Serving an LLM with a one-million-token context on a self-hosted machine is not only a VRAM problem. It is mostly a PCIe topology, BAR window, and system configuration problem. The report published on Reddit by Primary_Exchange21 shows this concretely: a build with 16 RTX 5060 Ti 16GB cards, spread across two islands connected by Broadcom/PLX PEX88096 switches, runs DeepSeek V4 Flash-0731 with reported generation between 100 and 150 tokens per second and an average of 140 tokens per second in the DeepSeek Harness in a tensor parallel 8 and pipeline parallel 2 configuration.

The base is an ASRock Rack SPC621D8U-2T/OVH motherboard with a Xeon Gold 6330. The key technical detail is not the GPU itself but the requirement for a 16,384 MiB BAR1 on every card. To reach it, the builder enabled Above 4G Decoding, set MMIO High Granularity to 1024G, set MMIO High Base around 56T, disabled SR-IOV, and added GRUB parameters such as intel_iommu=off, pci=realloc=on, and hpmmioprefsize=512G. The NVIDIA module also needs NVreg_EnableResizableBar=1. UEFI boot is enabled, CSM and Secure Boot are disabled because the locally built EFI application and the patched NVIDIA modules are unsigned. The driver is the patched Aikitoria open driver 610.43.02-p2p.

The less visible work is on the PLX switches. For every PEX bridge, the builder writes the value ECAP_ACS+0x6.w = 0000 to the ACS control register to prevent PCIe access controls from blocking peer-to-peer traffic inside each cluster. Then comes a custom all-reduce for the PLX clusters and DSpark work for pipeline parallel execution. The result depends on the configuration: with tensor parallel 8 and pipeline parallel 2, the available context is 500,000 tokens, with about 4000 tokens per second of prompt processing up to 500,000 tokens; with tensor parallel 4 and pipeline parallel 4, the context reaches 1 million tokens and prompt processing rises to about 7000 tokens per second up to 500,000 tokens, while generation drops to 80 tokens per second.

The news is not the benchmark. It is the kind of economics that emerges. According to the report, the whole system cost 0.6 times an RTX 6000 Pro: a very different TCO indicator compared with a traditional accelerated server. The budget shifts toward consumer components, but the real cost is system knowledge. This is not something replicable with a manual: it requires patched unsigned kernel modules, a locally compiled EFI application, aggressive boot parameters, and register-level changes on PCIe bridges. Anyone running such a machine must accept Secure Boot being off and low-level PCIe management.

With 16 cards of 16 GB each, total VRAM is 256 GB, but the real resource is not the sum of the memories: it is the ability to make them cooperate on long contexts. The bottleneck here is not GPU price but the platform: the MMIO window, BAR reallocation, and peer-to-peer traffic through the switches. For those evaluating on-premise deployments, AI-RADAR covers these trade-offs in the analytical framework at /llm-onpremise.

For labs and professionals with deep system skills, this experiment rewards the ability to turn consumer hardware into long-context inference capacity. Vendors that sell value only through single-GPU certification risk losing relevance: here the asset is platform control, not the individual component. The enterprise limitation remains: such a deployment requires maintenance, certified reproducibility, and an update strategy that vendor configurations offer in a more mature way. The structural signal is that data sovereignty and self-hosted inference on long contexts are becoming a platform engineering problem, not a race to buy more VRAM.