Today NVIDIA has officially released Maxwell Tegra X1 processor architecture, based on 20nm Process node. The NVIDIA Tegra X1 will work in a Drive PX dedicated module that will be provided for cars only, but two Tegra X1 Chips, for an impressive total of upto 2.3 TeraFlops of computing power on FP16.
Before about it even Internet-related reports are too small, because NVIDIA refused to disclose any information related to it and the media has also no data to say. Xiao Bian has collected a little information available. In this post we will look at the preliminary Tegra X1 architecture.
Two Tegra X1s working in parallel to power the Drive PX Module – 20nm is here
The first is the GPU, upgrade on the Maxwell architecture, a total 256 CUDA cores, because Maxwell architecture SMM unit reorganization, 256 CUDA cores has just two groups SMM unit. And previously we have seen that 192 CUDA core has been used in Tegra K1 which is actually a group of SMX unit, so that although the number of CUDA cores has been upgraded to that level, but X1 offers two groups SMM K1 cells, the performance boost is derived from this.
Let’s look at CPU architecture, NVIDIA here, try to put the most mysterious part, look at the number of cores NVIDIA does not want to lose the war, and to win the war of SoC Tegra X1 is featuring 8-core architecture, basically eight cores 64-bit architecture. These 8-core how they plot is very interesting. We know that Denver 64 is NVIDA’s own architecture, but the K1 only uses a dual-core design, which may be cautious in selecting NVIDIA control risk. It also means that the next 8 Denver 28nm process provides the core which is too crazy to choose, almost impossible, this self-study core hanging out.
For the 8-core architecture, you can determine 4 + 4, it is very easy to think big. LITTLE ARM architecture, NVIDIA has really abandoned its Denver architecture, these 64-bit 8-cores are basically based on ARM architecture and distributed in actually the four Cortex-A57 and four Cortex-A53 cores, same as Qualcomm Snapdragon 810.
Regardless of this kind of design, in fact, 8-core 64-bit CPU + Maxwell GPU performance will be a massive force. If we talk about the Graphics. In GPU part NVIDIA runs UE4 engine as an example to demonstrate the power of their chip. Previously we know that from NVIDIA Tegra 1, NVIDIA does such practices and every time they show how smooth performance is.
Other aspects, Tegra X1 can support 4K 60Hz with 10bit VP9 / H.265 decoding, video capability abnormally tough.
In addition, the NVIDIA Tegra X1 represent floating point performance to TFLOPS, first to achieve such a high-level of TeraFlops numbers on mobile SoC, but two years ago, when the Imagination PowerVR 6 release said their mobile GPU architecture, performance of up 1TFLOPS up, but floating-point performance of mobile SoC is actually very weak, to the 100-200GFLOPS on very high. Huang claimed that Tegra X1 floating point performance is two times better if you compare it with NVIDIA Tegra K1
If we talk about TDP, we know that NVIDIA domestic official micro chips does not exceed 15W. But now things have changed as the power consumption is about 10W, and this candy is kind enough to do a comparison with current Gen-Console from Microsoft Xbox One. Xbox One Consume 100W, while Tegra X1 gaming performance is comparable to consoles.
There is one question about the process node, although Huang did not disclose this secret, but Tegra X1 is indeed based on a 20nm process node.
Here I am leaving you with some words on the strength of this SoC. NVIDIA each conference is not out to kill the opponents NVIDIA has simply its own style, in the contrast we can compare it with an Apple A8X processor, but Tegra X1 can easily beat Apple A8X, at least 1.5 times better performance than A8X.
Huang also announced that NVIDIA Tegra X1 is for automotive electronics Drive CX, which is based on Tegra X1 digital car cockpit systems, not simply in-car entertainment / navigation system but It will also take care of various functions and actually very powerful.