Mac vs RTX 5090 for LLMs: M5 Max, M5 Ultra and the RTX 5090 compared on the same models (November 2026)

On the same llama.cpp benchmark (Llama 2 7B at Q4_0), an RTX 5090 generates 290 tokens/s against 179 for an 80-core M5 Ultra Mac Studio and 120 for a 40-core M5 Max, and processes prompts 2.8 and 4.4 times faster. That follows bandwidth (1,792 GB/s against 1,228 and 614) for generation and tensor compute for prompts. The Macs win on capacity: the M5 Max takes 128 GB and the M5 Ultra 512 GB, against the 5090's 32 GB, so by the bandwidth arithmetic a dense model above about 28 GB of weights will usually run faster on a big Mac than on a 5090 that has to spill layers into system RAM. Data checked 7 November 2026; Apple's M5-generation Macs are the current ones.

Benchmark: llama-bench, Llama 2 7B Q4_0, from the llama.cpp CUDA scoreboard (RTX 5090, without flash attention, commit 8cf6b42; 14,970 and 300 with it) and the Apple Silicon thread (M5 rows at commit c1d0e7a, build 10621, thread updated 25 to 30 August 2026). Bandwidth as listed in those threads; Apple quotes 1.2 TB/s. Community-submitted numbers. Mac prices from Apple's 25 August 2026 press release (USD before tax) and hardwaredealz.com's German price comparison (EUR including German VAT, card only), read 7 November 2026; they are not directly comparable.

Which Macs are current on 7 November 2026?

The M5 Max and M5 Ultra are the high-memory chips, and a base M6 arrived in late October for the cheaper machines. Apple's Mac Studio specs page lists the M5 Max (32 or 40 GPU cores, 36 to 128 GB, 460 or 614 GB/s) and the M5 Ultra (64 or 80 GPU cores, 96, 256 or 512 GB, 1.2 TB/s). Apple introduced them on 25 August 2026 at $2,499 and $5,499, with the 512 GB option "coming in late October".

• MacBook Pro 14 and 16 inch with M5 Pro or M5 Max, the new PROLED display model announced on 28 October 2026 (Apple newsroom): up to 128 GB at 614 GB/s on the 40-core M5 Max, per the specs page. A 128 GB M5 Max laptop is the only portable machine here that loads a 70B model whole.
• 14-inch MacBook Pro and iMac with M6, Apple's first 2 nm chip: 16 to 32 GB at up to 170 GB/s, from $1,999.
• Mac mini with M6 or M5 Pro: up to 32 GB at 170 GB/s or 64 GB at 307 GB/s, per the Mac mini specs.

The M5 Ultra is two M5 Max chips joined by UltraFusion, per Apple's chip announcement, which is why its bandwidth is double. If your question is a Mac against NVIDIA's own desktop box, the DGX Spark vs Mac Studio comparison covers that matchup; this page is about the RTX 5090.

How much faster does an RTX 5090 generate text than an M5 Max or M5 Ultra?

1.6 times an M5 Ultra and 2.4 times an M5 Max on the same 7B model, close to the bandwidth ratios of 1.46 and 2.92. Generating a token reads the whole model (or, for a mixture-of-experts model, its active experts) from memory once, so decode speed is roughly bandwidth divided by model bytes. Llama 2 7B at Q4_0 is about 3.8 GB (6.74B parameters at 4.5 bits, my arithmetic), which gives these ceilings and efficiencies:

ceiling = bandwidth / model bytes
RTX 5090:  1,792 GB/s / 3.79 GB = 473 tok/s  -> measured 290 = 61%
M5 Ultra:  1,228 GB/s / 3.79 GB = 324 tok/s  -> measured 179 = 55%
M5 Max:      614 GB/s / 3.79 GB = 162 tok/s  -> measured 120 = 74%

The M5 Max uses its bandwidth best and the M5 Ultra worst. The M3 Ultra shows the same pattern in that thread (92 tok/s from 800 GB/s, against 83 from 546 GB/s for a 40-core M4 Max), probably because the work is split across dies. Against the previous generation in the same thread, the 40-core M5 Max generates 1.44 times faster than a 40-core M4 Max (120 against 83 tok/s), and the 80-core M5 Ultra 1.94 times faster than an 80-core M3 Ultra (179 against 92).

For a model people run today, take Qwen3.8-27B, a dense model whose 4-bit GGUF is 18.97 GB in ggml-org's repository and which fits the 5090. By the same arithmetic the ceilings are 94 tok/s on a 5090, 32 on an M5 Max and 16 on an M5 Pro. Speculative decoding moves those numbers more than the hardware does: LM Studio reports 74 tok/s for Qwen3.8-27B on a 48 GB M5 Pro with its Splash engine and a DFlash 2 draft model (18 September 2026), about 4.6 times the plain-decode ceiling for a 19 GB 4-bit build; LM Studio does not state its quantization. DGX Spark reports for the same model are collected on the Qwen3.8-27B model page.

How big is the prompt-processing gap?

Smaller than it was, but still about 3 times against an M5 Ultra and 4.4 times against an M5 Max. Reading a prompt is compute-bound, and the M5 generation added a Neural Accelerator to every GPU core. I think that is the biggest change for Mac LLM users in two years: the 40-core M5 Max processes prompts 3.6 times faster than a 40-core M4 Max (3,220 against 886 tok/s) in the llama.cpp thread, and Apple's own MLX study found 3.3 to 4.1 times faster time to first token on a base M5 than an M4, with generation only 19 to 27 percent faster.

For long prompts the gap shows up as waiting time. No one has published M5 numbers for gpt-oss-20b at long context yet, so the table mixes measured older Macs with my extrapolation:

Measured rows from the llama.cpp gpt-oss performance discussion (August 2025 builds, MXFP4). Estimates: the M3 Ultra and M4 Max figures multiplied by two factors, Apple's LM Studio time-to-first-token claims on the Mac Studio page (M5 Ultra 2.4x M3 Ultra, M5 Max 2.8x M4 Max; 8K prompt, 14B 4-bit, LM Studio 0.4.19, tested July 2026) and the pp512 ratios from the llama.cpp thread (3.4x and 3.6x). Newer llama.cpp builds are faster on all machines.

A 5-second wait and a 20-second wait feel very different in a coding agent that resends 30,000 tokens of context every turn. For chat with short prompts, the Mac's prefill is now fast enough that most people will not notice.

What fits in 32 GB, and what needs a Mac?

A 5090 holds dense models up to about 30B parameters at 4-bit and MoE models up to about 45B at 4-bit, with some room for context. Beyond that you need either more GPUs, CPU offload, or unified memory. My arithmetic, weights only, with 4-bit at about 4.5 bits per parameter:

Ceilings are bandwidth divided by weight bytes for dense models, before efficiency losses of 25 to 45 percent seen above. macOS keeps part of unified memory for itself, so a 128 GB Mac does not give 128 GB to the model. KV cache comes on top; the memory calculator adds it for a given context length.

For scale, a 96 GB RTX PRO 6000 Max-Q runs gpt-oss-120b at 197 tok/s in the same llama.cpp gpt-oss thread (September 2025 build). A 5090 can run it only by moving expert layers to system RAM (DGX Spark figures for the same model are on the gpt-oss-120b page); how much that costs is the subject of the planned llama.cpp MoE offload guide. On a 128 GB M5 Max it simply loads.

Which software runs on each?

The 5090 runs everything built for CUDA: vLLM, SGLang, TensorRT-LLM, llama.cpp, Ollama, ExLlamaV3, and most fine-tuning libraries. That matters most for serving several people at once, where vLLM's continuous batching turns one card into a shared endpoint. On a Mac you use llama.cpp's Metal backend, Ollama on top of it, or Apple's MLX through mlx-lm or LM Studio. New model architectures usually land in CUDA engines first.

