Why memory bandwidth decides which local AI computer deserves your money

A $2,999 machine that ships at $3,999 and then jumps to $4,699 within a few months isn't a pricing strategy — it's a moving target. That's exactly what happened with NVIDIA's DGX Spark, and it's a big part of why the "supercomputer on your desk" conversation has gotten so heated among people actually trying to run local AI models instead of paying for another subscription.

If you're weighing a DGX Spark against a Mac Studio, a Mac mini, or an AMD Strix Halo mini-PC, the spec sheet alone won't tell you what you need to know. You need to look at one number specifically — memory bandwidth — and see how it plays out on a real benchmark. Here's how to actually make sense of it.

What is the NVIDIA DGX Spark, and why did the price change?

The DGX Spark is built around NVIDIA's GB10 Grace Blackwell Superchip — a combination of a 20-core Arm CPU and a Blackwell GPU connected through NVLink-C2C, which lets the CPU and GPU share one pool of memory instead of copying data back and forth. It combines a Blackwell-generation GPU with fifth-generation Tensor Cores and FP4 support with a 20-core Arm CPU connected via NVLink-C2C, and that interconnect is why unified memory works — the CPU and GPU share the same physical memory pool with coherent access.

On paper, that's genuinely impressive. It can deliver up to 1 petaflop of AI performance and 128 GB of unified CPU-GPU memory in a compact desktop form factor. That setup offers developers the power to run inference on AI models with up to 200 billion parameters and fine-tune models of up to 70 billion parameters locally. Need more headroom? Two DGX Spark units can be linked via ConnectX-7 to share 256GB and run models up to 405 billion parameters.

You can read NVIDIA's own announcement on the official DGX Spark launch page for the full technical rundown straight from the source.

Now the part that annoys buyers. Originally announced with a $2,999 starting price, Spark launched in late 2025 at $3,999, and in February 2026 NVIDIA raised the price to $4,699, citing global memory supply constraints. That's a 17.5% jump on top of an already-higher launch price, and it happened after the hardware was already in people's hands — not before.

How does memory bandwidth actually determine local AI performance?

Here's the thing nobody puts on the marketing slide: raw compute (measured in FLOPs) matters a lot less for token generation than memory bandwidth does. When a model generates text one token at a time, the bottleneck usually isn't how fast the chip can multiply matrices — it's how fast it can pull the model's weights out of memory for each token.

The DGX Spark's unified memory runs at 273GB/s. While memory bandwidth at 273 GB/s creates some trade-offs compared to discrete high-end GPUs, the unified memory architecture eliminates VRAM constraints that plague traditional setups. That's the trade NVIDIA is making: more usable memory capacity, at the cost of bandwidth compared to dedicated GPU VRAM.

Compare that to Apple's high-end silicon. The M3 Ultra Mac Studio starts with 96GB of unified memory with an 819GB/s bandwidth, while the Mac Studio with M4 Max starts with 36GB of unified memory with up to 546GB/s bandwidth, with a maximum of 128GB of memory. Both of those numbers dwarf the Spark's 273GB/s. That gap is the single biggest reason the Spark doesn't run away with every benchmark despite its petaflop rating.

How does the DGX Spark compare to Mac Studio and AMD Strix Halo in practice?

This is where the specs stop mattering and the actual test runs start to matter. GPT-OSS 120B is a useful benchmark model because it's large enough to stress unified-memory systems and small enough to actually fit and run on all of these boxes.

On llama.cpp, independent testing found the gpt-oss-120b model hitting PP (prefill/prompt-processing) of 1,821 tokens per second on the DGX Spark, which lines up closely with other reported runs. Recent benchmarks show the GPT-OSS 120B model achieving 1,723 tokens per second for prompt processing on the same setup. Prefill — the speed at which the model reads and processes your prompt — is genuinely strong on this hardware, largely because that stage leans on raw compute rather than memory bandwidth.

Decode speed (how fast it generates each new token) tells a very different story. Multiple community benchmarks land the DGX Spark's GPT-OSS 120B decode rate somewhere in the mid-30s to mid-40s tokens per second, depending on the inference engine. On the llama.cpp GitHub discussion thread where developers have been comparing notes, one tester reported Test 1000 tok Long (1000 tok) 1000 19.1s 52 tok/s using SGLang, versus 46 tok/s on llama.cpp for the same model — a real difference depending on which inference stack you run. You can follow that whole benchmarking thread yourself on llama.cpp's DGX Spark performance discussion, which is worth bookmarking if you're serious about squeezing more speed out of this hardware.

And here's the uncomfortable comparison for Spark owners: at around 38 tokens per second on the 120B model, it is only marginally faster than the AMD Strix Halo APU — a mini-PC platform that costs roughly half as much. That's the number that decides the whole argument. If decode speed on a 120B model is nearly tied between a $4,699 box and a $2,000-ish Strix Halo machine, the DGX Spark's premium has to be justified by something other than raw chat speed.

What about the software ecosystem?

This is the DGX Spark's real advantage, and it shouldn't get lost in the bandwidth argument. The DGX Spark stands out for its 128 GB unified memory, the largest among desktop GPUs, and its FP4 acceleration. It's also a full CUDA machine, which means you get access to NVIDIA's entire software stack — TensorRT-LLM, NIM microservices, and the broader CUDA ecosystem — none of which run natively on Apple Silicon or AMD's platform. And NVIDIA has kept improving it post-launch: TensorRT-LLM optimizations and speculative decoding have improved performance, with NVIDIA reporting up to a 2.5x throughput increase across a range of benchmarks compared with launch performance. If you're doing CUDA-dependent fine-tuning work, that ecosystem access is worth real money. If you just want a fast chat box, it isn't.

Does running multiple requests at once change the math?

Single-user chat throughput is only half the story if you're planning to serve a small team instead of chatting solo. Batching multiple simultaneous requests together lets a GPU spend more of its cycles on compute instead of waiting around for memory to catch up — which is exactly where the DGX Spark's petaflop-class FP4 performance starts to pull its weight back. If your use case is a handful of people hitting the same local model through an API rather than one person typing in a chat window, the concurrent-throughput picture looks meaningfully better for the Spark than the single-stream numbers suggest. That's the strongest rebuttal owners have, and it's worth testing with your own workload before writing the machine off entirely.

Is the price increase actually justified?

Community sentiment has shifted along with the price. On developer forums discussing the hardware, reactions ranged from praising the CUDA access to fl