M5 Ultra
Apple's first processor built on a quad-chip fusion architecture, connecting two M5 Max dies via the UltraFusion interconnect. It supports up to 512 GB of unified memory and 1.2 TB/s memory bandwidth, enabling local inference of models with hundreds of billions of parameters.
Core Facts
The UltraFusion inter-chip bandwidth of the M5 Ultra has increased from approximately 2.5 TB/s on the M3 Ultra to 4.4 TB/s, with a 6x improvement in connection density.
The new Mac Studio uses a unified memory design, with the 128 GB memory version starting at RMB 19,999 for the domestic Chinese market and the 512 GB version starting at RMB 46,999.
Under the M5 Ultra's 1.2 TB/s bandwidth, the theoretical speed of a 100 GB quantized model is approximately 12 tokens/s, a 200 GB model approximately 6 tokens/s, and a 300 GB model approximately 4 tokens/s.
Timeline (last 90 days)
The full version of the MuseSpark Open model requires 1TB of memory, roughly equivalent to two fully specced M5 Ultra Mac Studios, priced at about $40,000.
Muse Spark Open can run on two Mac Studios equipped with M5 Ultra chips, together with the latest MLX framework.
The M5 Ultra GPU neural accelerator improved the TTFT of Qwen 3.8 and Mistral Medium 3.5 by up to 4.6 times.
Four M5 Ultras networked with dual Thunderbolt cables reduced Muse Spark Open's TTFT from about 50.5 seconds to 40.6 seconds at 65536 tokens.
On two M5 Ultras, tensor parallelism for Muse Spark Open at 65536 tokens is inferior to pipeline parallelism; after scaling to four machines, tensor parallelism overtakes it.
With tensor parallelism across four M5 Ultras, Muse Spark Open's decoding speed rose from 15.74 tokens/s under two-machine pipeline parallelism to 36.73 tokens/s, and Qwen 3.8 rose from 19.44 to 50.89 tokens/s.
For users running quantized models below 70B, the RTX 5090 is the faster choice; for developers who need to run very large models, unquantized models, or MoE architectures, the M5 Ultra is almost the only option that can complete the task on a desktop-class device.
The M5 Ultra has a computing performance of about 110 TFLOPS, while the RTX 5090 can reach 419 TFLOPS, about four times that of the former.
A Reddit user ran benchmark tests of multiple models including GLM-5.3-Flash on an M5 Ultra 256GB 80-core Mac Studio.
If the GPU is already a bottleneck at the 256GB configuration, doubling the memory to 512GB will not improve GPU compute, and loading larger models will instead result in slower inference.
40 more timeline events
All Facts (20)
The UltraFusion inter-chip bandwidth of the M5 Ultra has increased from approximately 2.5 TB/s on the M3 Ultra to 4.4 TB/s, with a 6x improvement in connection density.
85%VerifiedThe new Mac Studio uses a unified memory design, with the 128 GB memory version starting at RMB 19,999 for the domestic Chinese market and the 512 GB version starting at RMB 46,999.
70%VerifiedUnder the M5 Ultra's 1.2 TB/s bandwidth, the theoretical speed of a 100 GB quantized model is approximately 12 tokens/s, a 200 GB model approximately 6 tokens/s, and a 300 GB model approximately 4 tokens/s.
65%VerifiedIf bandwidth reached 2.4 TB/s, the theoretical speed ceiling for a 100 GB model could reach 24 tokens/s, a 60 GB model 40 tokens/s, and a 35 GB model close to 69 tokens/s.
65%VerifiedThe M5 Ultra's peak AI compute performance is more than 4x higher than that of the M3 Ultra.
65%VerifiedThe UMA memory bus bandwidth of Apple's M2 Ultra has reached 800 GB/s.
65%VerifiedThe Mac Studio M5 Ultra starts at $5,499, with up to 512GB of memory and memory bandwidth of 1.2TB per second, with the 512GB version expected to ship in late October.
80%VerifiedApple introduced a new generation Mac mini equipped with the M6 chip.
80%VerifiedThe Mac mini M6 starts at $899, with up to 32GB of unified memory, going on sale September 22, 2026.
75%VerifiedThe M5 Ultra is very likely to adopt UltraFusion dual-chip interconnect packaging technology, combining two M5 Max-level chips into a single logical unit.
65%Partially VerifiedMuse Spark Open can run on two Mac Studios equipped with M5 Ultra chips, together with the latest MLX framework.
65%Partially VerifiedThe M5 embeds a dedicated Neural Accelerator for matrix multiplication inside every GPU core.
65%Partially VerifiedUnder heavy GPU load, the M3 Ultra consumes approximately 100 W, while the M5 Ultra surges to over 400 W.
65%Partially Verified统一内存架构使CPU、GPU与NPU共享同一块内存池，避免数据在处理单元间反复搬运的开销
65%UnverifiedThe M5 Ultra version of the Mac Studio is equipped with up to 80-core GPU and up to 512GB of unified memory
60%UnverifiedCompared to the M3 Ultra, the M5 Ultra shows approximately 33.5% improvement in Geekbench single-core and approximately 32.5% improvement in multi-core scores.
60%UnverifiedIn the time-to-first-token test on a 70-billion-parameter model, the M3 Ultra took 31.6 seconds while the M5 Ultra took only 7.7 seconds, achieving approximately 4.1x, close to Apple's claimed 4.3x.
60%UnverifiedOn the 70-billion-parameter DeepSeek R1 distillation model, because the model overflows the RTX 5090's 32 GB of VRAM, the M5 Ultra is more than 20 times faster than the 5090.
60%UnverifiedThe M5 Ultra's multi-core performance score is nearly twice that of the 9950X3D, and it maintains the same advantage in Cinebench 2026.
60%UnverifiedFor small-to-medium models that fit within 32 GB of VRAM, the RTX 5090 leads (approximately 45% on Gemma 4 12B and over 75% on GPT-OSS 20B), but the lead on Qwen 3 27B drops from 47% at Q4 to 23% at Q8.
60%