GPU Cluster Operations

Abstract After measuring a 292x cost gap between a rented B200 and frontier API providers on a batch inference workload, the logical next question was whether the same pattern held for operational intelligence: could a smaller, dedicated model handle the continuous judgment calls required to run a GPU fleet? The batch inference test had exposed KV cache contention as the dominant bottleneck on shared API infrastructure. Processing similar structured data at scale, but this time continuously rather than in batch, seemed like a valid test of whether that contention would degrade operational quality the same way it degraded throughput. ...

Three months ago, I plugged a Blackwell GPU into my lab bench and pointed it at a million PDFs. Corporate documents: contracts, engineering reports, compliance filings, vendor proposals, maintenance logs, insurance certificates. The kind of archive that accumulates over two decades of running critical infrastructure. A million files. Not sampled, not curated, not cleaned. Raw. The plan was straightforward: parse the documents, chunk them, embed them into a vector store, wire up a retrieval layer, and start asking questions no keyword search could answer. Retrieval-Augmented Generation. The acronym that launched a thousand vendor decks. ...

A few weeks ago we hit a production issue on a cloud environment — one XCP-ng host was showing IOPS contention caused by a single guest VM. The classic noisy-neighbor race condition on shared storage. The diagnostic path was obvious: cross the dom0 guest list with iostat on the host, find the VM hammering the disk, and work the problem from there. Straightforward correlation — the kind of thing an experienced operator resolves in fifteen minutes with two terminal windows. ...

Part 2 ended with a promise: find the cliff. Run the MoE model from four concurrent agents upward until the physics says stop. We scaled to eight. The cliff never came. This is Part 3 of the Local LLM Bench series. Part 1 covers the single-request baseline. Part 2 established the MoE advantage under concurrent load. The model: Qwen3-Coder-30B-A3B — a Mixture-of-Experts architecture that activates only 3.3B of its 30B parameters per token. On consumer GPUs, that sparse activation leaves ~90% of memory bandwidth idle at batch size 1, creating headroom that concurrent agents fill. Dense models activate all 32B parameters on every token — already at the bandwidth ceiling before the second agent connects. Part 1 explains why these specific models were chosen (best in class for each architecture); Part 2 conclusively eliminated Dense under concurrent load. This benchmark tests MoE only. ...

In Part 1, we established the baseline: MoE delivers 168 tok/s on a single RTX 3090, 4.1x faster than Dense. Clean single-request numbers. One prompt in, one response out. That’s not how swarms work. An orchestrator like Claude Code dispatches four coding tasks simultaneously. The local model serves all four. Under concurrency, memory bandwidth saturates, per-task throughput drops, and the architecture of the model — not the GPU, the model — determines whether you get useful parallelism or just contention. ...

I went looking for sustained-load benchmarks comparing MoE and Dense coding models on consumer GPUs. Not demo bursts on a Mac Mini — sustained autoregressive generation on real coding tasks, where architecture and interconnect are the only variables. I found plenty of one-shot numbers. Nobody had published the comparison that matters: same hardware, same quantization, same inference engine, MoE versus Dense, across GPU configurations. Methodology visible. Numbers verifiable. So I ran the tests. Dual RTX 3090s with NVLink, custom liquid cooling, a 6 kW isolation transformer feeding a double-conversion UPS. Not elegant, but thermally and electrically honest — sustained inference loads without throttling, no measurement fiction. The hardware details are below. ...

It was 2017. We had just deployed an additional ScaleIO cluster to handle the onboarding of a new customer with hundreds of VMs. Eight nodes, each with 40 Gbps at the backend. Beautiful. Efficient. The whole rack was a work of art—Dell R740s with MD1220 expansions, bezels removed so you could see all those drives blinking in perfect synchronization. The cluster was deployed less than two weeks ago. I told the customer to “burn it.” ...

I’m currently working on the design of a framework for GPU fleet management. We’re living in a crowded data center reality where everybody wants “hero” compute — dense GPUs, fast networking, and delivery that’s closer to the edge. We’re in a land-grab phase where every business wants to be everywhere, but most teams are discovering the same thing: buying GPUs is the easy part. Operating them as a coherent fleet is the hard part. ...

Right now, I’m working on an InfiniBand topology design for a GPU cluster. The math keeps pointing to the same conclusion: scale-out only makes sense when scale-in has topped out. It’s not about CUDA cores. It’s not about tensor throughput. It’s about tail latency. NVLink keeps GPU-to-GPU communication on-package or over short copper links — no NIC, no PCIe host traversal, no protocol stack. For small messages, that means sub-microsecond latency in the hundreds-of-nanoseconds range. InfiniBand NDR switches add sub-microsecond port-to-port latency, but once you include the full path — PCIe to the NIC, driver overhead, fabric hops, and back — real-world GPU-to-GPU latency across nodes often lands in the 3-10μs range depending on message size and topology. ...

The “Everything Is Green” Problem Here’s a realistic scenario I’ve seen in different forms across fleets (this is a composite, not a single true story with exact numbers): A training run is supposed to take ~3–4 weeks. Two weeks in, someone notices the timeline slipping. Not a crash. Not a failure. Just… slow. The job is running 10–30% behind plan, and nobody can point to a smoking gun. The dashboards look perfect: ...