Chip guide · with measured numbers
GB10 is NVIDIA's Grace-Blackwell superchip: an ARM64 Grace CPU and a Blackwell GPU sharing one 128 GB memory pool. We run it in production — every number on this page was measured on our own nodes.
The chip
| Design | Grace-Blackwell superchip — Grace CPU + Blackwell GPU on one package |
|---|---|
| Memory | 128 GB unified, shared CPU/GPU — one pool, no separate VRAM |
| CPU architecture | ARM64 (Grace) |
| Ships in | NVIDIA DGX Spark — the rentable product |
| At AxForge | 4 systems live in Sweden (eu-se-1), 3 more on the way — days |
Unified memory changes the sizing question. On a conventional GPU, the model must fit in VRAM. On GB10, weights, KV cache and the host share one 128 GB pool — so model weights up to roughly 100 GB can load on a single chip without sharding across devices, with room left for cache and the system. For the 7B–35B models we serve, that means the whole model plus a 65,536-token context fits with headroom.
Performance
| Workload | Result | Condition |
|---|---|---|
| Qwen3.8 27B decode | 15.9 tokens/s | single stream, multi-token-prediction speculative decoding on |
| Qwen3.8 27B decode (baseline) | 5.8 tokens/s | single stream, speculative decoding off |
| Qwen3.6 35B A3B decode | 30.7 tokens/s | single stream; MoE, 3B active parameters |
| ERNIE Image Turbo 1024×1024 | ~32 s / image | 8-step turbo fp8, current generation model |
| Qwen-Image 1024×1024 | ~26 s / image | 8-step lightning (now serving edits) |
Measured on our production DGX Spark node, single-stream, 2026-08. Serving config for Qwen3.8 27B: 65,536 context, up to 4 concurrent sequences. No third-party benchmark claims — for model quality, see each model's official card.
Reading the numbers
Speculative decoding matters
The same dense 27B model runs at 5.8 tokens/s plain and 15.9 tokens/s with multi-token-prediction speculative decoding. On GB10, serving-stack choices move the result as much as the silicon does.
MoE runs fast
Qwen3.6 35B A3B — 35B total, 3B active — decodes at 30.7 tokens/s, roughly twice the dense 27B. Sparse models are a natural fit for a chip whose memory pool is large relative to its compute.
Data & privacy
Our GB10 systems run in Sweden. On a dedicated machine: your model, your traffic, our hardware — prompts never persisted. Only request metadata (token counts, timestamps, status) is kept for billing and operations. Full policy at axforge.ai/privacy.
FAQ
GB10 is NVIDIA's Grace-Blackwell superchip: a Grace ARM64 CPU and a Blackwell GPU sharing 128 GB of unified memory in one pool. It powers the NVIDIA DGX Spark.
128 GB of unified memory, shared between CPU and GPU. There is no separate VRAM pool — weights, KV cache and the host share the same 128 GB.
On our production DGX Spark nodes (single-stream, 2026-08) we measured Qwen3.8 27B at 15.9 tokens/s with speculative decoding, Qwen3.6 35B A3B at 30.7 tokens/s, Qwen-Image at about 26 seconds per 1024×1024 image, and ERNIE Image Turbo (the current generation model) at about 32 seconds per 1024×1024 image.
Yes — as a dedicated NVIDIA DGX Spark from AxForge. 8 units are available now in Sweden on monthly rental, with 3 more arriving in days. €495/month launch pricing.
The 128 GB unified pool means model weights up to roughly 100 GB can load on a single GB10 without sharding across devices, leaving room for KV cache and the system. Our measured production workloads are 7B–35B class models.
Yes. The Grace CPU is ARM64, so software must be built for ARM64. Our own serving stack runs on it in production — every number we publish was measured there.