docs: GB10 (Blackwell sm_121) optimization guide

[CyberBrown]

Summary

• Adds a hardware-specific optimization guide for the NVIDIA GB10 (DGX Spark), the first Blackwell desktop GPU
• Covers the unified memory architecture (128 GB LPDDR5X shared with Grace CPU), which is fundamentally different from HBM-based data-center GPUs
• Includes measured benchmark numbers from real hardware, not theoretical estimates

What's in the guide

Section	Key content
Architecture overview	48 SMs, 1536 threads/SM, 100 KB shared/SM, 24 MB L2
Comparison table	Side-by-side with H100 and A100
Measured performance	91 TFLOPS BF16 tensor, 218 GB/s bandwidth, RMSNorm benchmarks
Memory hierarchy	Unified memory behavior, vectorization patterns, L2 sizing
Occupancy tuning	48 warps/SM math, block size recommendations
Compilation	sm_121 flags, TORCH_CUDA_ARCH_LIST, flash-attn warning
Best practices	8 actionable rules specific to GB10's bandwidth profile

Benchmark highlights (GB10, CUDA 13, PyTorch 2.10, BF16)

• BF16 MatMul: 92.3 TFLOPS (tensor core)
• Memory bandwidth: 218 GB/s measured (40% of ~546 GB/s theoretical)
• Vectorized RMSNorm: 2.59× average speedup over PyTorch baseline
• Key insight: 256-thread blocks give 100% occupancy; 1024-thread blocks drop to 67%

Motivation

The GB10/DGX Spark is shipping now and kernel authors need architecture-specific guidance. The unified memory model means many assumptions from HBM GPU guides (pinned memory, PCIe transfers, multi-TB/s bandwidth budgets) don't apply. This guide fills that gap with tested, practical advice.

A companion kernel (logos-flux/gb10-rmsnorm) is published on the Hub as the first sm_121 kernel.

Test plan

• Guide renders correctly in markdown
• Added to _toctree.yml under "Building kernels" section
• All benchmark numbers measured on real GB10 hardware
• Code snippets tested and verified

🤖 Generated with Claude Code

[danieldk]

cc @burtenshaw @sayakpaul