New 20 nm chip packs 256 Maxwell-based GPU cores and an A53/A57 octacore design

NVIDIA Corp. (NVDA) has announced the upcoming availability of the Tegra X1, a sixth generation mobile system-on-a-chip (SoC) design.  Developed under the codename "Erista", the Tegra X1 succeeds last year's underappreciated, overperforming Tegra K1 (codename: "Logan").

I. Tegra K1 and K1 Denver -- Promising Cores, Poor Adoption

In a market dominated by Qualcomm, Inc. (QCOM), NVIDIA has struggled in recent years to carve out a niche for its Tegra SoC.  Tegra always sounded promising on paper, thanks to the integration of NVIDIA's GeForce graphics technologies.  But ultimately recent Tegra SoCs underperformed on both the battery life and processing power front.
Tegra Shield
NVIDIA's own Shield tablet was one of the few devices to use the Tegra K1.

Tegra K1 on paper was very powerful, but was virtually ignored by OEMs.  In total it appeared in only four devices: Other than the ThinkVision monitor, these were all relatively high volume products, but overall chip volumes still were a far cry from Qualcomm's ubiquitous Snapdragon series.  Late last year NVIDIA introduced a new 64-bit Tegra K1 that with a new 64-bit proprietary core design developed under the codename "Project Denver".

Tegra K1
NVIDIA's Tegra K1 Denver was a promising design by like the original Tegra K1 saw weak adoption.
[Image Source: Android Central]

Denver appeared to be a terrific design.  While at times a bit mercurial, earning low marks in a handful of benchmarks, in most tests the Denver-based Tegra K1 found in HTC Corp.'s (TPE:2498) Nexus 9 tablet, was the top the performer in many graphics benchmarks.

Overall, the new Tegra K1 was a very fit challenger that truly was capable of vying with Apple, Inc.'s (AAPL) proprietary A8X SoC (found in the iPad Air 2) and in most cases blowing by Qualcomm's Snapdragon 805 (found in the Samsung Electronics Comp., Ltd. (KRX:005930) (KRX:005935) Galaxy Note 4 and other devices).

II. Tegra X1 -- Supercomputer in the Palm of Your Hand (or in Your Car)?

The new Tegra X1 has some big shoes to show after the strong showing with the second-generation Denver-based Tegra K1.  The Tegra X1's trump card is the culmination of its strong graphics heritage -- a GeForce on-die graphics processor with 256 Maxwell CUDA cores -- a third more cores than the Tegra K1's (192 CUDA cores).

Tegra X1

NVIDIA claims that with the new GPU configuration, the Tegra X1 will offer "teraflops" of general purpose GPU (GPGPU) computing power and double the performance of the previous generation.

Having failed to gain traction in the consumer mobile device market, NVIDIA seems to be trying a new tact, promoting its chip as an attractive option for the automotive space.  With Texas Instruments Inc. (TXN) and the Marvell Technology Group Ltd. (MRVL) winding down their automotive SoC, both NVIDIA and Qualcomm are vying for this growing market.  Qualcomm is expected to announce shortly a refresh to the automotive-aimed Snapdragon 602A, which was announced nearly a year ago. 

The company CEO and cofounder, Jen-Hsun Huang, asserts that the Tegra X1 is more than a capable competitor in the automotive space, bragging:

We see a future of autonomous cars, robots and drones that see and learn, with seeming intelligence that is hard to imagine.  They will make possible safer driving, more secure cities and great conveniences for all of us.

To achieve this dream, enormous advances in visual and parallel computing are required. The Tegra X1 mobile super chip, with its one teraflops of processing power, is a giant step into this revolution.

NVIDIA makes the comparison that its new chip is roughly as powerful as the ASCI Red supercomputer at U.S. Department of Energy's (DOE) Sandia National Laboratory (SNL).  It brags that its new chip draws only 10 watts, where ASCI Red sucked down 500,000 watts of power, plus another 500,000 watts for cooling (not to mention occupying 1,600 sq. ft.).

