Tegra 3: The Quad-Core SoC That Goes To Five
| SoC | Apple A5X | Tegra 3 |
|---|---|---|
| Fab Node | 45 nm | 40 nm |
| Processor | 1 GHz ARM Cortex-A9 (dual-core) | 1.4 GHz ARM Cortex-A9 (quad-core) |
| Graphics | PowerVR SGX543MP4 (quad-core) | ULP GeForce |
| L1 Cache(Instruction/Data) | 32 KB / 32 KB | 32 KB / 32 KB |
| L2 Cache | 1 MB | 1 MB |
Tegra 3 (code name: Kal-El) isn’t particularly new to us. We've already run demos on devices and discussed this architecture prior to today's review. However, the iPad 3's introduction as a rival sparks renewed interest in how Nvidia’s architecture compares to Apple’s A5X.
| GeekBench v2.2.7 Results | iPad 2 | iPad 3 | Dell Mini 1012 | LePan II | Transformer Prime |
|---|---|---|---|---|---|
| CPU | Apple A5 | Apple A5X | Atom N450 | APQ8060 | Tegra 3 |
| Architecture | Dual-core A9 | Dual-core A9 | Single-Core Atom | Dual-core Scorpion | Quad-core A9 |
| Speed | 1 GHz | 1 GHz | 1.66 GHz | 1.2 GHz | 1.4 GHz |
| Overall | 764 | 760 | 917 | 649 | 1194 |
| Integer | 691 | 687 | 910 | 709 | 1781 |
| Floating Point | 921 | 920 | 762 | 943 | 1781 |
| Memory | 830 | 825 | 1105 | 362 | 1091 |
In terms of raw processing potential, Tegra 3 leads the pack by a large margin. Software optimizations and clock rates aside, increased parallelization allows Nvidia's SoC to work on more data concurrently. Similar to the desktop space, adding cores doesn't turn out to have a multiplicative effect on most real-world applications. But an enhanced ability to multitask is nice, especially as resource-hungry background tasks pile up.
Of course, bolstering performance often incurs higher power consumption at the same time. Nvidia, anticipating this, addressed power from a creative angle.
| Kal-El | Companion CPU Core | Main CPU Cores (Symmetric Processing) |
|---|---|---|
| # of Cores | 1 | 4 |
| Function | Power-Optimized (Standby) | Performance |
| Architecture | Cortex A9 | Cortex A9 |
| Process Technology | Low Power (LP) | General |
| Operating Frequency | 0 MHz to 500 MHz | 0 MHz to Max Frequency |
Kal-El sports a fifth "companion" CPU core that operates at lower frequencies and handles background tasks like syncing email, playing ringtones, and keeping applications alive while the device is in standby mode. It's hard to quantify the exact benefit of Nvidia's implementation since there aren't any Tegra 3s that lack the fifth core. However, the company's engineers clearly felt strong enough about its effect (particularly coupled with low-power silicon) that they were willing to dedicate precious die space to what was considered a power-optimized design.
Borrowing a page from Qualcomm's book, Nvidia employs an asymmetrical clock scheme that's similar to Turbo Boost, except that it allows the companion core to operate at a different frequency. It also incorporates Advanced SIMD (called NEON), which lets the CPU perform certain tasks (like playing MP3 audio) at extremely low CPU speeds, generally between 10-20 MHz. Qualcomm made a name for itself using a similar design, and the result is a processor with very low power consumption that can deliver performance when it's needed. Read Third-Generation Snapdragon: The Dual-Core Scorpion for more information on Qualcomm's solution.
There’s every reason to believe that this hybrid approach should work well. However, realizing gains with this approach depends on Nvidia to work within the constraints of operating system design. Purposely, Tegra 3 doesn't expose the fifth CPU core to the OS. Rather, it operates in the background without any management from the operating system. That means “low-overhead tasks” have to be identified by the hardware and handled by its companion core.