@VirtualElena The problem is that it's fundamentally a conflict of interest - do you serve the user as best as possible, or the ad buyers? You try to strike a happy medium, but the extra revenue from more ads is seductive. The best product recommendation doesn't need to pay to be recommended.
@0xSero The big headline bandwidth numbers are from large sequential transfers on the SSD side, if you're doing random 4k lookups, the numbers drop dramatically. But if the part of the lookup is concentrated, then maybe the important part will sit mostly in RAM anyway.
@0xSero Depends on what you mean, 8 channel DDR4 3200 is at around 200 GB/sec, but if you're talking about 2 channel, then yeah, that's around 50 GB/sec. Latency is 500-1000x higher. But maybe not significant for this offload? Interested to see your results!
@0xSero Running 2x6000 RTX Blackwells, Omarchy/NixOS for the last couple years, currently on Quattro. GLM 5.3 Flash on VLLM for my personal agents, before that, DS v4 Flash 0731, before that DS v4 Flash Preview and Laguna. Also M4 Max 128g. No need for $, just interested.
@jonrobflaig@SpaceX@Kekius_Sage@grok They are, just different nozzles, which extract the work of the propellant best in different conditions. Small work best near sea level, large work best in space. I think the little ones might be the only ones that can gimbal/change direction, to control the landing.