Xiaomi AI Cube vs DGX Spark: prototype versus product

Xiaomi demonstrated a compact local-AI computer with three in-house processors, a claimed 1.22TB/s near-memory path, and local 120B plus 3B model deployment. It looks like a DGX Spark competitor, but the products are not at the same stage.

DGX Spark is available and documented. AI Cube is an engineering prototype. Xiaomi has not announced the Cube’s price, release date, installed memory, storage, ports, operating system, public runtime, or matched benchmark.

Buyer verdict: buy the platform you can test

Do not defer a current local-AI project for AI Cube. Buy DGX Spark only when its 128GB unified memory, CUDA stack, and documented deployment path match a named workload you need now. Choose a configured Mac or conventional GPU workstation when your preferred runtime, budget, or existing software points there. Watch AI Cube if its near-memory design interests you, but do not count an announced prototype as capacity, support, or a delivery date. Xiaomi’s public demonstration is useful architecture evidence; it is not a retail specification or a comparable performance result.

Verified specification versus unknown

Specification	Xiaomi AI Cube prototype	NVIDIA DGX Spark
Status	Verified: demonstrated prototype; no retail date or price announced	Verified: shipping developer system with support and documentation
Architecture	Verified: XRING O3 + O100 + D100	Verified: GB10 Grace Blackwell Superchip
System memory	Unknown: installed AI Cube memory and how chips share it. D100 itself supports up to 160GB.	Verified: 128GB coherent LPDDR5X
Published bandwidth	Verified but scoped: O100 claims 1.22TB/s near-memory; this is not a published system-memory figure	Verified: 273GB/s coherent system memory
Local-model claim	Verified as a demo: 120B + 3B dual-model deployment; exact model, precision, and residency are unknown	Verified capacity claim: models up to 200B, subject to format and runtime
Published token result	Verified as Xiaomi lab result: up to 330 tokens/s on MiMo 3B; context, batch, runtime, and wall power are unknown	Unknown as a universal figure: rate depends on model and runtime
Software and I/O	Unknown: OS, runtime, containers, storage, ports, networking, and public SDK	Verified: DGX OS, CUDA ecosystem, 4TB NVMe, 10GbE, Wi-Fi 7, and ConnectX-7
Price	Unknown: not announced	Published NVIDIA MSRP: $4,699; local reseller price and availability can differ

Component maximums are not system benchmarks. Xiaomi’s 1.22TB/s figure and NVIDIA’s 273GB/s figure describe different memory scopes.

Why Xiaomi uses three chips

• XRING O3 supplies general compute, graphics, and a low-power NPU.
• XRING O100 is the specialist inference accelerator with stacked DRAM and a claimed 1.22TB/s local interface.
• XRING D100 is an intelligent-driving processor with a 20-core CPU, 16-core NPU, and support for up to 160GB unified memory.

The missing detail is how those chips share work and memory. Xiaomi has not published the inter-chip fabric, coherency model, copy cost, or the chip used for each stage of the 120B workload. Three accelerators help only if data movement does not erase their individual advantages.

What O100 may solve

Large-model token generation often becomes limited by memory movement. O100 places two high-speed DRAM wafers above a 6nm logic layer with 14 NPU cores. Xiaomi reports 28,672 effective data connections and up to 1.22TB/s of local bandwidth.

That design could improve decode speed when the working model fits in the stacked memory. Xiaomi has not disclosed how much DRAM O100 contains. A small fast tier can make a 3B model quick while a 120B model still depends on a slower capacity tier.

What 330 tokens/s does not prove

Xiaomi’s 330 tokens/s result used its MiMo 3B model. The company did not publish the input length, output length, batch size, exact format, runtime version, time to first token, or wall power. It was not the 120B model.

Dense model example	Approximate raw weights at 4-bit	Ideal full-weight scans per second at 1.22TB/s*
3B	1.5GB	813
30B	15GB	81
70B	35GB	35
120B	60GB	20

*This is a bandwidth illustration, not a speed forecast. It excludes caches, mixture-of-experts routing, compression, compute limits, context traffic, and runtime overhead.

The table shows why a 3B result says little about a 120B workload. A defensible comparison requires the same public model, quantization, context, batch size, output length, and quality settings on both machines.

Why the 120B plus 3B design is useful

A small local model can classify a request, retrieve a file, or decide whether the task is simple. A larger model can handle deeper reasoning and difficult generation. Routing can reduce latency and avoid using the large model for every step.

Xiaomi has not shown whether both models stay resident, how memory is divided, whether context survives a handoff, or whether third-party models can use the same route. The concept is useful; the developer platform is not yet public.

Where DGX Spark is ahead

DGX Spark provides one coherent 128GB memory pool and a documented software path. Developers can use CUDA, PyTorch, NVIDIA containers, TensorRT-LLM, vLLM, llama.cpp, Ollama, and LM Studio. NVIDIA also publishes deployment and fine-tuning guides.

