返回资讯中心
外部精选
软件工程
#开发者实时流

Two Mac Studios Are Not One 1 TB Computer

Two Mac Studios Are Not One 1 TB Computer

Hacker News 最新durul2026/10/18 · 2026/10/1819 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Artificial Intelligence

LLM

Machine Learning

Apple

Two Mac Studios Are Not One 1 TB Computer

Durul Dalkanat

10 min read·

1 hour ago

--

Photo by Dima Solomin on Unsplash

Meta’s “one supercomputer” line is right about capacity and misleading about speed. That difference matters to anyone considering a second Mac.

On October 15, 2026, Meta released Muse Spark Open, its new open-weight model. What made it news was that a model with 816 billion parameters runs on two Mac Studios. Here is the line from Meta’s blog:

“The result is two Mac Studios behave like one supercomputer with a terabyte of unified memory.”

MacRumors and Engadget repeated the line word for word. The same day, Apple published its own measurements. Apple’s article makes no such claim: it describes a model that will not fit on a single machine, so half of it goes on one Mac and half on the other.

So do two Macs really behave like one computer? The short answer: there is a terabyte of memory, but it isn’t one memory. That distinction explains both why the setup works and what to check before buying a second Mac.

The 512 GB Mac Studio is not on sale yet; Apple says it is coming in late October. Apple had not announced its price as of October 15, but according to MacRumors, even the 256 GB model costs more than $10,000. Engadget puts the alternative, four 256 GB Mac Studios, at about $38,000.

Together, the two Mac Studios have 2 × 512 = 1,024 GB of memory: a terabyte. The model’s 4-bit version is about 493 GB. The weights alone fit under 512 GB, but that leaves no room for the operating system, the intermediate data created during computation, or the conversation’s context. 🧐

Apple says so plainly: the model will not fit on a single machine, so the solution is to “store half of the model on one Mac Studio and the other half on another” (Apple).

In terms of capacity, then, two Macs together hold a model that one Mac cannot. The “terabyte” in Meta’s line is accurate in that sense. 👏🏻

Inside a Mac, “unified memory” means the CPU and the graphics processor (GPU) share the same memory. On the M5 Ultra, the GPU reads that memory at 1.2 TB per second (Apple).

Between two Macs, there is a Thunderbolt 5 cable. At best, it carries about 10 GB per second; with two cables, as in Apple’s setup, about 20 GB (MacGeneration). That is 60 to 120 times slower than the Mac’s own memory.

The gap in latency is even larger: according to MacGeneration, memory responds roughly 1,000 times faster than this link. So each Mac’s GPU can read only its own memory quickly. The two machines’ memories are not one pool; they are two separate pools joined by a cable.

A thought experiment makes the difference concrete. What if one Mac’s GPU had to read the half of the model stored on the other Mac over the cable?

The 42 billion parameters active for each token come to about 24 GB, even in the 4-bit version. Half of that, about 12 GB, would have to cross the cable for every token. Even with two cables, that is about 0.6 seconds per token: fewer than 2 tokens per second.

Apple measured 25 tokens per second. In other words, the system works not because it behaves like one memory but because it avoids doing so.

Interestingly, the “supercomputer” comparison is apt. Supercomputers, too, are made of many nodes, each with its own memory, connected by a fast network. The misleading part is “unified memory.”

Three things come together: the model’s structure, the way it is split, and the latency of the link between the two Macs.

Muse Spark Open is a mixture-of-experts (MoE) model. Each layer has 256 “expert” subnetworks, and only 8 of them are chosen for each token (NVIDIA). That is why only 42 billion of the 816 billion parameters, roughly a twentieth, are active for each token.

The sentence just before Meta’s “supercomputer” line makes this point: “the chip’s 1.2 TB/s of memory bandwidth is fast enough to read the small slice each token needs.” That part is accurate. 👏🏻

There are two ways to split a model across two machines (Apple):

With Muse Spark Open on two Macs, the pipeline split produced 15.7 tokens per second and the tensor-parallel split 25.0 (Apple).

In a tensor-parallel split, the weights never move; each Mac reads its half from its own memory. All that crosses the cable is intermediate results, combined twice in every layer. A rough calculation puts this at 2–4 MB per token: 62 layers × 2 combinations × 8,192 values × 2–4 bytes (architecture).

A cable carrying 10 GB per second moves that in 0.2–0.4 milliseconds, while a token takes about 40 milliseconds. Cable capacity is not the problem.

In a tensor-parallel split, the two Macs wait for each other about 124 times per token. Over an ordinary network connection (TCP), a message on this link takes about 300 microseconds (Gigazine). That adds about 37 milliseconds of waiting per token; by a rough estimate, it would cut 25 tokens per second to 13–14, below even the pipeline split.

That is why RDMA is critical. With RDMA, added in macOS 26.2, the Thunderbolt controller writes data arriving from the other Mac directly into the application’s memory, without going through the operating system’s network stack (Apple TN3205). Latency drops below 50 microseconds (Gigazine); MacGeneration speaks of a few tens of microseconds (MacGeneration). Total waiting per token falls to at most about 6 milliseconds.

Independent measurements bear this out. In Jeff Geerling’s test with four Mac Studios, llama.cpp, which splits by layers over TCP, got slower as machines were added: Qwen3 235B ran at 20.4 tokens per second on one machine and 15.2 on four. By contrast, exo, which splits within layers and uses RDMA, sped up from 19.5 to 31.9 (AppleInsider).

Apple measured with an 8,192-token prompt and a single user (Apple).

Adding machines does not scale linearly. For Muse Spark Open, going from two Macs to four raises speed from 25.0 to 36.7 tokens per second: about 1.47×, not 2×.

Dense models gain more. Qwen3.8 27B, which uses all of its parameters for every token, produces 19.7 tokens per second on one M5 Ultra, 33.3 on two (1.69×), and 50.9 on four (2.58×). Apple’s claim of “up to 3x faster performance” than a single system (Apple) also holds for a dense model: Mistral Medium 3.5 goes from 7.7 to 23.4 tokens per second on four machines.

First token: With an 8,192-token prompt, Muse Spark Open’s first token arrives in about 8 seconds on two Macs and 4.8 seconds on four.

The 1-million-token context was not measured. The model supports 1 million tokens, but Apple’s measurements stop at 65,536. At that length, even four Macs take about 40.6 seconds to produce the first token, or about 50.5 seconds with one cable per link. One million tokens is 16 times longer; even an optimistic linear estimate puts the first token more than 10 minutes away.

What it means for the Macs on your desk

The first question many readers will have: if I buy a second Mac, can I run a model this big, too?

What Apple and Meta showed is real and impressive: an 816-billion-parameter model producing 25 tokens per second on two desktop computers. Three things make it possible: the model’s MoE design, a smart way of splitting it, and RDMA’s low latency.

“One supercomputer with a terabyte of unified memory,” by contrast, is marketing shorthand that squeezes this engineering into a single sentence. There is a terabyte of memory; it just isn’t one memory.

This article was written with AI assistance; all figures come from the linked sources.

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →