Darajavahus
macrumors regular
Jesus if the Pro have the same CPU, the multi is 2x my M3 Pro ...just two years later... ðŸ˜
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's M5 Max Chip Achieves a New Record in First Benchmark Result
- Thread starter MacRumors
- Start date
- Sort by reaction score
klasma
macrumors G4
That would likely be well into the five digits, with current RAM prices.
Hope so. But they discontinued even the 512GB M3 Ultra now. Only a 256GB config remains…
Please Apple don't wait 6 months to update the Mac Studio. Don't be greedy a$$holes who want as many people as possible to order much more expensive MacBooks Pro to get access to the latest fastest chip.
marco114
macrumors 6502
Any idea what 1TB of ram cost on the 3rd party market today? Roughly $30k. I can’t imagine how expensive that would be in Apple $$$
I wonder if the next version of macOS, that will be Apple Silicon only, will be faster without all the Intel code.
IceCool
macrumors 6502a
Cannot wait for mine to get delivered!
This M2 Max/64GB 14" has served me incredibly well in the last 3 years and I have quite literally zero complaints about it bar none. It definitely has enabled me to do (and learn!) things I wouldn't have been able to do on any of my previous Mac's. While it will now become a secondary computer since my work is exceeding its capabilities, I do strongly believe that this was the best $5k I have ever spent and that I have more than gotten my money's worth.
This M2 Max/64GB 14" has served me incredibly well in the last 3 years and I have quite literally zero complaints about it bar none. It definitely has enabled me to do (and learn!) things I wouldn't have been able to do on any of my previous Mac's. While it will now become a secondary computer since my work is exceeding its capabilities, I do strongly believe that this was the best $5k I have ever spent and that I have more than gotten my money's worth.
I'm waiting for the M6 series of processors. The M5 Pro and Max are the first set of processors using an "all-new Fusion Architecture". I'm interested in the refinements that come in the updated version. That and my M1 MacBook Pro is still fast without any lag or problems.
Going from 12 performance cores to 6 (yes, SIX "super cores" which ARE equivalent to the old performance cores) is quite an enormous drop.
I expect the architecutre will allow them to stretch their legs with the Mac Studio assuming it will last for 2 generations.
Otherwise, I think it's artifically limited for heat/power and... to sell more M6 Macbook Pros with OLED. And higher prices by then, likely.
It also may have been a financial decision to cut costs so they didn't have to jack up RAM and SSD prices due to the market quite yet.
NeekonMike
macrumors member
With identical 18 CPU cores, M5 Pro will likely score the same (synthetic benchmarks probably don't reflect Max's faster RAM). Insane.
conmee
macrumors 6502a
How are you managing to get by without any neural accelerators or ray tracing? 😉
ddhhddhh2
macrumors 6502
No need to guess. With how current LLM work, memory bandwidth basically determines the speed. If the bandwidth isn’t 4× higher, the speed won’t be 4× either. It’s just physics.
mansplains
macrumors 68000
ddhhddhh2
macrumors 6502
I’m actually more curious what people expect to run with 1TB for local LLM.
Even the biggest open-source model we have right now, like Llama 3.1 405B, would be around ~800GB if you ran it unquantized in FP16. But realistically almost nobody runs local models like that anymore, quantization is basically the norm now.
With something like Q4 / int4, even a 405B model would only need roughly ~200GB just for the weights.
And with MoE models, the parameter count can look huge on paper, but during inference only a subset of experts are actually active anyway, so the real runtime footprint is way smaller than the headline numbers.
Also another thing people forget: even if someone released an open model with parameter counts somewhere near GPT-4 scale, that still doesn’t mean the real world performance would match it. Model capability depends on way more than just parameter count : training data, post-training, alignment, architecture, all that stuff matters a lot.
And that’s even assuming the model would actually need anywhere close to 1TB of memory in the first place.
ddhhddhh2
macrumors 6502
Yeah, I’m not saying bandwidth is the only factor. It’s just a rule of thumb with current LLM, if bandwidth doesn’t scale close to 4×, it’s hard to expect real LLM speed to scale 4× either.
