Nvidia unveils 64-bit Denver-soc architecture

Nvidia has released more details of its upcoming Denver soc. This 64-bit ARMv8-soc will, among other things, have access to a mechanism called dynamic code optimization. According to Nvidia, the Denver soc is the fastest 64bit processor for Android hardware.

Plans for a 64-bit version based on the Tegra K1 processor were revealed at CES in January. In the meantime, Nvidia has announced more details about the architecture of the dual-core Denver soc in a blog posting and white paper.

What is special about the design is that no use is made of out of order execution, as is the case with almost every modern processor architecture, but of in order execution. OoO execution does not always process instructions in the order they are received. In cases where two consecutive instructions have to wait for each other, the processor can give priority to independent instructions, which improves efficiency. Modern socs from Samsung, Qualcomm and Intel, among others, use OoO execution.

To achieve comparable results, Nvidia uses software routines that analyze the incoming instructions and associated branch results and convert them into optimized microcode. The optimized instructions are then placed in a 128MB block of working memory, while a small cache on the chip contains a look-up table, allowing the processor to quickly find the optimized code.

Thanks to this technique, called dynamic code optimization by Nvidia, performance is achieved that is comparable to an out-of-order architecture, without the need for energy-consuming fix function logic in the chip design, according to Nvidia. An additional advantage, according to Nvidia, is that its software solution can analyze a larger set of instructions at once than the hardware in an OoO processor.

The design does not come without drawbacks. For example, the software optimization of the instructions results in overhead and in addition, due to the addition of extra cache, the Denver core is a lot larger than the Cortex A15 core in the 32-bit version of the K1 soc. Where that version of the chip manages to accommodate five Cortex A15 cores, only two Denver cores fit on the same surface with the 64-bit variant. The chips are also pin compatible.

Nvidia claims that the Denver-soc with a clock speed of 2.5GHz delivers performance that is equal to or better than that of Apple’s A7-soc. According to Nvidia, Celerons from Intel’s Haswell series can also be kept up with IPC in some tests. This would make it the fastest 64-bit ARM soc for the Android market. The first products in which the 64-bit version of the Tegra K1 soc is available should be available later this year.