Marvin

commentzilla said:

Sounds like a computer for K-12 education.

They'd be suitable for any general use cases but definitely popular with students. Even entry-level Apple Silicon is really fast. iPhone chips are essentially 1/2 the entry M-series chips of the same generation.

https://gfxbench.com/device.jsp?benchmark=gfx50&os=iOS&api=metal&cpu-arch=ARM&hwtype=iGPU&hwname=Apple%20A18%20Pro%20GPU&did=123295110 (A18 Pro)
https://gfxbench.com/device.jsp?benchmark=gfx50&os=OS%20X&api=metal&cpu-arch=ARM&hwtype=GPU&hwname=Apple%20M4&did=123984676 (M4)
https://gfxbench.com/device.jsp?benchmark=gfx50&os=OS%20X&api=metal&cpu-arch=ARM&hwtype=GPU&hwname=Apple%20M1&did=90754264&D=Apple%20M1 (M1)

A18 Pro is roughly 1/2 M4 and around the same as M1.

This would be like Apple selling an old M1 Air at $100-150 less than the entry Air that uses half the power. This would be $649-699 for education buyers.

surgefilter said:

So big set of micro needles in the back of it constantly taking blood sample?  Will that not get a bit messy? Interesting and useful addition if they can pull it off.

It seems there's some non-invasive alternatives:

https://www.healthline.com/health/diabetes/blood-sugar-monitor-without-finger-pricks
https://www.healthline.com/diabetesmine/non-invasive-diabetes-technology

This kind would be usable in a smartwatch:
https://www.hagartech.com/

Their devices use radio frequencies and described as 1/3 the size of a smartphone. Either Apple would use a smaller integrated sensor or an accessory device, maybe a custom watch band.

This one uses a light beam to heat up the glucose in the skin and the temperature change shows the amount of glucose:
https://www.diamontech.de/en/solutions/d-pocket

The sites say they are 95% as accurate as the blood testing strips. This would be a big improvement from having pins stuck in either regularly via finger testing or with the CGMs installed all the time.

michelb76 said:

Some results from my M1 Max, 10 CPU / 32 GPU, 64GB:
Steel Nomad - 1754
Solar Bay - 22008
Wild Life Extreme - 17910

Even with the M4 Max, Apple still has long way to go.

M4 Max is a laptop chip, it's comparable to a laptop Nvidia 5070.

5070 laptop (~100W):
https://www.3dmark.com/sn/6066166 (Steel Nomad - 3166)
https://www.3dmark.com/sb/301082 (Solar Bay - 67588)
https://www.3dmark.com/wl/477536 (Wild Life Extreme - 25707)

Desktop GPUs like the 4090-5090 are 400-500W GPUs.

M3 Ultra is a bit higher (about the same as a 5080 laptop):
https://www.3dmark.com/snmac/876 (Steel Nomad - 5519)
https://www.3dmark.com/sbmac/624 (Solar Bay - 81084)
https://www.3dmark.com/wlmac/340 (Wild Life Extreme - 51896)

If there was an M4 Ultra, this would be about 25% faster.
If there was an M4 Extreme (quad Max GPU), this would be competitive with the fastest PC GPU, the 5090, but this would also use 400-500W. Only the Mac Pro chassis would be able to handle this and the sales volume at that price point would't justify it.

The main thing is they are competitive in the laptop space as that's the majority of users and they have the advantage with unified memory.

beowulfschmidt said:

It seems to me that Apple's commitment to gaming extends only as far as their own app store.  If you ain't buying from one of Apple's apps stores, they really don't care much about gaming.  They've demonstrated this time and time again.

It serves as a showcase for how powerful and efficient their GPUs are. They had people from CD Projekt Red at WWDC showing off the Cyberpunk port:

https://geekculture.co/cyberpunk-2077-on-macbook-pro-m4-max-shown-at-apple-wwdc/

It was running at 120FPS on Ultra quality on M4 Max on battery. They said Apple worked with them to make some things more optimal like using FP16 to reduce load on the CPU.

They don't say if it was with path-tracing or just Ultra raytracing, I would assume the latter. This would make it similar to an Nvidia 5070, which reaches around 140FPS at 4K DLSS performance (1080p upscaled to 4K), 4x frame-gen with Ultra ray-tracing:

https://www.youtube.com/watch?v=no63Pq-UOXo

This would mean M4 Pro will do 60FPS on the same quality and M4 will do 30FPS and the quality can be lowered a bit for higher FPS.

Path-tracing cuts the FPS to 1/3-1/4 of Ultra so Max might still be able to handle path-tracing.

Eventually these games can generate some revenue in the store but it's more important that the games are available on the platform in much the same way having native pro software helps attract users to the platform.

Some games run smoother on the Mac due to the amount of graphics memory and SSD speed. Even higher-end PCs get stuck with 8GB of VRAM so they have to keep streaming assets in/out of memory causing stuttering. A 32GB Mac can use 16GB or more for video memory.

Hopefully CD Projekt Red ports their upcoming Witcher games. These are built on Unreal Engine so should be easy enough to port over:

https://www.youtube.com/watch?v=FJtF3wzPSrY

Real-time rendering is reaching the end-game for quality. Path-tracing allows photoreal rendering in real-time and approximated lighting like Unreal Lumen looks close to this:

https://www.youtube.com/watch?v=ZkuFNQIqcl4

This would be nice for interactive immersive environments on Apple Vision Pro.

Rogue01 said:

Siri and improved Apple AI reminds me of AirPower.  Schiller - we can do this.  Schiller a year later - no we can't.  I am not holding my breath for Siri.  I rarely use it because it isn't very reliable, and the response is usually the same thing, let me see what I can find on the internet.  I can do that myself and get a better response.  Apple has had 14 years to fix Siri, and hasn't done it yet.

Local AI is too resource-intensive for mobile just now. The 7 billion parameter models use about 10GB RAM. Apple's local models have around 3 billion parameters:

https://machinelearning.apple.com/research/introducing-apple-foundation-models

GPT 4 is 1.7 trillion parameters. GPT 4o mini is 8 billion:

https://deepnewz.com/ai-modeling/microsoft-paper-reveals-gpt-4o-mini-size-8b-parameters-gpt-4-1-76t-claude-3-5-a05ca5f5

If they can get something like GPT 4o mini to run locally, that would be an improvement. iPhones would likely need 16GB RAM.

They want agent-like capability with Siri to be able to do tasks (1:44) where someone can ask the device to do everything:



It needs a few things: more RAM, better reliability from smaller models, good training data and process, good QA testing.

It feels like a good default setup would be normal Siri for local tasks but instead of searching the web when it has difficulty, it uses GPT 4o on the server and has local Siri interpret the reply.