Apple goes all in on AI acceleration with M5 MacBook, iPad, and Vision Pros

Oh and the CPU is up to 15% faster for those that could care less about articifically intelligent Apple products and more about getting work done

Apple's fifth-generation of M-series silicon is starting to trickle out with the launch of the M5 MacBook, iPad, and Vision Pros this week.

However, those looking for an updated MacBook Air, Mac Mini, or Studio will have to wait a little longer to get ahold of Cupertino's upgraded system-on-a-chip (SoC).

On the surface, the M5 doesn't look all that different from the M4 it replaces. The chip features up to 10 CPU cores, 10 GPU cores, and a 16-core Neural Engine (NPU). 

Dig a bit deeper and it becomes obvious that Apple has managed to pack several significant architectural upgrades into the M5. Most notably, the GPU has been upgraded with improved shader and ray tracing cores that supposedly deliver between 30 and 45 percent higher performance in graphics heavy workloads like gaming over last year's M4 silicon.

Each GPU core now features an onboard neural processor, which you can think of a bit like the tensor cores in Nvidia's GPUs. These accelerate matrix operations common in generative AI and machine learning workloads.

And boy are Tim Cook and crew gung ho on AI with this launch. The subject of AI and how the M5 bolsters AI performance dominated the launch announcement, and the GPU is at the center of many of these claims.

Apple says that the GPU now delivers 4x the AI compute of the M4. How that translates into teraFLOPS of whatever-bit floating point performance, they conveniently didn't say. Even so, the inclusion of tensor cores in the GPU should translate into big improvements in large language model prompt processing and image generation — two areas where M-series silicon has historically struggled.

In theory, it should also benefit Apple Intelligence features now baked into the operating system.

Feeding the GPU is an improved memory subsystem. The M5 can be kitted out with 16, 24, or 32 GB of unified memory. This architecture is super fast compared to using SODIMMS, but means that whatever you spec the machine with is what you're stuck with. Compared to the M4, Apple has managed to boost memory bandwidth about 30 percent to 153 GB/s, which coincidentally is the same as Qualcomm's X2 Elite and Intel's Panther Lake.

For AI, faster memory should translate to improved generation rates for large language models running in apps like LM Studio.

And not to be forgotten is the M5's improved Neural Engine (NPU), which is used to accelerate smaller, less compute-intensive machine learning workloads which might run in the background. How Apple has improved the NPU isn't immediately obvious. The press release made no mention of TOPS, not that we're sure we'd notice the difference between a 38 TOPS NPU and one capable of 50 or more.

With that said, Apple's M-series has had an NPU from the very beginning and it's actually used for all kinds of subtle but helpful things like face and object detection in Photos or optical character recognition in Finder.

Not exactly a die-shot, but at least it gives some sense of how the M5's various cores are laid out - Click to enlarge

Setting aside the AI-focused improvements coming in the M5, Apple says it's also managed to deliver an even more powerful CPU this time around. The chip features the same 4 + 6 one-two punch of performance and efficiency cores as the M4, but says the cores are up to 15 percent faster in multi-threaded jobs.

The entire part is fabbed on TSMC's 3nm process node, and is currently available for preorder available in the MacBook, iPad, and Vision Pros starting at $1,599, $999, and $3,499 respectively.

The M5 arrives ahead of Intel and Qualcomm's latest silicon teased earlier this fall. Intel's Panther Lake SoCs announced last week won't debut until January, but will be available with up to 16 CPU cores and dedicated graphics options.

Qualcomm's X2 Elite chips, meanwhile, are expected to make their first appearance sometime in the first half of next year and will pack between 12 and 18 CPU cores in total. ®