Genuinely divided, recorded as a split: on whether the Strix Halo is a bad buy, the thread splits into owners who are delighted and buyers who are not, and the split is not resolvable by benchmarks

A post titled 'Why the Strix Halo is a poor purchase for most people' (103 upvotes, 394 comments) produced a thread where the top-voted response is a single line of mockery at 182 upvotes - 'This just in, more expensive computers are faster than less expensive computers. More at 11.' Underneath that the thread splits cleanly. Owners: 69 upvotes from someone who moved from a 2x3090 + 128GB DDR5 desktop and reports 'GLM 4.5 Air doing inference at 120w is faster than the same model running on my 800w desktop. And now my pc is free for gaming again'; 13 upvotes 'You'll love it. Don't listen to this guy. Everyone I know with Strix halos loves them... AMD is making ROCm better and better'; another 13 from a triple-GPU owner who says 'for the purpose for which I purchased it the Halo is WAY more useful and considerably faster than the 5090'. Buyers and detractors: 34 upvotes 'yeah no offense but you're advising people to go spend potentially thousands more and draw a heap more power for doing inference'; 22 upvotes conceding it is 'slightly overpriced and over hyped'. Thread-regulars also object to the methodology: 19 upvotes 'you kinda need to compare rigs of similar price, no?' and 20 upvotes 'is this bait?'

The disagreement is about which buyer the card is for, and both camps are internally consistent. The Halo's advantage is capacity and idle power in a small, quiet, always-on box; the 5090 box's advantage is prefill throughput, which is bounded by PCIe bandwidth on a Halo because it only offers x4. The OP's own measurements show decode roughly tied at context 0 and the 5090 pulling ahead as context grows - which is a real result, and one the author concedes he cannot explain on the prefill side ('Good question! I would love to know!'). A separate thread shows the numbers being actively contested as stale: 24 upvotes with a fresh ROCm 7.10 run, and 16 upvotes 'Those must be ancient numbers. Since the Strix Halo is better than that now and getting better everyday... In TG, I dare say it goes toe to toe with the 5090. Even at large context.' A third objection is that the thread never established the use case before declaring a verdict - which is the same refusal of a context-free verdict that shows up in the engine slice.

Decide on the workload, not the card. A Halo is the right answer for large-model capacity in a small always-on low-idle-power box; a 5090 desktop is the right answer for prefill throughput and anything that fits in VRAM. Compare at matched price and matched context depth. One owner's framing at 13 upvotes is the useful part: 'I have a Halo. I also have a 5090 (and an RTX Pro 6000 too).'

This is the flagship 'do not record a false consensus' record for the sentiment slice. Vote-share is not recorded as a percentage: the 182-upvote mockery is a reply, and counting replies against the post's 103 upvotes would misrepresent the split. What is defensible is the shape - a large volume of satisfied owners on one side, a substantial group calling it overpriced on the other, and a methodological dispute in between about price matching and the staleness of the numbers. Engine-side note that belongs here: vLLM does not work well on this class of hardware, which is part of why llama.cpp dominates the Halo conversation.