Apple’s M5 Ultra makes its clearest statement in workloads that can exploit its 36 CPU cores and enormous unified-memory pool, rather than simply trying to overpower a high-end PC GPU. The comparison establishes that distinction early by putting the Mac Studio against a Ryzen 9 9950X3D system with an RTX 5090, then testing CPU rendering, GPU acceleration, local AI, video production and gaming. That breadth makes the results more useful than a single benchmark showdown, although the two machines are not always direct equivalents in hardware, software paths or configuration.
CPU performance is presented as the Mac Studio’s most consistent advantage. The M5 Ultra reportedly beats the Ryzen system in single-core testing and approaches twice its multi-core score, while maintaining that performance during sustained testing. Blender CPU rendering, Houdini and HandBrake similarly favor the Mac, with the presentation also reporting roughly 150 watts of CPU power consumption against about 200 watts for the Ryzen. These results make a persuasive case for the M5 Ultra as a highly efficient workstation CPU within the tests performed, though the later acknowledgment that a Threadripper-class processor would be the more natural PC counterpart is important context that could have been emphasized earlier.
GPU testing is more complicated, and the presentation does a good job of exposing why benchmark configuration matters. The Mac Studio initially finishes a Blender task about 40 percent faster than the RTX 5090 using Nvidia’s CUDA path, but enabling OptiX reverses the result and puts the 5090 ahead. Rather than burying that qualification, the comparison explicitly explains that Metal and Nvidia’s acceleration options represent different software stacks. That transparency is valuable because it prevents an eye-catching initial number from becoming a misleading blanket claim about relative GPU performance.
Local AI provides the most interesting section because the results change dramatically with model size. Models that fit within the RTX 5090’s 32GB of VRAM generally run substantially faster on Nvidia, even in several cases where the Mac benefits from Apple’s MLX framework. The methodology also distinguishes MLX tests from GGUF comparisons instead of pretending every result is directly equivalent. Once the 70-billion-parameter model exceeds the 5090’s VRAM, however, the reported result swings enormously toward the M5 Ultra, which completes the workload more than 20 times faster because its much larger unified-memory pool avoids the same system-memory spillover. That is a particularly useful demonstration of how memory capacity and architecture can matter more than raw GPU throughput for specific local-model workloads.
Creative applications produce a similarly workload-dependent picture. DaVinci Resolve Studio has the two desktop systems essentially tied during the tested 4K H.265 export, while Premiere Pro heavily favors the M5 Ultra, with even the M5 Max MacBook Pro reportedly finishing around the same time as the RTX 5090 system. The presentation sensibly attributes this to platform and application optimization rather than treating hardware specifications alone as destiny. Gaming then delivers the opposite outcome: the RTX 5090 wins by large margins across the tested titles, while games requiring a translation layer on macOS impose another performance penalty. The acknowledgment that neither workstation configuration makes much sense purely as a gaming purchase keeps this section in perspective.
Price comparisons are informative but also the least stable part of the argument. The discussion relies heavily on unusually inflated RTX 5090 pricing and describes the tested 256GB Mac Studio as extremely expensive in its own right, including a $4,000 memory upgrade. It also notes lengthy reported Mac Studio delivery delays and concedes that a Threadripper Pro configuration would provide a more comparable high-core-count PC workstation, albeit at an even higher stated system cost when paired with the inflated GPU. These caveats strengthen the presentation, but they also demonstrate why the broad question of which platform offers better value cannot be settled by this particular hardware pairing alone.
Pros
- Tests a wide range of relevant workstation workloads instead of relying on synthetic benchmarks alone.
- Clearly explains the Blender CUDA-versus-OptiX result rather than using the more favorable Mac result without qualification.
- Distinguishes MLX and GGUF local-AI tests and identifies where the methodology prevents direct comparisons.
- The 70-billion-parameter test effectively demonstrates the practical advantage of large unified memory once a model exceeds the RTX 5090’s VRAM.
- Includes sustained CPU testing, power consumption and application-specific results alongside headline performance numbers.
- Acknowledges that Threadripper would be a more appropriate CPU-class comparison and that both tested configurations are expensive.
Cons
- The Ryzen 9 9950X3D is not a direct CPU-class counterpart to the 36-core M5 Ultra, limiting how broadly the large CPU victories can be interpreted.
- Several AI benchmarks use different backends between platforms, making part of the performance comparison inherently uneven despite the disclosure.
- The value argument depends heavily on unusually inflated RTX 5090 pricing that could change substantially.
- The tested Mac’s 256GB configuration gives the unified-memory architecture an especially favorable showcase without equally extensive testing of more typical configurations.
- Benchmark results are presented largely as percentages and conclusions, with limited deeper examination of repeatability, detailed settings or variability between runs.
A wide-ranging benchmark suite makes a convincing case that the M5 Ultra is particularly formidable for CPU-heavy production and local AI workloads that can exploit its unified memory, while the RTX 5090 retains major advantages when its raw GPU performance and optimized software paths can be fully used. Careful disclosures about OptiX, MLX, memory limits and the imperfect CPU matchup add credibility, though the rapidly changing hardware prices and unequal configurations make the broader value comparison less definitive.
