Key points
- Apple M5 Ultra GPU has 80 cores and 1.2 TB/s bandwidth versus Nvidia GB10's 48 SMs and 273 GB/s.
- Testing used Qwen 3.8 27B with multi-token prediction enabled on both platforms via llama.cpp.
- The Mac Studio configuration costs over $12,300 and faces 13 to 15 week shipping delays.
The claim
Apple's system offers 1.2 TB/s of memory bandwidth, providing more than four times the decode bandwidth of Nvidia's 273 GB/s.reported
Evidence: reported
Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Sparkaifleet-tomshardware
The response
They used Artificial Analysis' AgentPerf-Local benchmark to gauge agentic AI workload performance across both systems.
Why this matters now
The testing focused on filling the Mac Studio's memory pool with large models like DeepSeek V4.1 Flash 1017 and GLM 5.3 Flash.reported The outlet previously tested the Qwen 3.8 27B model but found it did not fully exercise the available memory.reported This new testing addresses the capability of the system to handle larger, more demanding models.
Evidence: reported
Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Sparkaifleet-tomshardware
Evidence: reported
Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Sparkaifleet-tomshardware
What buyers should weigh
Enterprise IT buyers considering a $12,300 Mac Studio must weigh the cost against the 13 to 15 week shipping delay according to the published report The decision hinges on whether the M5 Ultra's delivered performance in agentic tasks justifies the premium over Nvidia's DGX Spark.reported Buyers should verify if llama.cpp with multi-token prediction meets their specific inference backend requirements.
Evidence: reported
Going deep on local AI performance benchmarking with Apple’s Mac Studio with M5 Ultra taking on Nvidia DGX Sparkaifleet-tomshardware
Why it matters
- High unified memory options up to 512GB enable the loading of cutting-edge large models that exceed the capacity of smaller dense models.
- The M5 Ultra’s 80-core GPU and 1.2 TB/s memory bandwidth provide significantly higher raw compute and decode bandwidth than the Nvidia DGX Spark’s GB10 chip.
- Local AI inference performance varies greatly from paper specifications, making delivered performance in agentic tasks more critical than raw specs alone.
- A system cost of $12,300 or more raises questions about the value proposition for local AI inference setups compared to other hardware options.
Topics: Mac Studio, M5 Ultra, Nvidia DGX Spark, local AI, benchmarking
Updated
Ask the article
Local keyword search of this article; no generated answers.Did this article give you the evidence you needed?
Your choice stays on this device. No public tally.Questions readers ask
How does Mac Studio M5 Ultra memory bandwidth compare to Nvidia GB10?
The M5 Ultra offers 1.2 TB/s, which is over four times the 273 GB/s provided by the GB10 chip.
What is the current shipping timeline for the tested Mac Studio configuration?
Tom's Hardware reported that shipping times are currently 13 to 15 weeks for the configured system.
Reading receipt
- Apple M5 Ultra GPU has 80 cores and 1.2 TB/s bandwidth versus Nvidia GB10's 48 SMs and 273 GB/s.
- Testing used Qwen 3.8 27B with multi-token prediction enabled on both platforms via llama.cpp.
- The Mac Studio configuration costs over $12,300 and faces 13 to 15 week shipping delays.
Still open: Does the M5 Ultra's higher bandwidth consistently outperform GB10 in all agentic tasks?
Still open: Will shipping times for high-memory Mac Studio configurations decrease soon?
