AI Workstations (GPU+CPU Combos)

Some recent chips pair a CPU and GPU on one package with unified memory shared between them. Instead of a discrete GPU with its own dedicated VRAM, these systems let the GPU directly address tens or hundreds of gigabytes of system memory, enough to fit very large language models on a desktop or mini-PC. The tradeoffs aren't obvious from a spec sheet.

I'm calling this category AI Workstations (previously titled "GPU+CPU Combos" when this article was first published in May 2026). To be precise about terms: an AI Workstation (what NVIDIA calls an "AI Supercomputer") is a full working computer optimized for AI. You unbox it, plug it in, and run models; no additional components needed. That is fundamentally different from an SoC (what NVIDIA calls a "Superchip" and AMD ships as Ryzen AI Max and similar parts), which is a single chip combining CPU and GPU with unified memory shared between them, but still needs a full computer built around it (board, storage, power, chassis) before it runs anything. The shared idea across both: one memory pool directly addressable by the GPU, holding tens or hundreds of gigabytes, enough to fit very large language models on a desk instead of in a datacenter.

Prices below show both the original May 2026 figure and the latest September 2026 figure, so history doesn't get silently rewritten: memory shortages and demand moved several of these substantially.

Updates

September 2026: renamed the category to AI Workstations. Previously "GPU+CPU Combos." Same table, clearer name: these systems compete against a full multi-GPU build, not against one card.

September 2026: Apple introduces new Mac Studio with M5 Max and M5 Ultra. M5 Max (18-core CPU, up to 40-core GPU) keeps the 128 GB ceiling but lifts bandwidth to 614 GB/s; M5 Ultra (up to 36-core CPU, 80-core GPU) keeps 512 GB and jumps to 1.2 TB/s. Both add per-GPU-core Neural Accelerators (up to 4.3x peak AI compute vs M3 Ultra, per Apple). No independent benchmarks yet. I treat Apple's 9.8x LM-Studio prompt-processing claim as a vendor figure until third parties measure tokens/sec.

September 2026: added RTX Spark N1X and shipping DGX Station GB300. NVIDIA's Arm chip for Windows laptops and mini PCs arrives October 2026 in two configs: full 20-core CPU + 6144-core Blackwell GPU with up to 128 GB, and laptop-only 18-core + 5120-core with 24-32 GB (The Verge). NVIDIA markets these for gaming too (100 FPS @1440p "with RT & DLSS"), so "AI only" describes where they're competitive, not what the silicon can do. The deskside DGX Station tower (Grace 72-core + Blackwell Ultra, ConnectX-8) now ships at around $85,000 via system integrators. Firmly enterprise, but the ceiling this category is heading toward.

September 2026: prices now carry dates. DRAM shortages pushed Apple to trim M4 Max configs (96 GB max) and hike prices mid-cycle, so the table shows May 2026 alongside September 2026 instead of silently overwriting history.

May 2026: original publication. First version of this guide, then titled "GPU+CPU Combos," covering DGX Spark, Strix Halo, Mac Studio M4 Max/M3 Ultra, and the enterprise GH200/GB200/MI300A parts.

The Bandwidth vs Capacity Tradeoff

The defining design choice is still which memory technology to use:

• LPDDR5X is cheap (~$3-5/GB), dense, and low-power. You can solder 128 GB next to an SoC for under $600 in materials. But on a 256-bit bus it tops out around 273 GB/s.
• HBM3/HBM3e is fast (3-8 TB/s) but expensive: contract pricing sits around $11-14/GB in 2026 after the ~20% hike suppliers pushed through at renewal (versus ~$8-10/GB in 2025), with spot product reportedly changing hands far higher and that's before the advanced 2.5D packaging needed to attach it to a GPU. 128 GB of HBM dies at contract rates runs ~$1,500+ before assembly, up from ~$1,000+ in 2025.

That's why a $3,999 DGX Spark uses LPDDR5X and a $15,000+ AMD MI300A uses HBM. There's no $3,000 box with HBM bandwidth: advanced packaging (TSMC's CoWoS) and HBM supply were the primary bottleneck for AI chip production in 2025, and that capacity is allocated to datacenter parts.

Apple is now blurring the line: the M5 Ultra's 1.2 TB/s at 512 GB is the first deskside unified-memory system that doesn't force the capacity-vs-bandwidth choice, at a $5,499+ entry price.

This matters for AI workloads: large language model decoding (generating one token at a time) is memory-bandwidth-bound, while prefill (processing your prompt) is compute-bound (a well-documented split that production inference systems increasingly handle on separate GPU pools). Tokens per second during decode scales almost linearly with memory bandwidth; prefill scales with raw tensor core throughput.

A real example on a 120-billion-parameter model: DGX Spark generates ~39 tokens/sec, while a 3× RTX 3090 setup (with ~3.4× the aggregate bandwidth) does ~124 tokens/sec, a near-perfect linear scaling. So if "wait time for the AI to respond" is what you care about, bandwidth is the number that matters most, and a maxed-out Mac Studio or a stack of discrete GPUs will outperform a Spark or Strix Halo at the same model size.

The Ecosystem Story

Hardware is only half the picture. The software stack you're locked into shapes what's possible:

• CUDA (NVIDIA): every ML framework supports it natively. Lowest porting friction. Spark, N1X, Jetson, GH200, GB200 all share this stack.
• ROCm (AMD): improving rapidly but still has rough edges for niche operators. Strix Halo and MI300A run here.
• MLX / Metal (Apple): excellent for Apple-native workflows (MLX, ollama on Mac, Core ML, and the new Core AI framework), but most published research code targets CUDA and needs porting. Worth watching: oMLX (open-source, Apache 2.0), an MLX inference server for Apple Silicon with continuous batching and tiered KV caching (RAM plus SSD) that claims big latency wins for agentic coding workloads. It's still new, and its benchmarks are project-published, but one to watch on the Mac side.

For many buyers, the ecosystem question dominates the spec sheet. A Mac Studio with 3× the bandwidth of a Spark doesn't help if the model you want to run only ships CUDA kernels.

Where Each One Fits

• Spark / Strix Halo / RTX Spark N1X ($2-4K): fit big models locally, accept slower generation, value low power and small form factor. Spark wins on software, Strix Halo on price and x86 compatibility, N1X on Windows-native portability.
• Mac Studio M3 Ultra / M5 Ultra ($4K+): best bandwidth for local LLM decode if you can live with MLX/Metal. M5 Ultra is the new ceiling: 512 GB at 1.2 TB/s.
• Mac Studio M4 Max / M5 Max ($2-3K): the sensible middle: 128 GB is enough for 70B-class models, and 614 GB/s (M5) meaningfully improves decode over the Spark/Halo tier.
• Discrete GPUs (RTX 5090, RTX Pro 6000 Blackwell): 6× the bandwidth and much higher FLOPS, but capacity is capped (32-96 GB) and you need a host PC. Better for models that fit, worse for ones that don't.
• DGX Station / MI300A / GH200 / GB200: enterprise-only. If your budget starts with five figures per chip, you're not shopping at this end of the market.

Should These Be on GPU Poet?

Yes, but not in the GPU tables. Comparing a single graphics card against a full AI workstation in the same ranking would mislead more buyers than it would help: system price vs card price, unified memory vs VRAM, and missing gaming benchmarks make every shared column dishonest. A workstation competes against a machine (multi-GPU build plus CPU), not a card.

So my plan is a standalone AI Workstation section with its own workstation-vs-workstation rankings, shop pages, and comparisons. Let me know which comparisons you'd want to see first (Spark vs Strix Halo vs Mac Studio is the obvious starting grid), and I'll prioritize accordingly.