Hacker News Reader

Comment by gardaani | Hacker News Reader

Rumors say that Apple will only release M6 base variant and skips M6 Pro, M6 Max and M6 Ultra variants to concentrate all efforts to create a good AI capable M7:

"According to reports from Bloomberg, Apple will be skipping its M6 Pro, M6 Max, and M6 Ultra chips to accelerate development of the M7 chip. That means the only chip to be released from the M6 family will be the base M6.

The reason for this break with tradition: AI. Apple had been planning major neural-processing upgrades for the M7 family and ultimately decided those improvements were important enough to justify accelerating the next generation rather than completing the M6 lineup." https://9to5mac.com/2026/08/08/apple-m7-chip-heres-why-it-ma...

I'd skip M5 and M6 chips for LLM work and wait for a year for M7.

View on HN

I wonder when/if we will start to see Apple include a silicon encoded model into their chips. Similar to Taalas build Llama 3.1 silicon with 17,000 tok/s inference.

So could the M7 actually include an AFM 3B model, alongside a generic neural engine?

That model would be larger and more expensive than the entire M7 chip.

Why would a local model for a consumer device need 17k tok/s?

Apple is better off building chips with generalizable TPUs (or equivalent) so they can upgrade/patch models.

I cant shake the feeling of "640KB is enough for everybody". Imagine not one AI answering over 1 minute but a team of 100+ agents in hieararchical structure taking care of your request in seconds, checking each other.

It would open heaps of use cases, you could almost pass it over frames of images the camera sees in real time for example...

It would make zero sense. These things are improving by leaps and bounds every week, we are not at the point where you can burn weights into silicon and put it on one of the largest consumer devices on the planet yet.

On the other hand, models these days are getting to the point where even if all development halted permanently, they would continue to be useful long into the future. (At least until their knowledge base or linguistics become too outdated.)

The M6 doubled the neural engine from 16 to 32 cores. I would expect that the M7 doubles that again to 64 from 32? That would make sense.

I believe that the CPUs are actually limited by ram bandwidth more than the neural engine right when it comes to LLM processing?

Maybe the M7 introduces something new to get around the current ram bandwidth problems on the non-Ultra chips.

LPDDR6 is coming.

I understand that will boost read rates to around 14 Gbps as compared to the current 10 Gbps for LPDDR5X, so a 40% improvement.

The bus is also 50% wider so the bandwidth is double.

Or they could make use of LPDDR5X-PIM? That would be such a killer feature and competitive advantage.

https://www.tomshardware.com/pc-components/dram/hot-chips-20...

Super cool!

Apple's biggest bottleneck for real-world inference is prefill processing. They need a better GPGPU architecture, which is what I'm expecting M7 to reveal.

Agreed, seems like the M5 has already made steps in that direction, with 4x prompt processing / prefill performance vs. M4. [1] And token generation also got a 10% boost. The data below is only for Pro & Max but I think the base M5 got the same relative boosts vs. M4 base.

    Chip         BW (GB/s)   GPU Cores   Q4_0 Prompt   Q4_0 Gen
    M4 Pro (20c)    273         20          439.78        50.74
    M4 Max (40c)    546         40          885.68        83.06
    M5 Pro (20c)    307         20      ~1500 to 1700    ~56
    M5 Max (40c)    614         40      ~3000 to 3500    ~92

[1] https://www.hardware-corner.net/m5-pro-m5-max-local-llm-4x-f...

ELI5?

Look at the "Q4_0 Prompt" column. That's the tokens-per-second processing the initial prompt/system prompt. This is where the "neural accelerators inside each GPU core" is seen most prominently.

The M5 GPU added better matmul and dot product support in hardware (IIRC) that really boosted performance of those kernels over previous GPUs. Current transformer models rely heavily on matmuls. Prefill (processing the current context) is mostly limited by raw GPU processing power where token generation (predicting the next token from the current context) is mostly limited by memory bandwidth.

While an older M-series might run a matmul or dot product kernel just fine the M5 can run them much faster.

The Apple chips are quick when the model is "warm" on your second and subsequent requests. It's also quick if you don't have a harness, and feed "Hi, how are you?" as a prompt. However, a lot of common usage is to have a harness with lots of instructions and context. Loading that context is slow. So, the first response may take ~1 minute to return the first token (varies widely by hardware, model, etc.).

  Apple's biggest bottleneck for real-world inference is prefill processing.

Much less true since the M5 generation. Prefill, aka prompt processing, got a 4x increase.

I'll upgrade M3 Air only when Mx Pro/Ultra can run Opus level perf locally. Otherwise what's the point.

i sold my m3 ultra /512 for £14000, or about $18000. The difference is what is important to me for a new one, and assuming $6k or so for the 256->512 boost, my cost will be about $16k, so i’ll save $2k by upgrading to the top-of-the-range model, bar it being a 4TB drive

Their chip release cycle doesn't make any sense to me. Compared to their other hardware like iPhones or Apple Watches they just release a new chip when they feel like it. It might have a pro, ultra, or max version, or it might not. Sometimes that better version only releases when the next generation of the silicon is already out. I think that they really need some structure in their releases.

The difference makes a lot of sense to me.

The phones are watches are primarily sold on annual leases through carriers. And no one is buying them for the chips, it's all camera and form factor.

The macs, on the other hand, can more easily swap between chips in manufacturing. And the form factor is pretty static. So Apple, in a race to have the best chip in the field at any given time, is just releasing them as soon as they're ready.

I agree the whole "m5 max > m6" is confusing, but I think they figure only the nerds who learn the specifics are going to care anyway.

Do current models run on the NPU or GPU? Wondering if Apple will have something like a TPU.

Apple has a dedicated "neural engine" which is designed as an inference NPU. Where as Google's TPU has a dual focus, both inference and training, which is a more complex design.

That TPU training part I get, but from what I have seen the NPU is rarely used for inference by LLMs. They still use the GPU, no?

> I'd skip M5 and M6 chips for LLM work and wait for a year for M7.

Or you could lease an M5 max/ultra until the M7 equivalent comes out. At least in the US a leasing option is available.

I'd be worried about securely erasing my data when returning a leased model.

What do you do with older hardware that you've bought? Do you not wipe it and resell / donate?

If you encrypt from day one then a lease isn't much different.

Just delete the encryption key.

> I'd skip M5 and M6 chips for LLM work and wait for a year for M7.

Another note on this, it's likely the M7 ultra will be released around 6 months after the M7 Max, if the M1 and M5 are taken as reference. So the M7 ultra may be delayed until early 2028.

> I'd skip M5 and M6 chips for LLM work and wait for a year for M7.

Please say more? Is it because it is a one-time cost, unlike a recurring subscription of Claude/Codex?

If you expect a large leap in performance, buying now means that (a) you will want to buy again in a year but (b) the large leap in performance hurts the resale of what you buy today.

Local models will be that much better in a year, so unless you have to have a top-of-the-line local rig today (the argument goes), stick with Claude for another year.

That makes sense. I also think that we are at the point where maybe the expectation of a yearly refresh on everything should be reconsidered. They can end up on a Tick-Tock schedule when the Pro/Max/Ultra can be done every second year.

The true-local AI chip, codename "buddy", will be the M8

"buddy" -> "mate" -> M8. Nice.

Reply on news.ycombinator.com