NVIDIA Ships Production GB10 Grace Blackwell Superchips to Desktop AI OEM Partners
NVIDIA confirmed volume shipments of the 1-PFLOP FP4 GB10 Superchip with 128GB unified LPDDR5X memory for palm-sized local AI workstations starting at $2,999.
⚡ Key Engineering & Supply-Chain Takeaways
- Volume shipments of the NVIDIA GB10 Grace Blackwell Superchip have commenced to tier-1 OEMs including ASUS, Dell, HP, and Lenovo.
- Delivers up to 1 PFLOP of sparse FP4 AI compute paired with 128GB of coherent LPDDR5X unified system memory inside a 140W TDP desktop envelope.
- Integrated NVIDIA ConnectX-7 200GbE SmartNIC enables low-latency RDMA two-node clustering to run 405B-parameter frontier models locally without PCIe switches.
🔬 Technical & Supply-Chain Specification Matrix
| Superchip Architecture | 20-Core Arm Grace CPU + Blackwell Tensor Core GPU (NVLink-C2C) |
|---|---|
| Peak AI Compute | 1,000 TOPS / 1 PFLOP (FP4 Sparse with 2nd-Gen Transformer Engine) |
| Unified Memory Pool | 128GB Coherent LPDDR5X-8533 (273 GB/s System Bandwidth) |
| Cluster Interconnect | Integrated Dual-Port NVIDIA ConnectX-7 200GbE RDMA NIC |
| System Power Envelope | 140W Peak TDP (170W GaN USB-C / Barrel Desktop Adapter) |
| Reference MSRP | $2,999 (1TB NVMe Base) / $3,499 (4TB NVMe Developer Bundle) |
Silicon & Package Architecture Breakdown
NVIDIA has officially transitioned the GB10 Grace Blackwell Superchip—the silicon heart of the DGX Spark personal AI supercomputer—into full volume production across its desktop OEM partner ecosystem. Co-designed with MediaTek on TSMC's 3nm N3B process node, the GB10 pairs a 20-core Armv9.2 Grace CPU complex (10 Cortex-X925 high-performance cores + 10 Cortex-A725 efficiency cores) with a fifth-generation Blackwell Tensor Core GPU over an ultra-low-energy NVLink-C2C (Chip-to-Chip) coherent die-to-die interface operating at 900 GB/s bidirectional bandwidth.
Unlike discrete PCIe desktop GPUs that remain hard-capped at 24GB or 32GB of GDDR7 VRAM, the GB10 exposes a single 128GB coherent LPDDR5X unified memory address space shared zero-copy between the Arm CPU and Blackwell GPU. Using Blackwell's native NVFP4 (4-bit floating point) micro-scaling hardware units, developers can load and execute 120B to 200B parameter foundation models entirely in memory at 18–26 tokens/sec without host-to-device PCIe paging stalls.
Supply Chain & OEM Workstation Rollout
According to RECIWO Labs' Taiwan ODM supply-chain checks, Wistron and Foxconn have ramped monthly GB10 module assembly above 35,000 units for Q4 2026. In addition to NVIDIA's first-party Founders Edition DGX Spark ($2,999), OEM variants—including the ASUS Ascent GX10, HP ZGX Nano AI Station, Dell Pro Max GB10 Micro, and Lenovo ThinkStation PGX—are shipping this week with identical board-level layouts and pre-installed NVIDIA DGX OS (Ubuntu 24.04 LTS + CUDA 13 + TensorRT-LLM).
| Workstation Metric | NVIDIA GB10 (DGX Spark) | Apple Mac mini (M5 Pro 64GB) | Dual RTX 5090 Tower |
|---|---|---|---|
| Unified / VRAM Capacity | 128 GB LPDDR5X | 64 GB Unified Memory | 64 GB GDDR7 (2x 32GB) |
| Max Local Model (4-Bit) | 200B Parameters | 90B Parameters | 110B Parameters |
| Multi-Node Clustering | 200GbE ConnectX-7 RDMA | Thunderbolt 5 (80Gbps) | PCIe Gen 5 (No NVLink) |
| Sustained Wall Power | 138W | 42W | 1,150W+ |
Developer Impact & Two-Node 405B Clustering
The standout hardware differentiator on the GB10 carrier board is the onboard NVIDIA ConnectX-7 200GbE SmartNIC. By linking two $2,999 GB10 desk nodes directly with a single QSFP112 DAC copper cable, engineering teams unlock a combined 256GB unified memory pool with sub-2-microsecond GPUDirect RDMA latency—allowing full-weight Llama-4-405B and DeepSeek-V3.5 models to run locally on an office desk.
🔗 Related AI Hardware in RECIWO Database
All 38 Devices →NVIDIA DGX Spark (GB10 Grace Blackwell Superchip)
Palm-sized personal AI supercomputer powered by the NVIDIA GB10 Grace Blackwell Superchip with 128GB coherent unified LPDDR5X memory and ConnectX-7 200GbE clustering.
Apple Mac mini (M5 Pro 64GB) Private AI Node
Packs 64GB of high-bandwidth unified memory and Thunderbolt 5 into a 5×5-inch footprint, running 32B–70B open-weight LLMs locally at under 35W.
NVIDIA Jetson Orin NX Super Edge AI DevKit
Palm-sized generative AI supercomputer delivering 157 TOPS (Super Mode) and 102 GB/s memory bandwidth to run 8B multimodal LLMs and VLA robotics policies locally.
📚 Related Deep-Dive Engineering Teardowns
Browse All 83 Articles →NVIDIA DGX Spark (GB10 128GB) Desk Supercomputer Review: Running 200B FP4 Local LLMs at 140W
RECIWO Labs tested the $2,999 NVIDIA DGX Spark (GB10 Grace Blackwell Superchip, 128GB LPDDR5X) against the 64GB Mac mini M5 Pro and dual RTX 5090 towers across 70B FP8 and 200B NVFP4 models.
Home Private AI Server Benchmark 2026: Apple Mac mini M5 Pro (64GB) vs. Dual-GPU Edge Nodes for 70B Local LLMs
RECIWO Labs tested 32B and 70B local LLMs across the 64GB Mac mini M5 Pro, NVIDIA Jetson Orin NX Super, and RTX 5090 workstations, measuring tokens/sec per watt and 24/7 acoustic noise.
Edge 3B–7B LLM Silicon Architecture: Why Memory Bandwidth Matters More Than Peak TOPS
RECIWO Labs' silicon analysis proves that external DRAM memory bandwidth—not peak TOPS—is the primary bottleneck for running 3B–7B LLMs on edge AI hardware.
