M3 Ultra Chip's Massive 80-Core Graphics Performance Revealed in Benchmark Result

[harmonthe3rd]

That misses the point of my comment. My comment is that having an M3 chip come out after an M4 doesn't make sense.

I fully understand that someone with an iPad has different needs than someone who wants a Mac Studio. Thats obvious.

Maybe I didn't articulate my thought process well. Having a M3 Ultra come out when there is a M4 line up makes sense for the following reason. Base, Pro, Max then Ultra chipsets. 1, 2, 3 and 4 are generations. A person buying a M3 Ultra knows, the M4 chipset can't be made into an Ultra. The person buying a M4 Max knows or has an idea of the processing power of the M3 Ultra and doesn't need, or want to send the extra money. If we were to use Intel as an example and I said something like in need an i5, you could give me a computer with an i5. a 2008 i5 and a 2024 i5 have the same name. It's the generation that is differentiator of the two. As with Intel and also in Apple's case, the expectation is for the consumer to know what type of processing power they need or want. The Mac mini skipped the M3 altogether, the Mac Studio went from a M2 Ultra to a M3 Ultra. It is likely that the Mac Studio does not see another Ultra labeling for a few years. For these reasons and may others it makes sense for the M3 Ultra to be released at the same time as the M4 Max.

 

[SpotOnT]

You "...can’t think of any real reason for Apple not to offer it..." You mean other than physics? Apple is building a different, more efficient paradigm. We will see how it plays out long term.

😂😂😂😂😂

There is no “physics” preventing macOS from using a second (or third or 16th) GPU for user defined tasks. That is just a software setting.

I am fine with Apples paradigm. I am not wanting Apple to change their paradigm. I am just wishing Apple would add a few software features in addition to their paradigm.

 

[skippermonkey]

Has anyone got their machine yet? Would love to see some actual Blender render times compared to M2 Ultra. Should be twice as fast, but be good to see realworld evidence.

 

[StuBeck]

Maybe I didn't articulate my thought process well. Having a M3 Ultra come out when there is a M4 line up makes sense for the following reason. Base, Pro, Max then Ultra chipsets. 1, 2, 3 and 4 are generations. A person buying a M3 Ultra knows, the M4 chipset can't be made into an Ultra. The person buying a M4 Max knows or has an idea of the processing power of the M3 Ultra and doesn't need, or want to send the extra money. If we were to use Intel as an example and I said something like in need an i5, you could give me a computer with an i5. a 2008 i5 and a 2024 i5 have the same name. It's the generation that is differentiator of the two. As with Intel and also in Apple's case, the expectation is for the consumer to know what type of processing power they need or want. The Mac mini skipped the M3 altogether, the Mac Studio went from a M2 Ultra to a M3 Ultra. It is likely that the Mac Studio does not see another Ultra labeling for a few years. For these reasons and may others it makes sense for the M3 Ultra to be released at the same time as the M4 Max.

I still don't think it makes much sense, but we aren't going to agree. Intel isn't coming out with 2008 i5's in 2024 for example.

If you're coming out with an M3 Ultra, come out with it when the M3 Pro and M3 Max processors came out. Thats all I'm saying.

 

[tweedledee1234]

I think the M3 Ultra is going to mainly appeal to the A.I. crowd with 512MB memory and 32 neural engine cores.

This machine is the AI workhorse. If you need 512GB of Unified (GPU) memory this is your baby.

These two beautiful people in this forum understand the point of the M3 Ultra.

Many people here discussing the M3 Ultra seem to misunderstand its primary purpose i.e running large language models (LLMs) locally. The M3 Ultra (particularly the 512GB version) isn’t designed specifically for gaming or rendering (although it would perform well in those areas), but it excels in local LLM deployment. If you need something for games or rendering just get the M4 Max. For the small minority of badass renderers dealing with massive objects, I guess the M3 Ultra with 256GB makes sense.

For example, an M3 Ultra with 512GB of RAM lets us run models like DeepSeek R1 or Qwen locally without needing to buy two or more RTX 4090s (which needs FP6 or FP8 quantization) or shell out around $70k for a DGX A100 to run tasks in BF16 precision. The huge memory capacity is crucial because the reasoning process generates a massive number of output tokens. The memory footprint scales nearly quadratic as you increase the k:v cache size.