The configuration previously considered unstable—four Radeon AI PRO R9700 cards attached to a Threadripper PRO 9955WX—has passed its first real test. The affinity engine, which refused to start with more than two GPUs, ran DeepSeek-V4-Flash-0731 at 2.875 bits per weight without degenerate output, hitting 79 tok/s in greedy decoding and about 90 tok/s on a chat response with thinking enabled.

The problem was not in the multi-rank software but in the validation setup used by someone who had only two cards and simulated four ranks on the same pair. The simulation produced repeated tokens, leading the author to block configurations with more than two cards. We now know that block was an artifact of the test bed, not an architectural limit.

On the same system, two-card decoding stopped at 36 tok/s: 45% of the experts fit in VRAM, the rest streamed from a 51 GB pinned host pool, and the profiler attributed 82% of each token's time to that wait. With four cards, all 11,008 experts reside entirely in VRAM and the host pool is not even allocated. The jump from 36 to 79 tok/s happens because execution no longer depends on system memory.

Anyone repeating the test needs to pass --placement-strategy static: the default placement strategy tries to demote experts to the host pool, but when everything is resident there is nothing to move and the engine stops with an error. Loading takes about a minute, 42 seconds of which are spent transferring 96 GB of experts to the cards.

The story matters beyond the single benchmark. It signals that the barrier to using AMD GPUs on self-hosted MoE workloads is no longer compute power, but the maturity of validation procedures. Testing with two cards simulating four ranks created a false negative that kept a working mode locked. When real hardware does not produce repeated tokens, the problem disappears. For teams building on-premises nodes with AMD cards, this changes incentives: they do not need to wait for a single GPU with massive VRAM; they can aggregate four accelerators on CPU PCIe lanes with peer-to-peer between every pair, a cheaper and more modular configuration.

The winners are teams that prioritize data sovereignty and TCO: they can run an 11,008-expert model on local hardware without renting cloud instances. Potential losers are vendors selling single high-capacity GPUs as the only path for large-model inference; if four smaller cards are enough, the market shifts toward more flexible aggregates.

There is also a structural signal for frameworks: affinity's built-in cross-rank check, which compares cards every few samples, passed every test. This type of internal verification reduces the risk of silently corrupted outputs in multi-GPU deployments, a critical aspect when the hardware does not have a track record as established as NVIDIA's.

The test also corrects a previous assumption: affinity's collective was already written for N ranks and tested on 3, 4, and 8 cards, not limited to two like Radiance before the patch. The author is leaving the system available for further tests and offering logs. The next step will be to see whether behavior remains stable on longer sequences and mixed workloads, but the starting data indicates the problem was in the test bed, not in the multi-rank architecture.