Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
Apple is falling behind, M4 Max nowhere close to RTX 5090.
- Thread starter Zest28
- Start date
- Sort by reaction score
If Apple was to move away from UMA, they would also lose some of the advantages of the current design, which will become more important as AI/ML usage continues to grow. The added latency associated with copying data twoce and comparing the data after the CPU and CPU are finished processing it would negate many of the advantages Apple Silicon has now. Ditching UMA would be a massive step backwards at a time where Apple can't afford to stand still, let alone regress.
the GPU uplift here will probably mean that the m5 will have metal scores around 70000. and if the pro/max also have some variant of SOIC, it could mean an even bigger lift in general. Not unthinkable that a m5 max will land at about 200 000 in metal.
and with 2x f16 for some ops it will me a measurable uplift. my m3max actually is just the same for f32 as f16 for tests I have done in pytorch and mlx. That will not bring the perf inline with nvidia for ML loads but it is an improvement. Maybe there is more niceties not known at this time, but all in all, the m5 series seems to be a bigger improvement than expected which is good. Now just add in faster mem as well and a m5 ultra that scales better...
and with 2x f16 for some ops it will me a measurable uplift. my m3max actually is just the same for f32 as f16 for tests I have done in pytorch and mlx. That will not bring the perf inline with nvidia for ML loads but it is an improvement. Maybe there is more niceties not known at this time, but all in all, the m5 series seems to be a bigger improvement than expected which is good. Now just add in faster mem as well and a m5 ultra that scales better...
I think if you read my entire post you’ll be able to suss out what I meant, but I’ll reiterate: the criticism is valid, particularly around the lack of self-learning and correcting errors. The technology has a fundamental flaw in its design and token prediction no matter how much you bolt onto it can not solve this intrinsic problem, e.g. if it makes a mistake and I correct it, it will never change all of its future behavior, even with the frankly rudimentary “memory” capabilities in the latest Frontier models. I personally believe future world models with some different development paradigms may be able to solve this, but we’ll see.
I use 2-3 AI tools every single day and am working on a pretty complex pipeline for generative objects as a side project with a centralized orchestrator and human in the loop review. I am absolutely not against the advances in technology.
For me personally, the tooling became enormously more useful once it got web search and some of the memory / context length etc. improvements, but there is a long way to go. I do think there is a bubble for non-frontier model companies, even ones as large as Perplexity which from my POV has nothing unique to offer. Anthropic, OpenAI, xAI, Google, and Meta are all going to continue to exist and move forward for the next 5+ years at least. To what degree, we’ll see.
Hope this helps. I can both enjoy the benefits of the technology and criticize what I see as obvious limitations. It’s not an attack on you at all, but the marketing that says they’ll eventually scale and solve the problem, because the answer doesn’t exist there.
I genuinely am hopeful about the advances we’re seeing, as I said elsewhere in the forum my view is that this is the most exciting time for technology in 2+ decades, outside of the sociological issues which I don’t want to get into in this forum but recognize do exist. I’m having a great time and learning a ton I would not have otherwise without these things, and can’t wait until Apple’s hardware gets a significant bump in the next generation or two to support even faster development turn-around.
I’m so hopeful in fact that I’m transitioning professionally from traditional engineering management to re-focusing on the fundamentals with a medium-term goal of getting more deeply involved in the RL space, I see a lot of very exciting challenges there I’d like to work on but we’ll see how that pans out. It’s good to have stretch goals
.…
Once again on the thread topic I can’t wait to see what M5 Pro / Max bring. Here’s a metal score directly for A19: https://browser.geekbench.com/v6/compute/4765325
And scores for the 17 Pro Max: https://browser.geekbench.com/v6/cpu/13736665
I think the full gpu core M5 Max may hit 250,000 which would be phenomenal. We’ll see in ~4-6 months!
Last edited:
I think it is possible that they are working on stacked SoCs where GPU could be on a separate die (but still part of the SoC), allowing more flexibility. I very much doubt we will see this tech this generation however.
I certainly don’t see them abandoning UMA, that wouldn’t make any sense. What would make sense is larger L2 for the GPU.
Up to 40% increased GPU performance bodes well for M5.
iPhone 17 Pro and iPhone Air Benchmarks Reveal Speed of A19 Pro Chip
The first benchmark results for the A19 Pro chip in the iPhone 17 Pro, iPhone 17 Pro Max, and iPhone Air surfaced in the Geekbench 6 database today. ...
www.macrumors.com
I think there is a difference between me saying that LLMs are not hype and that they are extremely useful and will continue to get more useful quickly vs me saying that LLMs are perfect.
I never said they're perfect. I don't understand why you're directing these things at me. I'm honestly baffled.
