Dimensity 9600 Pro Explained: Everything MediaTek Packed Into Its New 2nm Flagship

 

Every year a new flagship chip shows up, every year the slides say "faster and cooler," and every year most of us shrug and wait for the actual phones. So let me save you the slide deck. Here's what the Dimensity 9600 Pro actually is, in plain language, and which parts of it you'll genuinely feel.

Short version: it's MediaTek's new top-end phone chip, it's built on a 2nm process, it has two separate AI engines instead of one, and the whole thing is designed around the idea that your phone is going to be running an AI assistant in the background all day whether you asked it to or not.

Two brains for AI, not one!

This is the headline, so let's start here.

Most phone chips have a single NPU (the AI bit). The 9600 Pro has two. There's the NPU 1090, which is the muscle, meant for the heavy stuff like running a language model on your phone, generating images, doing the actual thinking. And there's the Super Efficient NPU 2.0, which is the quiet one that sits in the background doing the boring always-on work without eating your battery.

Think of it like a restaurant kitchen. The head chef handles the complicated orders. The prep cook keeps chopping onions all day so the kitchen never stalls. You don't want your head chef chopping onions, and you don't want your prep cook attempting the tasting menu. Splitting the job is just smarter.

The numbers MediaTek is claiming for the big NPU, all compared to last year's flagship Dimensity:

Double the INT4 compute (INT4 is the low-precision math that on-device AI models love)

55% more tokens generated per watt, meaning the AI talks longer before your battery cries

51% faster LLM prefill, which is the delay between hitting send and the AI starting to respond

Better KV-cache compression, so models take up less memory while running

Support for Mixture of Experts models running locally, and some on-device model training

That last one is more interesting than it sounds. MoE models are how the big cloud AI systems stay fast without being enormous, and getting them running on a phone is a real step.

The CPU: eight cores, all of them big

MediaTek has gone all-big-core for a few generations now, meaning no tiny efficiency cores. This is the 4th generation of that, in a 2+3+3 layout:

2x Arm C2-Ultra with 2MB L2 cache each

3x Arm C2-Pro with 1MB L2

3x Arm C2-Pro with 512KB L2

16MB L3 cache and 10MB system-level cache on top

Total L2 cache is up about 21% over the previous generation. Cache is basically the countertop space next to the stove. More countertop means less running back to the fridge, which is why cache bumps matter more than people assume.

Performance claims: 17% faster single-core while using 37% less power, and 15% faster multi-core at 61% better efficiency. If the efficiency numbers hold up in shipping phones, that's the part you'll actually notice, because it shows up as a phone that doesn't get hot and doesn't die at 6pm.

Feeding all this is LPDDR6 memory at 10667 and UFS 5.0 storage, both first-wave support. UFS 5.0 is claimed to be twice as fast at reads and writes, and LPDDR6 pushes about 33% more bandwidth at the same clock speed. Translation: apps load faster and big AI models get pulled into memory quicker.

Gaming: Mali-G2 Ultra NX and some genuinely fun claims

The GPU is Arm's Mali-G2 Ultra NX, and MediaTek has wired it directly to the NPU so graphics and AI can talk to each other without going the long way round.

27% more peak performance, 24% better efficiency at peak

Up to 18% faster ray tracing, with hardware-accelerated OMM and multi-material reflections

Native gameplay above 185fps

Neural super-resolution, which is the phone version of DLSS: render at lower resolution, let AI upscale it, get the frames for free

Ray tracing on phones has been a bit of a party trick so far. Multi-material reflections and OMM acceleration suggest it's slowly becoming something developers can actually ship.

Camera stuff, and one first

The ISP supports sensors up to 200MP and can do 8K60 video, 4K120 with stabilisation, and 4K240 slow motion. Video HDR goes up to 17EV, with 60fps continuous subject tracking so the focus stops hunting when your subject moves.

The neat trick is that the ISP can pull directly from the NPU, so AI processing happens inside the imaging pipeline rather than after the fact. MediaTek's example is a 4K live photo with full depth of field in a single tap.

There's also a genuine industry first here: this is the first phone chip with hardware H.266 (VVC) decoding. H.266 is the next-gen video codec, roughly half the bitrate of HEVC for the same quality. Nobody's streaming much H.266 today, but hardware has to come before the content does, and doing it in hardware instead of software is the difference between smooth playback and a warm phone. MV-HEVC for spatial video is in there too.

Screens, signal, and the boring-but-important bits

Display support goes up to WQHD+ at 180Hz, with a tri-port MIPI setup specifically for tri-fold phones, which tells you where MediaTek thinks the market is heading. There's a full BT.2020 colour pipeline end to end, plus multi-frame overdrive to clean up motion blur.

On connectivity: a 5G Advanced modem with 5-carrier aggregation across 350MHz for peak speeds past 7.4Gbps, and AI running inside the modem to cut gaming latency. Wi-Fi 7 with triple-band triple concurrency up to 7.3Gbps, Wi-Fi roaming handoffs that are 90% quicker, and 30% better positioning accuracy. Bluetooth 6.2 with a dual engine. Triple SIM triple active, if you're one of those people with three numbers.

For anyone in India, worth noting the GNSS block includes NavIC on both L5 and L1.

Security gets a hardware root of trust, post-quantum cryptography support, and something called AISeal, which uses hypervisor isolation to keep AI workloads walled off from everything else. Given that the whole pitch is an assistant watching your screen all day, that isolation isn't decoration.

So, is it worth caring about?

A couple of honest caveats. Every percentage above is MediaTek's own figure measured against its own previous flagship, not against Qualcomm. And a chip is only as good as the phone built around it. Same silicon, two different vendors, and you can get wildly different thermals and battery life.

That said, the direction here is clear and it's the right one. The interesting fight in phone chips has moved from "how high can the benchmark score go" to "how much AI can you run continuously without melting the phone." Splitting AI across a big NPU and a tiny always-on one, obsessing over performance per watt, wiring the NPU into the GPU and the camera pipeline, all of that is a chip designed for a phone that's quietly working even when you're not looking at it.

Whether that phone is actually useful is a software question, not a silicon one. But if it does turn out to be useful, this is the kind of chip that'll be running it.

What about the first phones?

The Dimensity 9600 Pro is expected to make its smartphone debut with the vivo X500 Pro and X500 Pro Max, which are expected to be among the first devices using the new chip.

The OPPO Find X10 Pro Max is also expected to feature the Dimensity 9600 Pro.

That means we probably won't have to wait too long to see whether MediaTek's performance and efficiency claims actually translate into something noticeable in everyday use.

Credit : MediaTek
Page: LINK