Its limitations are known. The 273GB/s memory interface can constrain single-user decode on large dense models. The Arm64 host can complicate x86-only software. The $4,699 price is high, and the 1 PFLOP headline assumes sparse FP4 work.

Those limits can be measured because the product exists.

Buying decision

Buy DGX Spark if you need 128GB of local model capacity now, rely on CUDA or NVIDIA containers, or need a documented deployment path.

Watch AI Cube if near-memory inference, local model routing, or Xiaomi’s edge ecosystem matters to you. There is nothing to buy yet.

Keep a conventional GPU workstation if your models fit in GPU memory and image, video, training, or x86 compatibility matter more than compact unified memory.

Before spending, compare the workload rather than the enclosure: the Mac mini configuration guide explains local-memory headroom, the Mac Studio guide covers higher-capacity Apple options, and the model comparer keeps model-cost and benchmark evidence separate from hardware headlines.

Which machine fits which local-AI user?

Your requirement	Best current fit	Why	Do not overread
A 128GB local model host with CUDA containers now	DGX Spark	Retail hardware, coherent memory, DGX OS, CUDA, and published playbooks exist	A capacity claim is not a guarantee that every 200B artifact is fast or compatible
A 7B–35B assistant on a quiet general-purpose desktop	Configured Mac mini or Mac Studio	Unified memory can fit selected local artifacts while retaining macOS workflows	Memory headroom, context cache, and runtime support still decide fit
Image, video, training, or x86-first software	Conventional GPU workstation	Mature discrete-GPU and x86 software paths may matter more than compact unified memory	Price, VRAM, power, and noise can outweigh a headline accelerator spec
Researching near-memory inference or Xiaomi’s edge stack	AI Cube watchlist	O100 is an interesting disclosed memory architecture	There is no public buying, runtime, or benchmark path yet

A reproducible local-AI benchmark plan

Run this plan only when AI Cube hardware and a public runtime exist. It deliberately publishes no invented result.

1. Use one public checkpoint and immutable revision on both systems. Record the quantization, tokenizer, context length, prompt tokens, generated tokens, sampling settings, batch size, concurrency, and stop condition.
2. Pin the runtime, compiler/driver, container image, command line, and model file hash. Run a warm-up, then at least three timed runs per workload.
3. Measure prefill throughput, time to first token, decode tokens/s, and aggregate tokens/s separately. A short 3B decode demo cannot stand in for a long-context 120B workload.
4. Capture peak host and accelerator memory, model residency, cache behavior, and any inter-chip transfer or offload. For AI Cube, disclose installed O100/D100 memory and the data path used by every stage.
5. Measure wall power with an external meter at idle, during prefill, during decode, and after 30 minutes of sustained load. Record ambient temperature, thermal mode, and throttling.
6. Compare task quality with fixed public prompts and a blinded human pass/fail rubric, not just token rate. Publish raw logs and failures.
7. Repeat the exact harness on DGX Spark using a public optimized runtime, then publish the hardware configuration, total cost, and the full result table.

The closest current reference implementation is NVIDIA’s DGX Spark vLLM playbook; adapt its version-pinned command into a public repository rather than presenting a screenshot or a single console number as a benchmark.

Until those results exist, AI Cube is an interesting architecture, not a proven DGX Spark replacement.

Sources

• IT Home: AI Cube specifications and three-chip configuration
• The Paper: XRING announcement and commercial timeline
• 36Kr: O100 architecture and stacked-memory details
• CnEVPost: D100 and the AI Cube demonstration
• NVIDIA: DGX Spark specifications
• NVIDIA: DGX Spark system overview
• NVIDIA: DGX Spark playbooks
• Tom’s Hardware: independent DGX Spark review

Put this to work

Separate memory capacity, memory bandwidth, precision, model size, prefill, and decode when reading local-AI claims.

Try

Write a matched benchmark using the same model, quantization, context, batch size, runtime, and wall-power measurement.

Prove it worked

Publish time to first token, decode rate, throughput, memory use, power, and task quality from both machines.

Where it can pay

A local-inference setup is valuable when it completes a private workflow reliably, not when it wins a component-spec comparison.

Keep in view

• Xiaomi's 150W AI Cube prototype combines XRING O3, O100, and D100. It is not a retail product.
• The claimed 1.22TB/s figure belongs to O100's near-memory interface; Xiaomi has not disclosed its fast-memory capacity.
• The 330 tokens/s demo used MiMo 3B, not the unnamed 120B model also shown on the Cube.
• DGX Spark provides 128GB coherent memory, a documented CUDA stack, and a current $4,699 MSRP.

Learn the workflow: choosing and self-hosting Chinese open-weight models →