More Details

Looking at the specifications, the standout change is the new 4+4 processor configuration, featuring four powerful ARM Cortex-X4 cores and four high-performance Cortex-A720 cores. Unlike traditional mobile chipsets used in Android phones, including the latest Google Tensor G3 and Snapdragon 8 Gen 3, the less powerful and energy-efficient ARM Cortex-A5XX series cores are absent at all. Surely this can’t be better for battery life? Well, actually, everything could be fine.

To understand why, we dug up a slide from Arm’s A720 announcement. Not only did the presentation point to a smaller core count for mobile devices, as is the case with the 8 Gen 3, but a closer look at Arm’s DVFS curve indicates that the A520 has a lower minimum power/performance ratio while the A720 runs closer to the maximum than the A520. Consumes less energy in the condition. The question then becomes whether the A520 will be able to realize any power benefits for real-world workloads that are becoming richer and more complex, or whether it will actually be just as well or more efficient to run the A720 in lower power conditions.

MediaTek’sFinbarr Moynihan provided further explanation, noting that it may be more efficient to power up a more powerful core, quickly complete a task, and then power down, rather than running for longer on a less powerful core. Google has provided similar logic in Core Tensor and G2 for its dual Cortex-X1 approach. However, since then the Tensor G3 has used a more traditional single large core. MediaTek backs this up with numbers, claiming that “normal” tasks such as web browsing and social apps, as well as more demanding applications, offer a 10-15% reduction in power consumption compared to the Dimensity 9200.

The Dimensity 9300 still uses a sort of three-tier processor system. A large Cortex-X4 core operates at up to 3.25 GHz, while the other three are clocked at up to 2.85 GHz. MediaTek notes that there is no difference between these cores in terms of cache, only that the higher clocked cores have more silicon area to accommodate the higher frequencies. The four Cortex-A720 cores are identical, each with a peak clock speed of just 2.0 GHz, indicating a more energy-efficient implementation of these cores than other chipsets. For comparison, the Snapdragon 8 Gen 3 overclocks the A720 core to 3.2 GHz.

MediaTek’s processor adds 8MB of L3 cache and 10MB of system cache, increasing the cache size by 29% compared to last year. Together, the Dimension 9300 delivers 15% more performance for the same power, or 33% less power for the same performance, as the Dimension 9200. Alternatively, at maximum usage, the CPU configuration delivers 40% more peak performance than last year’s model. . This all sounds promising, but we’ll definitely be keeping an eye on standby power consumption and temperatures with this new design.

It’s impossible to launch a chip in 2023/2024 without mentioning AI capabilities, and MediaTek has some equally significant changes here. The Dimension 9300 comes with an updated 790 APU, which doubles integer and floating point capacity and reduces power consumption by 45%. It is very useful for a variety of machine learning tasks, from voice recognition to image segmentation.

In keeping with the trend, the Dimension 9300 includes specific enhancements to on-device generative AI, which MediaTek claims results in an eight-fold improvement in performance over the previous generation. The problem with generative AI models like LLM is that they require a lot of memory, which makes it much more difficult to run on a phone than on a cloud server. To get around this problem, the 790 APU supports INT4 (A16W4) for running small quantized models and a dedicated hardware memory decompression unit that feeds the APU. In MediaTek’s example, the 13GB INT8 model can be pre-compressed to 5GB to fit into RAM and then decompressed by hardware en route to the APU.

In terms of applications, the 790 APU can process LLMs with 7 billion parameters at 20 tokens per second, which is fast enough for real-time use. By comparison, Qualcomm says its Snapdragon 8 Gen 3 can process an LLM with 10 billion parameters at about 15 tokens per second, which seems pretty comparable. The Dimension 9300 can extend this capability to run LLM with 13 billion parameters on 16 GB RAM, up to 33 billion parameters with 24 GB RAM, albeit at a much slower 3-4 tokens per second.

While compact models won’t be as accurate as their data center counterparts, the prospect of running generative AI on a single device, with all the security and throughput benefits that entails, is now a real possibility. Seen in initiatives like the Gemini Nano and Samsung’s Galaxy AI suite It is important to note that MediaTek fully expects its partners to take advantage of these opportunities immediately.

Processors and artificial intelligence are two big pieces of the latest dimensionality pie, but there are plenty of toppings as well. Gamers, for example, are offered the latest Arm Immortalis-G720 GPU in a 12-core configuration from the latest generation 11-core processor). MediaTek estimates a peak performance increase of up to 23% over the 9200, combined with improvements in architecture and technology nodes, as well as up to 46% improvement in ray tracing. Perhaps even more enticing is the prospect of a 40% reduction in power consumption while maintaining the same level of performance as last year’s model, resulting in much longer gaming sessions on a single charge.

The new GPU tweak includes support for global illumination effects, 40% memory bandwidth savings for geometry-heavy games, and 2x MSAA support without the performance penalty of 4x MSAA (the lowest AA level previously supported on the G715). In terms of display, WQHD panels with refresh rates up to 180Hz or 4K at 120Hz are now supported, as well as Dual Active Display for foldable devices and Google’s Ultra HDR display format from Android 14.

For photographers, MediaTek’sImagiq 990 ISP is tightly coupled to the APU, supporting up to 16 levels of object segmentation even when recording video. However, ISP supports continuous HDR shooting at 4K resolution, as well as software bokeh and depth effects. A hardware-based dual-lossless upscaling engine reduces quality when capturing frames from high-resolution sensors, and the ISP now has a dedicated OIS sensor core instead of sharing resources with sensor hubs or other components.

The Dimensity 9300 features a new secure boot chip, an isolated secure computing environment and an Armv9 memory tagging extension that helps developers avoid memory exploitation and improve the security of future smartphones. Network updates are more repetitive. You get native support for Wi-Fi 7 and sub-6GHz 5G bands for downlink speeds of up to 7Gbps.

With the devices in hand, we can now see where the MediaTek Dimension 9300 chip stands out from the competition. As expected based on the specifications, MediaTek’s flagship chip offers solid processor capabilities. Although it doesn’t quite match the Apple A17 Pro in terms of power, its powerful processor narrowly beats the Qualcomm Snapdragon 8 Gen 3 in multi-core workloads.

Archrival Qualcomm takes the lead in graphics workloads, with the Snapdragon 8 Gen 3 holding a solid lead in one-shot tests (including ray tracing) and generally outperforming long-term workloads in our tests. Not that Dimension fails when it comes to that. For gaming its GPU handles the demands of modern games without any issues and significantly outperforms Apple A17 Pro and Google Tensor G3.

The MediaTek Dimension 9300 based Android smartphones were announced in late 2023 and reached customers in early 2024. The high-end vivo X100 Pro is perhaps the best example of a global flagship powered by the latest and greatest MediaTek technology, while several China-only phones also use the processor, such as the OPPO Find X7.