-->

NVIDIA Jetson AGX Investigation: Views of Xavier and Carmel Cores

Today's work is a bit of an unusual review; NVIDIA's new Jetson AGX embedded system kit is not really a device platform that we expect the average reader to think about, let alone buy. NVIDIA's shift over the past few years from offering consumer-level Tegra chipsets to more specialized silicon applications is no more clear than the new Tegra Xavier that supports Jetson AGX. Although the capabilities of the board are clearly beyond the use of most consumers, it is still a very attractive platform with many functions and silicon IPs that we have not found on other devices to this day. So when NVIDIA reached out to offer us a sample, we decided to try to make a high-level overview of what the new board and Xavier chips could do.




First of all, we must describe what this really is. Jetson AGX is a small computer with a small form factor / embedded system, with the shape of the entire unit no larger than 105x105mm. The AGX module itself is designed to be a complete commercial system of shelves (COTS) for use in finished products, with NVIDIA directing it to cases of centric AI use (read: neural networking) such as robotics and industrial automation. Jetson boards usually occupy the small to medium volume end of the market, appearing in disposable products and limited production items, which makes no sense for manufacturers to develop and use their own special hardware.

But of course the naked module is only half of the story. You cannot develop nude modules, and this is where NVIDIA's complete Jetson AGX development kit is. The AGX dev kit is equipped with everything needed to run one module, including power supply, heatsink, and more importantly of all, a breakout board. Escape boards offer a variety of I / O headers and ports, ranging from your standard dual USB-C 3.1 ports, HDMI connectors and Gigabit Ethernet ports, to more specific connectivity such as the MIPI CSI-2 connector for camera connectivity and various typical header board developments such as 40 pin GPIO connector.



The more unusual connectivity options of the Jetson AGX are the PCIe Gen4 x16 slot and the M.2 PCIe x1 extension slot that are intended to be used for add-on connectivity such as WiFi or cellular modules, both features that are not common among Arm development boards because most SoCs does not have a PCIe backup controller.



The board is equipped with many other connectors, and that is one thing where the new AGX Jetson does not lack flexibility. Power is supplied by an external generic 19V power supply - the stock supplied by NVIDIA is the LiteOn 65W unit which seems no different from most laptop charger bricks.

NVIDIA's Jetson AGX Investigation: A View of Xavier and Carmel Cores

Under the heavy and dense aluminum heatsinks we found what really strengthened the AGX Jetson board: the XXX AGX module. This is a system module that is above the Jetson motherboard - the module does not have its own I / O port and only functions as a system brain, integrating core components around the Xavier chip, such as GBDDR4x 16GB memory. , a small 32GB eMMC storage chip and all power delivery circuits to power different DRAM power rails and IP blocks from Xavier SoC

NVIDIA's Jetson AGX Investigation: A View of Xavier and Carmel Cores

Chip Xavier, as said, is the brain platform and represents the biggest and most complex SoC at NVIDIA to date. With 9 billion transistors on a 350mm² die size, this is one of the heavy weights of the Arm ecosystem, though between the initial announcement and today Apple has NVIDIA one-up in terms of the number of transistors as the new A12X is a 10B chip - in manufacturing knots that are far more small.

NVIDIA's Jetson AGX Investigation: A View of Xavier and Carmel Cores

Originating from the traditional PC industry, NVIDIA did not hesitate to show photos of their products, which is something very rare lately among Arm SoC vendors. Xavier SoC is mainly dominated by two large IP blocks consisting of a large portion of the space allocated to die: "Carmel" 8-core CPU Complex and Volta four-cluster GPU.

NVIDIA's Jetson AGX Investigation: A View of Xavier and Carmel Cores


Originating from the traditional PC industry, NVIDIA cannot display photos of their products, which is something very rare lately among Arm SoC vendors. Xavier SoC is largely determined by two large IP blocks which comprise most of the space allocated to die: "Carmel" 8-core CPU Complex and Volta four-cluster GPU.



The GPU at Xavier is rooted in Volta's architecture. Here we find GPUs configured into four TPC (texture processing clusters), each with two SM (multiprocessor streams), for a total of 8 BC or 512 ALU / CUDA core paths. The most interesting aspect of the GPU is that it is based on Volta, the GPU also inherits the Tensor processing unit from its larger siblings. This adds to the total GPU processing power up to 22.6 8-bit TOPs or 11.3 TOPS FP16 in the core Tensor, above 2.8 and 1.4 TFLOP respectively for the CUDA FP16 and FP32 operations provided by SM.




Along with DLA, programmable vision accelerators are also a key component of the Xavier system that enables it to focus on vision and specifically robotics, embedded AI, and automotive casings. PVA is a more traditional IP vision block that handles simpler tasks such as object detection in ways that are far more efficient than those that can be handled by GPUs or algorithms that conclude machines. Here the PVA will be the first IP block after the ISP in the vision pipeline will function to divide the image parts into objects which will then be forwarded to other algorithms which will then occur on the GPU or DLA.


Also Read More :

NEXT ARTICLE Next Post
PREVIOUS ARTICLE Previous Post
NEXT ARTICLE Next Post
PREVIOUS ARTICLE Previous Post

Composure

 

Delivered by FeedBurner

-->