Apple's 512GB unified memory is the real star: M5 Ultra's benchmark scores beat the RTX 5080 — is it worth buying? These past two days, the hardware community has been flooded by a leaked benchmark: Apple's M5 Ultra GPU scored 360,000 in Geekbench, and some media outlets calculated that it "beats the RTX 5080 by 26%," with headlines getting more sensational by the minute. I've already broken down the inflated nature of these benchmark scores in another article.

💡 What You Will Learn

Apple's 512GB unified memory is the real star: M5 Ultra's benchmark scores beat the RTX 5080 — is it worth buying? These past two days, the hardware community has been flooded by a leaked benchmark: A

📜 Table of Contents

• Let's Set the Benchmark Aside First (Leaked Numbers)
• What 512GB Unified Memory Actually Means (For Readers Who Don't Use Apple)
• Running Large Models Locally — VRAM Is Always the Bottleneck (A Hard Ceiling)
• Why the RTX 5080 Can't Keep Up (The 16GB Ceiling)
• Time to Pour Some Cold Water (Bandwidth and Ecosystem)
• Should You Wait for It (From a Buyer's Perspective)

Apple's 512GB Unified Memory Is the Real Star: M5 Ultra's Leaked Benchmark Beats RTX 5080 — Is It Worth Buying?

The hardware community has been buzzing these past couple of days over a leaked benchmark: Apple's M5 Ultra GPU scored 360,000 in Geekbench, and some outlets calculated it "beats the RTX 5080 by 26%," with increasingly dramatic headlines. I already broke down the benchmark inflation in another piece — ranking Metal and OpenCL scores side by side in the same chart doesn't hold up. What I want to talk about here is the real star that got buried under the headlines: 512GB of unified memory.

Let's Set the Benchmark Aside First (Leaked Numbers)

The M5 Ultra is Apple's flagship chip launched alongside the Mac Studio at the end of August, with an official max configuration of 36-core CPU, 80-core GPU, and 512GB unified memory. This round of leaked benchmarks comes from an engineering sample with a 40-core GPU and 256GB of memory — a single database submission, sample size of one. The "41% faster than M3 Ultra" figure is comparable on the same interface; the rest of the cross-brand comparisons will have to wait for third-party hands-on testing.

What 512GB Unified Memory Actually Means (For Readers Who Don't Use Apple)

Unified memory (a design where the CPU and GPU share the same memory pool) is Apple's signature feature. On a regular PC, "RAM" and "VRAM" are two separate accounts, and the ceiling the GPU can touch is just that one VRAM slot. A common high-end build is 64GB of RAM plus a card with 32GB of VRAM — so the GPU has 32GB readily available. The M5 Ultra tops out at 512GB, which is 16 times that, with no boundary — give the model as much as it needs.

Running Large Models Locally — VRAM Is Always the Bottleneck (A Hard Ceiling)

Last month I wrote about the llama.cpp 0.3.0 update, where single-card inference got 30–50% faster — the local deployment toolchain keeps getting faster. But no matter how fast the tools get, if the model won't fit in VRAM, none of it matters. A 70B-class model still needs around 40GB of space after quantization (the practice of compressing a model to a smaller size), and general-purpose models in the hundreds of billions of parameters typically need hundreds of GB. Most enthusiasts go the multi-GPU split route, slicing the model across several cards — the configuration barrier and the cost of hitting pitfalls are both high. 512GB means a several-hundred-GB model can fit entirely into one machine, no splitting required.

Why the RTX 5080 Can't Keep Up (The 16GB Ceiling)

The RTX 5080 in those benchmark comparisons only has 16GB of VRAM — great for running 7B and 14B models, but beyond that you're just staring. Two cards get you 32GB, four cards get you 64GB, and both the money and the case space double each time. Apple's machine does it in one step, at the cost of the total system price and a closed ecosystem.

Time to Pour Some Cold Water (Bandwidth and Ecosystem)

Large unified memory is one thing, but bandwidth and the software ecosystem are two other hurdles. Nvidia's CUDA ecosystem has accumulated for nearly twenty years — training, inference, and framework support are all ready-made; Apple's MLX ecosystem is growing and sufficient for inference, but deep tinkering still takes effort. Large memory solves "does it fit" — how fast it runs and how smooth the ecosystem is will have to wait for retail-unit hands-on testing.

Should You Wait for It (From a Buyer's Perspective)

If your budget is 30,000 RMB or more and you just want to run large models at home, the M5 Ultra Mac Studio deserves a spot on your shortlist; if you already have a single 5090 or a multi-GPU setup, there's no need to get excited — what you're missing might just be another card. Take the engineering-sample benchmarks with a grain of salt, and don't rush to spend until retail units are tested.

If you had a 30,000 RMB budget, would you put it toward an RTX 5090, or wait for a 512GB Mac Studio? Let's talk in the comments.