I also never talked about junior devs that you guys keep bringing up. It's someone else, not me, who brought up the topic of junior devs.
Meanwhile, I'm responding to mr_roboto who doesn't understand why anyone would use LLMs.
And I'm sure they said the same thing about using calculators vs doing hand calculators in the past.
People who think AI is in a bubble or "bubble hype", just short Nvidia. Put your money where your mouth is.
And for whatever reason, he's now attacking my environmental stance? What the heck?
There is nothing from mr_roboto's post that is worth discussing over. It's the same old regurgitated anti-AI stance that you read everywhere. It's all the same. They cite some blog/article on how AI is not perfect but ignore the huge benefits. They say LLMs hallucinate but ignore the extremely fast progress in reducing them and giving LLMs tools to verify. They say it's not "real AI" because it's not AGI. They say it's bad for the environment but they're continuing to buy new Apple devices every year.
The truth is, a lot of these people are actually afraid of LLMs. They're afraid that it's changing things too fast. They don't work at OpenAI or Anthropic so they feel like it's out of their control. They're secretly afraid that it could make them redundant or take away their earning power. They're afraid that it'd give big tech too much power.
These are all valid reasons to be afraid of LLMs. My problem is that people who are take a strong anti-LLM stance do not disclose this hidden feeling. They usually let it slip though. For example, you can see in this message that mr_roboto thinks there's a decent chance LLMs can indeed replace a lot of junior devs:
Last edited:
Up to 40% increased GPU performance bodes well for M5.
iPhone 17 Pro and iPhone Air Benchmarks Reveal Speed of A19 Pro Chip
The first benchmark results for the A19 Pro chip in the iPhone 17 Pro, iPhone 17 Pro Max, and iPhone Air surfaced in the Geekbench 6 database today. ...
Yea, it was mentioned here earlier. I wonder how much of the 40% improvement is the cooling upgrade and how much of it is the GPU improvement.
It matters because cooling for Macs will likely stay the same.
Geekbench is usually unaffected by the sustained performance drop-off, as the tests are very short. So I’d say most of it is performance improvements. And it’s in line with what I’d expect from dual issue FMA+ADD plus some clock boosts and other tweaks.
Btw, I’d also think that gaming improvements will be even better since there are more opportunities for FP16. Unless GB6 already uses half-precision extensively in compute shaders.
Yeah this is kinda the opposite of where things are going.
The next step is building the CPU/GPU cores directly into the memory to get better performance.
Nah thats probably more a case of the M4 still being competitive, and the iPad getting m5 first.
The M1 and M4 shipped in November, q1 is not that far off. Apple could have been working on "optimising a cooling system" for a decade or more, whether the CPU is out yet is irrelevant as its a simple case of "how do we dissipate X watts in Y^2 space" and that's something they can play with using a heating element.
The next new cooling system will be with the redesign, and some suggested late 2026 with oled
I think spring is ipad pro and macbook air all with M5
The MB Air's could get vapor chamber cooling like the Pro phones. The MB Pro's already have vapor chamber cooling (which is all a heatpipe is). The next steps would be things like Liquid Metal thermal TIM to "increase" thermal conductivity. Or maybe coming up with some exotic liquid that can carry more heat load for the pipe to dissipate.
I might be reading too much into the IPhone 17/A19, but apple is adding the vapor chamber to the phone for a reason - Its quite possible that the A19 runs warmer then its predecessor. The implications are that the M5 will be a hotter running chip that will need better cooling technology. Again, that's a giant assumption that may not have any basis in reality.
MBA would need something since it draws more power than the iPhone. A passive HP/VC system is kind of their only choice.
I agree, but I was also thinking more along the lines of the MBP, Mini and to a lesser extend Studio - if the M5/M5 Pro/M5 Max are indeed a hotter running chip
The 16" unit cooling seems fine, the 14" will suffer. I can't speak to the Mac Mini, and yea the Studio has an overbuilt cooling system.
The mini runs warm, particularly the M4 Pro. This was one of the reasons why I returned the M4 Pro Mini and bought a M4 max Studio
Here's the truth: Apple had a LARGE lead and blew it. When AMD started making strides on power efficiency, most people here dismissed it saying X86 is a legacy architecture.
The problem is that they STILL make progress, and now they have the AI Max+ 395, which can have up to 128GB which can not only be used for gaming (it runs some games smoothly at 4K!), Sure, this memory won't run AI as fast as an actual, dedicated 3D card. But it can be used for large language models at an OK speed run with as little as 45W. Those language models will still run slower than an Apple Silicon processor, but since we have Thunderbolt 4 and Windows GPU support, we have the option to connect external GPUs if running a laptop (or straight up plugging a GPU if we have a desktop option).
