3 ms·
> M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks. It also provides up t
by egonschiele 1mo ago
> M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks. It also provides up to 170GB/s of unified memory bandwidth — a 10 percent increase over M5 and a 2.5x increase over M1.
Isn't 170GB/s slow for bandwidth?
- tristor 1mo agoKinda. Strix Halo does 256GB/s of memory bandwidth, and is significantly slower than an M5 Max (614GB/s). Feels like intentional market segmentation?
- kamranjon 1mo agoYou can get the m5 pro in the Mac mini with 307GB/s at 64gb of memory it’s $2899
- blints 1mo agoIt is. That's the mac mini. For local LLMs you would want the Mac Studio, which tops out at 1.2TB/s.
- scosman 1mo agoIt's fast by computer standards and excellent for entry level chip. The Pro/Max/Ultra chips are always faster. Compared to something like VRAM it's slow.
- mkesper 1mo agoIt's actually quite fast, more bandwidth than NVIDIA RTX 4090 at 1.0 TB/s according to https://www.digitalocean.com/community/tutorials/gpu-memory-bandwidth https://www.digitalocean.com/community/tutorials/gpu-memory-... Compute power will still be different, though.
- angoragoats 1mo agoThe thread you're replying to is asking about the 170GB/sec memory bandwidth of the base M6, not the faster spec of the Ultra chips.
- mkesper 1mo agoFor max memory bandwidth you need to buy the Ultra versions (M5 Ultra: 1,2TB/s, this gets comparable to real GPUs regarding the memory bandwidth).
- matja 1mo agoIt's higher bandwidth than any dual-channel DDR5 desktop machine, but Apple never quote the memory latency, so hard to compare otherwise.
- geraneum 1mo agoWould that be a difficult comparison to do fairly sine one is SoC and the other isn’t?
- matja 1mo agoI meant that it's a difficult comparison between platforms without having all the details like memory latency, how the memory controller load increases latency (can increase by 4x on some platforms), cache latency, TLB entries, etc. Extremely high bandwidth is great for copying data, but not as relevant for walking chains of pointers, where latency/cache/TLB entries are more important. So just saying "high bandwidth == better" is true when other variables are the same, but they rarely are, especially in comparison to x86-64 offerings. All of these are independent that it is a SoC with on-package DRAM.
- kamranjon 1mo agoYou would want to get the M5 pro version with 307gb/s if you were interested in running local LLMs.
- diabllicseagull 1mo agoM6 and M5 Ultra both have comparable memory bandwidth per GPU core. I think they will perform well.