Tegra X1

This comparison is catchy, certainly, but also a tad bit misleading, given the limitations of NVIDIA's small-core computation approach.

Built in the mid-1990s ASCI Red was built on Intel Corp. (INTC) x86 CPUs.  Hence it was a multiple instruction multiple data (MIMD) design, where as the Tegra X1 is a single instruction multiple data (SIMD) design.  In some applications (particle simulations or graphics acceleration) the teraflop SIMD design will truly be comparable, but in many applications the MIMD teraflop-scale machine is far more flexible -- part of why ASCI Red was kept operational until 2006.

In a single application scenario, the Tegra X1 certainly might beat ASCI Red in some cases.  But in a multi-user, multi-application scenario ASCI Red -- who boasted over 4500 nodes each with a pair of (single-core) 333 MHz Pentium II processors and 256 MB of memory -- would obliterate the Tegra X1 (obviously).

III. Mobile Market Outlook

In the traditional mobile market the Tegra X1 will face some stiff competition -- namely the Qualcomm Snapdragon 810, which appears to be headed for a Q2 or Q3 2015 release.

Both chips share several similarities.  First off, they're both manufactured on Taiwan Semiconductor Manufacturing Comp., Ltd.'s (TPE:2330) (TSMC) new 20 nanometer (nm ) process -- although Qualcomm is sourcing some production to Samsung's 20 nm lines, as well.

Both also take a step back from their predecessors, dropping more fine-tuned proprietary core designs in favor of ARM Holdings plc's (LON:ARM) licensable ubiquitous Cortex intellectual property (IP) cores.  Both the Snapdragon 810 and the Tegra X1 are big.LITTLE octacore designs with a quartet of Cortex-A53 cores (the "LITTLE" ones) and a quartet of more power Cortex-A57 cores (the "big" ones).


In NVIDIA's case the Cortex-A57 cores are given 2 MB of L2 cache to work with, while the Cortex-A53 are given a leaner 512 KB cache.  The cores are on isolated rails and support "fourth generation cluster switching".  Both technologies should help NVIDIA be more competitive with Apple and Qualcomm on the power consumption front.

NVIDIA's best differentiation is on the GPU front.  

Its new chip supports a variety of cutting edge graphics techniques introduced with Maxwell, including Voxel Global Illumination (VXGI) and Multi-Frame Anti-Aliasing, (MFAA).  It also supports 1.3 gigapixels of video throughput, which can either be used to support 4K video recording or to process up to 12 separate lower resolution camera streams for automotive applications.

NVIDIA has two automotive platforms that are debuting based on the Tegra X1:

NVIDIA will be selling so-called "DRIVE Computers", which are basicaly prepackaged systems with a Tegra X1 SoC and supporting chips like memory and I/O communications integrated circuits (ICs).  NVIDIA is offering a developer platform called "Jetson Pro" to help automakers plug in the new Tegra X1-based DRIVE computers.

NVIDIA sells "DRIVE" computers, which can double as infotainment or self-driving hardware.

On the mobile front, Tegra supports:
  • Acceleration of 4K video codecs
  • Graphics and Compute APIs
    • Unreal Engine 4
    • DirectX 12
    • OpenGL 4.5
    • CUDA
    • OpenGL ES 3.1
    • Android Extension Pack
It should be interesting to see how the Tegra X1 does adoption-wise both on the mobile and the automotive fronts.

On the one hand the competition hasn't gotten any easier and it still lacks an on-die LTE chipset (a key advantage held by Qualcomm).  On the other hand, it has one of the fastest mobile processors on the market and may have the best available graphics package on an ARM-driven chip.  Those strengths make the Tegra X1 a real threat to Qualcomm and Apple -- on paper, at least.

Sources: NVIDIA [press release], [product page]

"The Space Elevator will be built about 50 years after everyone stops laughing" -- Sir Arthur C. Clarke

Copyright 2017 DailyTech LLC. - RSS Feed | Advertise | About Us | Ethics | FAQ | Terms, Conditions & Privacy Information | Kristopher Kubicki