Hacker News Reader

Comment by jdprgm | Hacker News Reader

I don't understand why they don't update the entire lineup whenever a new chip comes out, especially if really only swapping the chip in the macbook pro and imac here.

On a separate note does anyone have any insight on how unified memory compares to vram in the context of machine learning performance? Considering an H100 with 80GB costs like $30k a maxed out macbook pro with 128gb unified memory for $5k is interesting. Is it remotely comparable or compelling for large models considering realistically most prosumers are capped at the 24gb nvidia cards for anything resembling a reasonable budget.

View on HN

I expect it's a combination:

1. Engineering bandwidth (it's work to upgrade a lineup! They probably need at least a year)

2. Manufacturing bandwidth - they probably spent the first half of the year manufacturing the 3nm iphone processor while researching the mbp processor. I expect it's difficult to ramp manufacturing on many chips simultaneously

3. Sure, consider the demand side. If my parents are going to buy a new laptop and a new phone this year, I think they would be more likely to do that 6 months apart. Similarly, I expect keeping a cadence of announce products a,b,c in this quarter and d,e,f in that quarter helps keep apple in the news (for good/exciting things)

Wrt machine learning: I can't wait for the results to come out for this once we can get our hands on it

> 2. Manufacturing bandwidth - they probably spent the first half of the year manufacturing the 3nm iphone processor while researching the mbp processor. I expect it's difficult to ramp manufacturing on many chips simultaneously

Sort of. It's that the whole concept of a "ramp" implies that manufacturing things (not just chips!) starts expensive and risky and get cheaper and reliable over time as people work out the kinks in the various processes.

The M2 is mature. TSMC can make them reliably and in high quantity at low marginal price. Why would you *not* sell a product like that if there's a market for it?

Also cost.

Apple introduced the 15 inch MacBook Air on June 6th.

If they upgraded it to the M3 right now that machine would have only lasted 4 months and 4 days.

Sure they could do it, but how many sales are they really losing because they haven’t? People buying the Air are not exactly looking for the absolute best in technology. They’re probably much more likely to be price sensitive, and putting a brand new M3 in would likely make that worse.

So if you consider that you have to make new boxes and manuals and new hardware and test it and change the lines and everything else… what are the chances you would come out positive after such a short amount of time?

It would make some more sense for the 13 inch air which is about 18 months old, but again if you don’t have to maybe you keep selling the M2 for another six months.

Speculating, but I'd wager it's to clear out existing inventory by phasing out the old chips on other SKUs over time. E.g., you start with the expensive Macbook Pro models when you have fewer M3 chips in production, since those will sell slowly, and then you clear out your remaining M2 inventory in the Air lines while waiting for a "refresh" when M3 production is more ramped up.

I have no idea what Apple's sales are for the macbooks or if Airs sell more than Pros, etc., but they're definitely making profit this way to be this consistent in their approach.

Besides wanting to clear out existing M1/M2 capacity, I don't there's enough M3 stock right now.

I forget how many fabs TSMC has that have 3-nanometer capability, but it's gotta be low. Apple's supposedly got _all_ of TSMC's capacity[0], but I doubt it's enough.

[0]: Apple is saving “billions” on chips thanks to unique deal with TSMC | https://news.ycombinator.com/item?id=37040722

They just updated their macbook air and mac studios and mac pros like 6 months ago? If they updated right now, the previous owners would feel utterly screwed over.

Yup I bought an m2 mbp about 3 weeks ago.

Feels like it's moved super fast. I just bought an M1 Studio earlier this year.

Almost four years now, for M1. I get the impression that people are used to Intel’s stagnation that we were stuck with for some time.

And more importantly, these improvements are mostly thanks to TSMC's progress. You could predict the M3 already late 2022 when TSMC announced that they were starting mass production of their 3nm process. We knew that it would be a huge improvement based on the specs we had back then and roughly how long it takes to ramp up.

https://pr.tsmc.com/english/news/2986

Only 3, unless you're counting the developer boxes? M1 machines were released in November 2020. However, M1 Pro and M1 Max were released only 2 years ago. Things are definitely picking up now...

Right but why did they just update them then? Why not do it earlier or wait for the M3?

Not even. If you ordered a non-baseline M2 Ultra Studio at launch, you likely didn't receive it til the first week of August. So we're a week shy of three months. So I'm glad for not being screwed, I think. Although who is to say they're not just being sat on (M3 Max chips)...

This is a perspective that comes from stagnation. Tech is progressing again, finally. This is normal for tech, and a good thing.

The M2 Ultra tops out at 800GB/s unified memory bandwidth. A 4090, on the other hand, has 1,008 GB/s. On the PC side, dual-channel DDR4-6400 offers a bandwidth of 102 GB/s.

They're not really directly comparable due to the architectural differences, right? For pure rendering workflows I'm sure the 4090 is faster, but for anything that hits the CPU ...

Yes, the tiled-based architecture of their "GPU" is undeniably less performant. Even comparing to lesser GPUs.

Nvidia is tile based since Maxwell

Yes, I should've been more specific. Mobile GPUs often employed in ARM architectures often use a technique called "Tile-Based Deferred Rendering" (TBDR). This involves not just splitting the frame into tiles but also delaying the actual shading operations. This two-step approach minimizes overdraw but can create bottlenecks in scenes with intricate geometry or advanced lighting, shadows, and reflections.

Nvidia's approach doesn't have this.

How is the deferred rendering in TBDR different from the deferred shading and deferred lighting that have almost completely supplanted forward rendering in commercial games and engines?

Why exactly is tile-based a bad idea? Most rendering engines are tile-aware already.

See my response to the sibling post

I guess the question is: in terms of performance per watt, how valuable is the "unified" aspect?

Bandwidth means much less if those numbers aren't crunched faster.

Depends on if you're talking about training performance or inference performance. It's probably a decent deal for inference on such large models, but I doubt it's anywhere near competitive for training. In the 'consumer' space there's also the A6000 with 48GB for ~$4k.

It's probably due to a combination of the processor manufacture rate and the quarterly sales boost they get throughout the year

It makes a world of sense if you think of it from a logistics point of view, especially considering Apple sells insane volumes of everything they sell.

It costs less to manufacture the current generation ship and the last generation chip concurrently, it allows them to push the newest generation chip out the door faster without running out of supply or having to open up excessive production lines.

>H100 vs apple silicon

while it's only one spec, H100 has over 2TB/s of memory bandwidth and Apple silicon caps out at 400GB/s. H100 in general kicks the pants out of Apple silicon performance and has far more support. These parts aren't really in the same league but Apple Silicon has become a popular hobbyist choice for LLM inferencing.

Also see: The Jetson AGX Orin 64gb, Nvidia RTX A6000 48gb, AMD Radeon Pro W7900 48gb, two nvlinked 24gb 3090's.

Regarding updating the lineup, I would guess that it is completely intentional to keep the carrot just out of reach across the lineup. Each new chip release rotates the product categories in just the right way to keep the consumers consuming.

Right? I’ve just setup my office and for now I’m using my MBP hooked up to a couple of Studio Displays - which I adore for software development. I was planning on picking up a Mac Mini or Mac Studio so that I could still have the MBP spare.

Maybe not. It’s a bit of a pain having to unplug and replug everything in, not to mention having windows and such shuffle around each time.

My gut tells me the Macbook Air sales have been cannibalizing Macbook pro sales.

For the M1 and M2 generations, the airs released in the fall and the pros a few months later in the spring. The airs have gotten so good that many people that would have gotten the more expensive (and more lucrative) pros got airs instead.

Reply on news.ycombinator.com