The $9,500 Question

The DGX Spark's MSRP just jumped $700 — from $3,999 to $4,699 — and NVIDIA is blaming LPDDR5x supply constraints. That's an 18% price hike that fundamentally reshapes the math for anyone weighing these two platforms.

Two DGX Sparks now run $9,398. A Mac Studio M3 Ultra with 512GB unified memory starts at $9,499. We're talking a $101 difference for dramatically different hardware: the Mac gives you 4x the memory (512GB vs. 256GB combined), 3x the bandwidth per node (819 GB/s vs. 273 GB/s), and silence. The Sparks give you roughly 77x the AI compute (2 PFLOPS sparse FP4 vs. 26 TFLOPS FP16) and the entire CUDA ecosystem.

The price parity is almost poetic. NVIDIA positioned the Spark as a "personal AI supercomputer on your desk," but at $4,699, that desk needs deep pockets. For developers whose primary workflow is running coding models — not training them — the value proposition has shifted decisively toward Cupertino.

EXO Labs did something clever: they wired two DGX Sparks to a Mac Studio M3 Ultra and let each system do what it does best. The Spark handles the prefill phase (processing your prompt — compute-heavy), and the Mac handles the decode phase (generating tokens — bandwidth-heavy). The result? A 2.8x speedup over the Mac alone, and 1.9x faster than the Spark alone.

This isn't theoretical. Running Llama 3.1 8B at FP16 with an 8,192-token prompt, the hybrid cluster matched the Spark's blistering prefill speed while keeping the Mac's quick token generation. The trick is EXO's disaggregated inference framework, which streams the KV cache layer-by-layer between machines so both systems work simultaneously instead of waiting on each other.

The catch: EXO 1.0 with automated scheduling and KV streaming isn't publicly released yet (the alpha is at 0.0.15). But the architecture proves something important — the "Mac vs. Spark" question might be a false binary. The real power move could be the $19,000 heterogeneous cluster that demolishes either platform running solo.

When John Carmack says your hardware underperforms, people listen. The legendary programmer reported that his DGX Spark maxed out at just 100 watts — less than half its 240W rating — delivering roughly half the quoted performance. The system ran hot, crashed during extended workloads, and was rebooting spontaneously under sustained AI loads.

The NVIDIA developer forums quickly filled with corroborating reports. Users documented GPU crashes, thermal shutdowns, and the realization that the Spark's beautiful 1.13-liter chassis might be its own worst enemy. The GB10 Grace Blackwell superchip generates serious heat, and the compact enclosure simply can't dissipate it fast enough under sustained load.

Compare this to the Mac Studio, which runs at roughly 70W under full LLM inference and produces zero audible noise. Apple's thermal engineering, born from years of fanless laptop design, gives the Mac Studio a decisive quality-of-life advantage. You can run 70B models all day in a quiet home office. The Spark? Some users report keeping "a small USB fan pointed at mine."

AMD seized the moment, offering Carmack a Strix Halo box as an alternative. The episode raised pointed questions about NVIDIA's plans to rebrand the GB10 as the "N1" APU for laptops — if it throttles in a desktop box, what happens in a slim chassis?

The LMSYS in-depth review of the DGX Spark remains the most rigorous benchmark available, and the numbers reveal a fascinating split personality. On Llama 3.1 8B (FP8), the Spark hits 7,991 tokens/sec on prefill but only 20.5 tokens/sec on decode at batch 1. Crank up concurrency to batch 32 and decode jumps to 368 tokens/sec — but that's a server workload, not a coding assistant on your desk.

The llama.cpp community benchmarks tell an even more interesting story. Mixture-of-Experts models are the killer app for the Spark: Qwen3 30B MoE hits approximately 89 tokens/sec, while the similarly-sized Qwen3 32B Dense slams into a "bandwidth wall" at just 10.7 tokens/sec. The MoE architecture means only a fraction of the model's parameters are active per token, dramatically reducing the bandwidth requirement.

For the Llama 3.1 70B at FP8 — the kind of model a serious coding assistant would run — the Spark manages 803 tokens/sec prefill but only 2.7 tokens/sec on decode. That's painful for interactive use. On the Mac Studio, the 70B model benefits from 819 GB/s bandwidth to deliver noticeably faster token-by-token generation, though it needs 4-bit quantization (GGUF via llama.cpp) to fit in memory comfortably.

Here's the insight most benchmarks miss: coding AI workloads are decode-dominated. You paste a file into your prompt once (prefill), then the model generates hundreds of tokens of code, explanation, or refactoring suggestions (decode). A 2x prefill advantage matters far less than a 2-3x decode advantage when you're waiting for the model to finish writing a function.

Sebastian Raschka — the Machine Learning with PyTorch author — tested the DGX Spark against his Mac Mini M4 Pro for local PyTorch development, and his conclusion cuts to the heart of the ecosystem debate: "It is a CUDA device and thus much better supported in PyTorch."

For Ollama-based inference with optimized MXFP4 precision, both platforms hit roughly 45 tokens/sec on GPT-OSS 20B. Performance parity. But Raschka's real use case is coding LLMs from scratch in pure PyTorch, and here the Spark's CUDA support is decisive. MPS on macOS is still unstable for training workloads — fine-tuning jobs that converge perfectly on CUDA will sometimes fail to converge on MPS.

That said, Apple's MLX framework is evolving rapidly. MLX 2.0 added native agent support optimized for the M-series Neural Engine, and Apple is adding CUDA backend support to MLX — meaning you could eventually write and test locally on Apple Silicon, then deploy to NVIDIA hardware without rewriting. That future isn't here yet, but it's visible on the horizon.

When ServeTheHome called the 512GB Mac Studio "the local AI play," they weren't being hyperbolic. At 819 GB/s of memory bandwidth and half a terabyte of unified memory, the M3 Ultra Mac Studio can load DeepSeek-V3 — a 671-billion-parameter model — entirely in memory. The DGX Spark? Its 128GB ceiling means you're quantizing everything above 200B parameters into oblivion, or you're not running it at all.

The bandwidth advantage is structural, not incremental. LLM token generation speed is fundamentally limited by how fast you can read model weights from memory. The Mac's 819 GB/s means it reads the entire weight matrix roughly 3x faster than the Spark's 273 GB/s per clock cycle. For a 70B model, that translates directly into 3x faster token generation — the thing you actually feel while coding.

There's also the "it just works" factor. The Mac Studio is a general-purpose workstation. You run your IDE, your browser, your Slack, and a 70B coding model simultaneously — all sharing unified memory, all running silently. The DGX Spark is a dedicated AI appliance that runs Linux, needs its own monitor-keyboard-mouse setup (or SSH), and exists solely to serve models. For developers who want one machine that does everything, the Mac Studio is the only real option in this price range.