RTX 5090 Outruns M5 Ultra on Models That Fit in VRAM
A comparison of Apple’s M5 Ultra Mac Studio and NVIDIA’s RTX 5090 finds the graphics card faster on a Qwen3.8-27B test, leading in prompt processing and token generation. Its advantage depends on the model fitting within 32 GB of VRAM; the Mac’s 256 GB of unified memory can hold models that exceed that limit, making capacity—not benchmark speed—the deciding factor for larger local AI workloads.
The 5090 is faster when the model fits
The choice between NVIDIA’s RTX 5090 and Apple’s M5 Ultra depends on two competing advantages: the 5090 has more memory bandwidth and compute, while the M5 Ultra can hold much larger models in memory. On the benchmarked model, the 5090 is faster. The Mac’s advantage is capacity: it can accommodate workloads that exceed the 5090’s 32 GB of video memory.
The comparison lists the RTX 5090 at 1.792 terabytes per second of memory bandwidth and about 419 FP16 teraflops. The M5 Ultra is listed at 1.2 terabytes per second and approximately 110 FP16 teraflops. That M5 Ultra compute figure is an estimate implied by Apple’s claim of up to 4.3 times the M2 Ultra’s performance; Apple does not publish a corresponding number directly.
Power figures depend on what is being measured. The table gives about 214 watts for the M5 Ultra while running a model, and less than 200 watts for the M3 Ultra. For the RTX 5090 it gives about 300 watts for the card alone. The spoken comparison adds that the 5090 can draw from 300 to 575 watts at sustained peak load. Those figures are not like-for-like system measurements, but they indicate the difference in power demands described.
Memory capacity reverses the performance comparison. The 5090 has 32 GB of VRAM. The M5 Ultra is currently orderable with 256 GB of unified memory; a 512 GB option is described as coming soon. That extra capacity matters when the model and its working overhead cannot fit in the 5090’s memory.
The benchmark favors the 5090 on speed
The head-to-head uses Qwen3.8-27B at 4-bit quantization with a 6,000-token prompt. Prompt processing reaches about 3,031 tokens per second on the RTX 5090 and 1,701 on the M5 Ultra. Time to first token is two seconds on the 5090 and four seconds on the Mac.
For token generation with multi-token prediction (MTP) turned off, the reported rates are about 74 tokens per second on the 5090 and 52 on the M5 Ultra. With MTP on, the estimates rise to about 136 and 100 tokens per second, respectively. The 5090 leads on every measure shown, though the MTP-on figures are estimates rather than settled results. The comparison cautions that llama.cpp and MLX are changing quickly as support and performance improve.
The rule of thumb offered for inference speed is bandwidth divided by model size. It helps explain why the 5090’s higher bandwidth translates into an advantage on a model that fits in its VRAM. It is not a complete description of every workload: the comparison’s practical qualification is that the model must fit along with its overhead. If it does not, the capacity constraint changes the calculation.
These measurements are a snapshot, not a guarantee of performance on every model or software setup. The reported results are tied to Qwen3.8-27B, its 4-bit quantization, the prompt length, and the inference engines being used. The source specifically describes the MTP-on figures as estimates and notes that the software is improving. The direction of the comparison is clear for this test; the exact rates may move.
Beyond 32 GB, capacity changes the decision
The memory comparison gives three model-size examples: 55 GB for 27B BF16, 155 GB for Flash-Next, and 162 GB for DeepSeek V4. Each exceeds the 5090’s 32 GB of VRAM, but each falls below the M5 Ultra’s listed 256 GB capacity. On the comparison’s figures, these models can fit in the Mac’s memory but not entirely in the 5090’s VRAM.
The chart says that once a model exceeds 32 GB, the 5090 spills into system RAM, with generation rates shown between 4.6 and 1.5 tokens per second. It does not assign an individual rate to each of the three model examples, so those figures should be read as the chart’s range rather than as exact per-model benchmarks. The practical point is the sharp distinction between running a model within VRAM and continuing after the model exceeds that limit.
Capacity comes at a price in the comparison: $5,000 for the RTX 5090 and $10,799 for the M5 Ultra, summarized on screen as twice the price for eight times the memory. Those prices make the decision more than a simple speed contest. For workloads that fit in 32 GB, the 5090 combines higher bandwidth with better results on the Qwen benchmark. For larger models, unquantized models, or mixture-of-experts models, the M5 Ultra’s 256 GB gives room the 5090 does not have in its own VRAM.