Concurrency on a Mac is improving. LM Studio reports 170 tok/s combined across four parallel requests on the M5 Pro with Splash, 3.9 times the next engine it tested. That is a vendor figure for its own engine, but it shows multi-user serving on Apple Silicon is moving. A rough rule I use: one user and big models, Mac; several users or any training, CUDA.

What do they cost to buy and run?

The 5090 is no longer the cheap option. hardwaredealz.com's German comparison listed the cheapest RTX 5090 at €6,249 (INNO3D X3) on 7 November 2026, with most cards between €7,800 and €9,000 and no offer for the Founders Edition, and you still need a PC around it. That card price alone sits in the range of the M5 Ultra Mac Studio's $5,499 US starting price, although currency and tax treatment differ. The M5 Max Mac Studio starts at $2,499; the 128 GB M5 Max and 512 GB M5 Ultra configurations cost more, and I could not read Apple's configurator prices.

Power is the other running cost. NVIDIA rates the 5090 at 575 W total graphics power and recommends a 1,000 W power supply (NVIDIA's comparison page); Apple lists 480 W maximum continuous power for the whole Mac Studio and 155 W for the Mac mini. Neither publishes a typical draw while generating text. Euro-per-month figures for each machine are in the planned electricity cost article.

Which should you buy?

If your models fit in 32 GB, or you serve more than one person, or you fine-tune, I'd take the RTX 5090: it is 1.6 to 2.4 times faster at generating and 3 to 4 times faster at reading prompts than the Macs. If you want 70B dense models or 100B-plus MoE models on one quiet desk machine, I'd take a 128 GB M5 Max Mac Studio (or, if you need CUDA at that size, read the DGX Spark vs RTX 5090 comparison), and the M5 Ultra only if you need more than 128 GB or the extra 1.6 times of decode speed. For the 5090's model shortlist, see the planned best LLMs for an RTX 5090; GPUwerk rents one RTX 5090 machine by the minute if you want to time your own prompts before spending €6,000 or more. GPUwerk has published no RTX 5090 or Mac measurements of its own; every number above comes from the linked threads and vendor pages. For a Windows laptop with up to 128 GB of unified memory, see DGX Spark vs N1X laptop.

FAQ

Sources

Every figure above is attributed to its source in the sentence or table where it appears. Pages were read on 7 November 2026 unless a date says otherwise.

• Mac Studio technical specifications, Apple
• Mac Studio product page (LM Studio test footnotes), Apple
• Apple introduces new Mac Studio with M5 Max and M5 Ultra, Apple Newsroom, 2026-08-25
• Apple introduces M6 and M5 Ultra for a big leap in performance and AI compute, Apple Newsroom, 2026-08-25
• Apple Newsroom archive (October 2026 Mac releases), Apple Newsroom
• Apple introduces the new 14-inch MacBook Pro with M6, Apple Newsroom, 2026-10-27
• MacBook Pro technical specifications, Apple
• Mac mini technical specifications, Apple
• Exploring LLMs with MLX and the Neural Accelerators in the M5 GPU, Apple Machine Learning Research, 2025-11-19
• Performance of llama.cpp on Apple Silicon M-series (discussion #4167), ggml-org/llama.cpp on GitHub, 2026-08-30
• Performance of llama.cpp with CUDA (discussion #15013), ggml-org/llama.cpp on GitHub
• gpt-oss performance with llama.cpp (discussion #15396), ggml-org/llama.cpp on GitHub, 2025-08-18
• Splash Engine: the fastest local Qwen3.8 on Apple Silicon, LM Studio, 2026-09-18
• Compare GeForce graphics cards, NVIDIA
• GeForce RTX 5090, NVIDIA
• NVIDIA GeForce RTX 5090 32GB price comparison, hardwaredealz.com
• ggml-org/Qwen3.8-27B-GGUF file listing, Hugging Face