I am an M1 16 Max waiter as well. Very excited by the improvements, especially internal disk speeds and the TB5 improvements, but I just know I would have big time buyers remorse if I didn’t wait for the redesign and the new panels. Was a twitch away from ordering a Max 40c with 4TB and 128GB, now just praying that they don’t cripple hammer anything in the M6 transition and that the price hike isn’t insane.
How do I put this gently... This comment is very "2 years ago." A lot has changed. Too much to explain in a comment. The current SOTA open weight models perform extremely well - and they need this amount of memory to run at or close to trained weights.
I have no interest in running LLMs, therefore, not everyone wants to run LLMs.
But I am pleased that you are willing to do your part to stimulate the economy.
ddhhddhh2
macrumors 6502
I don’t really want to derail this thread.
I totally agree that if you're chasing absolute peak performance, running models close to their native precision has advantages, and in that case huge memory obviously becomes a hard requirement.
What I’ve personally noticed though is that quantization has improved a lot over the last couple years. It’s really not the same as when I was working on RAG about two years ago.
I can understand that for people like you who need to squeeze out the very last bit of performance from a model, quantization probably isn’t even on the table. Maybe that’s exactly where something like 1TB of memory actually makes sense for a certain group of users.
chars1ub0w
macrumors 6502a
Hang on. It was written:
"With four additional CPU cores compared to the M4 Pro, the new CPU architecture significantly boosts multithreaded performance by up to 30%."
"The M5 Max chip pairs the 18-core CPU [...]. The new CPU architecture offers up to 15% higher multithreaded performance when compared to the M4 Max."
Seems strange. But then:
M4 Pro = 10 P-cores, 4 E-cores => M5 Pro = 6 S-cores, 12 P-cores
M4 Max = 12 P-cores, 4 E-cores => M5 Max = 6 S-cores, 12 P-cores
This means the M5 Pro is equal to the M5 Max in CPU cores. And the better buy from the CPU core perspective?
$3,399 14" M5 Pro MBP 64GB RAM 2TB SSD (20 GPU cores)
$4,299 14" M5 Max MBP 64GB RAM 2TB SSD (40 GPU cores)
$900 difference for 20 GPU cores
"With four additional CPU cores compared to the M4 Pro, the new CPU architecture significantly boosts multithreaded performance by up to 30%."
"The M5 Max chip pairs the 18-core CPU [...]. The new CPU architecture offers up to 15% higher multithreaded performance when compared to the M4 Max."
Seems strange. But then:
M4 Pro = 10 P-cores, 4 E-cores => M5 Pro = 6 S-cores, 12 P-cores
M4 Max = 12 P-cores, 4 E-cores => M5 Max = 6 S-cores, 12 P-cores
This means the M5 Pro is equal to the M5 Max in CPU cores. And the better buy from the CPU core perspective?
$3,399 14" M5 Pro MBP 64GB RAM 2TB SSD (20 GPU cores)
$4,299 14" M5 Max MBP 64GB RAM 2TB SSD (40 GPU cores)
$900 difference for 20 GPU cores
conmee
macrumors 6502a
Also the Max has double the bandwidth and double media encoders. So if AI or encoding is important, Max is the preferred choice.
chars1ub0w
macrumors 6502a
Thanks! Then, looking at the CPU performance only, I wonder how much double the memory bandwidth affects the benchmarks.
Going from 20 GPU cores to 40 GPU cores doubles the memory bandwidth because the chipset has double the memory controllers?
bradman83
macrumors 68000
The M1-M2 Pros/Maxes had the same CPU core count. In addition to double the GPU cores, you also get double the memory bandwidth, a second media encoder, double the number of displays it can drive. Probably still not worth $900 but it's more than just the GPU cores.
EDIT: The guy above me beat me to the punch!