DGX Spark Deep Dive on Brandon's Technical Blog https://brandongeraci.com/projects/dgx-spark/ Recent content in DGX Spark Deep Dive on Brandon's Technical Blog Hugo -- 0.146.0 en-us Sat, 08 Nov 2025 15:00:00 +0000 KV Cache Deep Dive: The 2x Reduction Mystery https://brandongeraci.com/projects/dgx-spark/06-kv-cache-deep-dive/ Sat, 08 Nov 2025 15:00:00 +0000 https://brandongeraci.com/projects/dgx-spark/06-kv-cache-deep-dive/ Why do Docker containers allocate roughly half the KV cache of native execution? And why do all models converge to ~44 GiB in native mode? Let's dig deep into the memory allocation mechanics. What I Learned (And What's Next) https://brandongeraci.com/projects/dgx-spark/05-what-we-learned/ Sat, 08 Nov 2025 14:00:00 +0000 https://brandongeraci.com/projects/dgx-spark/05-what-we-learned/ After 60 benchmarks and hours of debugging, here's what I learned about Grace Blackwell, Docker, and the importance of understanding your full stack. Plus: what I'm investigating next. The Data: 60 Runs Don't Lie https://brandongeraci.com/projects/dgx-spark/04-the-data/ Sat, 08 Nov 2025 13:00:00 +0000 https://brandongeraci.com/projects/dgx-spark/04-the-data/ I ran 60 comprehensive benchmarks across three different model sizes. The results are consistent, surprising, and prove that unified memory + Docker cgroups = significant overhead. The Unified Memory Revelation: Why Docker Double-Counts https://brandongeraci.com/projects/dgx-spark/03-unified-memory-revelation/ Sat, 08 Nov 2025 12:00:00 +0000 https://brandongeraci.com/projects/dgx-spark/03-unified-memory-revelation/ The 'aha!' moment: Docker's cgroups were designed for discrete GPUs, but Grace Blackwell has unified memory. The result? Docker counts GPU memory twice, creating 20-30GB of phantom overhead. Down the Rabbit Hole: The MPI and Chroot Nightmare https://brandongeraci.com/projects/dgx-spark/02-mpi-chroot-nightmare/ Sat, 08 Nov 2025 11:00:00 +0000 https://brandongeraci.com/projects/dgx-spark/02-mpi-chroot-nightmare/ Sometimes the hardest part of debugging isn't finding the bug - it's just getting to a clean test environment. Here's how we spent hours fighting MPI libraries and chroot configurations before we could even start benchmarking. The Mystery: Don't Just Blame the Hardware https://brandongeraci.com/projects/dgx-spark/01-the-mystery/ Sat, 08 Nov 2025 10:00:00 +0000 https://brandongeraci.com/projects/dgx-spark/01-the-mystery/ YouTube tech reviewers were quick to blame NVIDIA's new DGX Spark for being slow. But was it really the hardware's fault, or was I missing something deeper in the software stack?