addition, the maximum internal display resolution moves from 3200x2000 at 60 Hz to 3840x2160 at 60 Hz with support for VESA’s display stream compression. For external displays, we see a significant improvement with support for HDMI 2.0 and HDCP 2.2, which means that 4K60 is supported in contrast with the Tegra K1, which only supported 4K30.

The ISP is largely similar to the one we see in the Tegra K1 when it comes to feature set, but JPEG encode and decode rate is now five times as fast, going from 120 MP/s to 600 MP/s. For video encode and decode, we see support for 4K60 H.265 and VP9, with support for 10 bit color on decode on H.265. In addition to support for new codecs in the hardware blocks, we see that H.264 and VP8 now support 4K60, a notable improvement over the 4K30 that was the maximum for Tegra K1. We also see that the storage controller now supports eMMC 5.1 for faster storage performance.

Tegra X1's GPU: Maxwell for Mobile
Comments Locked


View All Comments

  • harrybadass - Monday, January 5, 2015 - link

    Nvidia X1 is somehow already obsolete when compared to A8x.

    Clusters 8
    FP32 ALUs 256
    FP32 FLOPs/Clock 512
    FP16 FLOPs/Clock 1024
    Pixels/Clock (ROPs) 16
    Texels/Clock 16
  • psychobriggsy - Monday, January 5, 2015 - link

    NVIDIA are claiming power savings compared to the A8X, at the same performance level.

    And additionally, they can run the X1 GPU at ~1GHz to achieve greater performance than the A8X. However the A8X's lower GPU clock is just a design decision by Apple so they can guarantee battery life isn't sucky when playing games.

    But yet, hardware-wise the X1's GPU specification isn't that amazing when compared to the A8X's GPU.

    Last up, how does a quad-A57 at 2+ GHz compare to a dual 1.5GHz Cyclone...
  • techconc - Monday, January 5, 2015 - link

    Isn't always amazing how company A's future products compete so well against company B's current products? The X1 won't be competing with the A8X, it will be competing against the A9X. If you're familiar with the PowerVR Rogue 7 series GPUs, you'd wouldn't be terribly impressed with this recent nVidia announcement. It keeps them in the game as a competitor, but they will not be on top. Further, I'm quite certain that Apple's custom A9 chip will compare well to the off the shelf reference designs in the A57 in terms of performance, efficiency or both. If there were no benefits to Apple's custom design, they would simply use the reference designs as nVidia has chosen to do.
  • Yojimbo - Monday, January 5, 2015 - link

    Yes but how do you compare your product to something that isn't out yet? You can't test it against rumors. It must be compared with the best of what is out there and then one must judge if the margin of improvement over the existing product is impressive or not. The PowerVR Rogue 7 series is due to be in products when? I doubt it will be any time in 2015 (maybe I'm wrong). When I read the Anandtech article on the details of IMG's upcoming architecture a few months back I had a feeling they were trying to set themselves up as a takeover target. I don't remember exactly why but it just struck me that way. I wonder if anyone would want to risk taking them over while this NVIDIA patent suit is going on, however.
  • OreoCookie - Tuesday, January 6, 2015 - link

    The Tegra X1 isn't out yet either!
    If you look at Apple's product cycle it's clear that in the summer Apple will release an A9 when they launch the new iPhone. And you can look at Apple's history to estimate the increase in CPU and GPU horsepower.
  • Yojimbo - Tuesday, January 6, 2015 - link

    But NVIDIA HAS the Tegra X1. They are the ones making the comparisons and the Tegra X1 is the product which they are comparing! Apple seems to be releasing their phones in the fall recently, but NVIDIA nor the rest of the world outside Apple and their partners has no idea what the A9 is like and so it can't be used for a comparison! It's the same for everyone. When Qualcomm announced the Snapdragon 810 in April of 2014 they couldn't have compared it to the Tegra X1, even though that's what it will end up competing with for much of its life cycle.
  • Yojimbo - Monday, January 5, 2015 - link

    Perhaps those are the raw max-throughput numbers, but if it were that simple there would be no reason for benchmarks. Now let's see how they actually perform.
  • edzieba - Monday, January 5, 2015 - link

    12 cameras at 720p120?! VERY interested in DRIVE PX, even if it'd never end up near a car.
  • ihakh - Monday, January 5, 2015 - link

    about the intel chip I have to say that it is a very good CPU (think about sse and avx) + a little GPU
    but nvidia chip is a good GPU+ reasonable CPU

    you can have windows x86 on intel chip and run something like MATLAB (also android)
    and you can have a good gaming experience with nvidia's

    each of them has its use for certain users
    its not like that every program can use 1TFLOPS of tegra GPU
    and its not like every user is "game crazy"
    intel core M have its own users

    and of course tegra chip is very hot for mobiles and it is a hard decision for engineers who design mobiles and tablet to migrate from a known chip like snapdragon to an unknown and new chip like tegra

    I think both nvidia and intel are doing good and nor deserve blaming
    but it is a good idea for nvidia to make a cooler chip for mobiles
  • Morawka - Monday, January 5, 2015 - link

    So compared to the K1 it's twice as fast, And it also uses Twice as less energy.

    So does that mean it will still be a 7w SOC? albeit twice as fast.

Log in

Don't have an account? Sign up now