And because we ALSO have regular Windows, we have a portable system with all legacy applications it offers (I dislike Microsoft, but they did something right here).
The problem is that they STILL make progress, and now they have the AI Max+ 395, which can have up to 128GB which can not only be used for gaming (it runs some games smoothly at 4K!), Sure, this memory won't run AI as fast as an actual, dedicated 3D card. But it can be used for large language models at an OK speed run with as little as 45W. Those language models will still run slower than an Apple Silicon processor, but since we have Thunderbolt 4 and Windows GPU support, we have the option to connect external GPUs if running a laptop (or straight up plugging a GPU if we have a desktop option).
And because we ALSO have regular Windows, we have a portable system with all legacy applications it offers (I dislike Microsoft, but they did something right here).
I'm glad AMD is making gains, but they are not nearly caught up. Wait a few months and re-assess Apple after M5 Max is released, you may be surprised.
I think it will take another year or two for Intel or AMD to catch up to Apple Silicon across the board including power usage, and that's assuming Apple hits some snag with M6 or M7 and doesn't keep iterating so quickly.
Meanwhile the inverse is true with Graphics and Compute, I believe Apple will close the gap(s) significantly and wind up maybe a 12-18 months behind nvidia which isn't bad given where they were a few years ago. This is all consumer grade I'm talking about, I'm not sure anyone can catch nvidia in the large datacenter space given their enormous lead in large scale interconnect technology.
We'll see
.PC wise I'd run linux primarily anyhow but I do recognize there are a lot of benefits to the CUDA development toolchain and yes some Microsoft apps are useful.
Development wise, I swore off ROCm after horribly buggy experiences a while ago and to my mind MLX etc. are already at parity or better for consumer grade hardware development.
For regular desktop use and especially gaming, AMD is doing very well. I just hope they quit screwing around with their V-Cache and actually put it on every damn CCD next time. It's ridiculous they keep holding out on that; we used to pay $1,000 for HEDT processors and $600 for graphics cards and now we're paying $2,000 for graphics cards and $600 for processors.
They can throw prosumers a bone and give us a full chip without software load balancing compromises, it's a huge part of why I haven't upgraded my PC in years because I keep waiting for something that isn't compromised in some obvious way or a nuclear furnace like Intel's latest underperforming processors. I'm considering getting a 9800x3D to just hold me over for a year since my system is aging but has weirdly valuable components, but I hate stopgap solutions.
Last edited:
I'm not sure why you are laughing. Sure, if you have an Apple-only workflow with Final Cut, I'm not going to convince you to change. But AMD is not to be underestimated with their low-power devices. I have an ROG Ally, which is not even as good as that new processor anymore, and it's incredibly useful as a portable computer as it is.
I'm glad AMD is making gains, but they are not nearly caught up. Wait a few months and you'll see.
I think it will take another year or two for Intel or AMD to catch up to Apple Silicon, and that's assuming Apple hits some snag with M6 or M7 and doesn't keep iterating so quickly.
Meanwhile the inverse is true with Graphics and Compute, I believe Apple will close the gap(s) significantly and wind up maybe a 12-18 months behind nvidia which isn't bad given where they were a few years ago.
We'll see
PC wise I'd run linux primarily anyhow but I do recognize there are a lot of benefits to the CUDA development toolchain. I swore off ROCm after horribly buggy experiences a while ago and to my mind MLX etc. are already at parity or better for consumer grade hardware development.
On paper, it's true that the GPU of a portable, state-of-the-art AMD device is not as powerful as a state-of-the-art Apple Silicon processor.
BUT the gap now is VERY SMALL. You are going to miss that extra GPU power if you do rendering, but there are ways to more than make up for it if you are willing to sacrifice some portability (e.g, Thunderbolt 4 + eGPU, which Apple USE to have as an option).
Frankly, for regular usage, there's nothing that an Apple GPU can do anymore that I can't do with my devices. Of course, everyone's use case is different. I'm describing mine here, but I doubt that AMD wouldn't suffice for many video editing and even light video fx effects / rendering (or an eGPU if you want to do heavier rendering).
By the way, some people report 10 hours battery life is easily achievable with those devices: ( ). Pretty good for an "outdated" architecture.
Last edited:
Absolutely, I think AMD deserves credit for pushing here and I hope they eventually fuse this technology with their desktop architecture, I'm not sure what the plan there is I haven't seen rumors about it. I don't want to get to far off topic but something that combined HEDT type features with that new integrated architecture for the memory bandwidth improvements and "AI" acceleration would be really cool.
Based on the A19 Pro slides it seems like we're in for some really good things for M5 Pro / M5 Max. It's a great time for this technology across the board. Unless you're Intel, lol.