M3 Ultra
A high-performance chip developed by Apple, representing the top-tier model in the M3 series. It is featured in professional devices such as the Mac Studio. The chip's GPU cores do not include a dedicated Neural Engine accelerator, distinguishing it in certain respects from other configurations within the M3 lineup.
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 M5 Ultra's peak AI compute performance is more than 4x higher than that of the M3 Ultra.
Under heavy GPU load, the M3 Ultra consumes approximately 100 W, while the M5 Ultra surges to over 400 W.
Timeline (last 90 days)
One user spent $11,000 to buy an M3 Ultra.
The Mac Studio is configured with an M3 Ultra chip and 96GB of unified memory, leaving more than 30GB to simultaneously run Immich, Nextcloud and a Kubernetes cluster.
Mac users should note: rapid-mlx reports that the M3 Ultra measured a peak of 72.7 GB at 1K context, while also stating that 96 GB Macs have not been verified and that image input and MTP are not enabled; to use vision capabilities on a Mac, you need to go through llama.cpp with an mmproj file.
The M3 Ultra's 80-core GPU can directly hold all 32 layers of the model.
The speculative decoding gains observed on the M5 are only between 4% and 40%, and the M3 Ultra shows almost no improvement.
The M3 Ultra begins to degrade at around 16K tokens, and by 500K tokens its incremental speed drops to one-third to one-quarter of its peak, whereas the Spark maintains about one-half at the same point.
After two to three minutes of full load, the DGX Spark's CPU temperature exceeds 90°C, whereas under the same load the M3 Ultra only reaches 60°C and stabilizes at around 72°C under sustained full load.
At a 4K×4K×16K matrix multiplication scale, the M5 Ultra achieved 91 TFLOPS, while the M5 Max and M3 Ultra achieved 34 and 22 TFLOPS respectively.
In an FP16-precision 4K×4K×4K transposed matrix multiplication task, the M5 Ultra with Neural Accelerator enabled surpassed 107 TFLOPS, which is 3.51 times its own non-accelerated performance and 4.79 times that of the M3 Ultra.
When generating a 5-second 768P video with Fast Cache enabled, the M3 Ultra took 37 minutes and 14 seconds, while the M5 Ultra took 9 minutes and 24 seconds, nearly 4 times faster.
19 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 M5 Ultra's peak AI compute performance is more than 4x higher than that of the M3 Ultra.
65%Partially VerifiedUnder heavy GPU load, the M3 Ultra consumes approximately 100 W, while the M5 Ultra surges to over 400 W.
65%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%UnverifiedIn AmorphousDiskMark tests, the M5 Ultra achieves sequential read speeds of 14,888 MB/s and write speeds approaching 20,000 MB/s, while the M3 Ultra reads at approximately 6,500 MB/s and writes at 2,700 MB/s.
60%UnverifiedThe Mac Studio is equipped with M-series chips, such as the M2 Max, M2 Ultra, and subsequent updated versions.
60%UnverifiedMini PCs equipped with a unified memory architecture can run large models with 70 billion parameters (70B), whereas high-end discrete graphics cards cannot fully load them due to insufficient VRAM.
60%UnverifiedOne user spent $11,000 to buy an M3 Ultra.
50%UnverifiedThe Mac Studio is configured with an M3 Ultra chip and 96GB of unified memory, leaving more than 30GB to simultaneously run Immich, Nextcloud and a Kubernetes cluster.
50%UnverifiedMac users should note: rapid-mlx reports that the M3 Ultra measured a peak of 72.7 GB at 1K context, while also stating that 96 GB Macs have not been verified and that image input and MTP are not enabled; to use vision capabilities on a Mac, you need to go through llama.cpp with an mmproj file.
50%UnverifiedThe M3 Ultra's 80-core GPU can directly hold all 32 layers of the model.
50%UnverifiedThe M3 Ultra begins to degrade at around 16K tokens, and by 500K tokens its incremental speed drops to one-third to one-quarter of its peak, whereas the Spark maintains about one-half at the same point.
50%UnverifiedAfter two to three minutes of full load, the DGX Spark's CPU temperature exceeds 90°C, whereas under the same load the M3 Ultra only reaches 60°C and stabilizes at around 72°C under sustained full load.
50%UnverifiedThe speculative decoding gains observed on the M5 are only between 4% and 40%, and the M3 Ultra shows almost no improvement.
50%UnverifiedIn an FP16-precision 4K×4K×4K transposed matrix multiplication task, the M5 Ultra with Neural Accelerator enabled surpassed 107 TFLOPS, which is 3.51 times its own non-accelerated performance and 4.79 times that of the M3 Ultra.
50%UnverifiedAt a 4K×4K×16K matrix multiplication scale, the M5 Ultra achieved 91 TFLOPS, while the M5 Max and M3 Ultra achieved 34 and 22 TFLOPS respectively.
50%UnverifiedWhen generating a 5-second 768P video with Fast Cache enabled, the M3 Ultra took 37 minutes and 14 seconds, while the M5 Ultra took 9 minutes and 24 seconds, nearly 4 times faster.
50%UnverifiedThe M5 Ultra GPU's Solar Bay Extreme score is 108.4% higher than that of the M3 Ultra.
50%UnverifiedIn the Cinebench R26 multi-core test, the M5 Ultra scored 17,943 points versus 10,425 for the M3 Ultra, representing approximately 1.72x, far exceeding Apple's claimed 1.3x.
50%