The fall is so stupid U don't really need 1.5TB per CPU for training, chatbot interactions, and light agentic workloads Provisioning that much DRAM for them will just lead to underutilization and waste of resources (500GB-750GB per CPU is enough) My biggest fear was Nvidia being unable to sell finished Rubin NVL72 racks becuz of the DRAM shortage Now they can handle demand in a much better way
Many users see NVIDIA's SOCAMM capacity cuts and dual DRAM configs for Rubin as a smart optimization that supports long-term memory demand and makes waste optional, while one calls the moves very dumb.
Nvidia is offering dual configs 1.5TB per CPU for racks assigned to agentic workloads 750GB per CPU for the rest of the stuff
The fall is so stupid U don't really need 1.5TB per CPU for training, chatbot interactions, and light agentic workloads Provisioning that much DRAM for them will just lead to underutilization and waste of resources (500GB-750GB per CPU is enough) My biggest fear was Nvidia being unable to sell finished Rubin NVL72 racks becuz of the DRAM shortage Now they can handle demand in a much better way
@zephyr_z9 Sk hynix prob way more down than this since Korea closed before the blood started. The European ADR closed -20,53% yesterday
@zephyr_z9 For those who need a cheat sheet btw this is VERY BULLISH $MU
@cryptoendgamer @zephyr_z9 Okay $7k level before aping
@zephyr_z9 So resume up next week?
@zephyr_z9 Many stocks went down 10%, not just memory stocks. Out of curiosity, how are correlating this specific news to the broader selloff?
@zephyr_z9 @paurooteri 秋？ってなったけど 文脈からThe fall の誤訳と思ったら案の定…w
@zephyr_z9 I think a lot of it has to do with temporary mechanical leverage unwinding rather than any fundamentals
@zephyr_z9 Panic sell-off amidst all this SpaceX shit we will be back soon
@zephyr_z9 Agentic workloads need massive memory for multi-step reasoning context. Nvidia's split makes sense: prioritize bandwidth for inference, not storage. Shows they're optimizing for LLM memory walls, not GPU count. Smart segmentation.
@zephyr_z9 I mean it meaningfully changes supply demand forecasting if you don’t need 1.5TB per CPU but I hear you
@zephyr_z9 It is time for On-Prem.
All companies I know don't want to donate their data lake to anthropic / openai / amazon / microsoft
Signing LOI with the cloud service providers fell of a cliff.
I see 100th and 1000th of small installations.
@zephyr_z9 The utilization point makes sense for current workloads
Wonder how this shapes infrastructure design as models keep scaling
@zephyr_z9 damn right they don't need 1.5tb per cpu for basic chat. my tracker caught this supply chain shift early.
@zephyr_z9 So you crashed the market for nothing? :D
@zephyr_z9 Obvious all of this is VERY dumb
@zephyr_z9 Provisioning 1.5tb per cpu for chatbots is overkill
underutilization wastes resources and inflates costs efficiency gains from right sizing hardware are significant for scaling ai workloads
@zephyr_z9 Written October 11th. https://open.substack.com/pub/jamiefollese/p/o3o2o1?r=22e59f&utm_medium=ios nobody cares until it’s over as usual
@zephyr_z9 Yep, been a matter of time for quite some time. Just wasn’t a popular position to take until about a week ago. https://open.substack.com/pub/jamiefollese/p/the-era-of-anti-scaling-is-here?r=22e59f&utm_medium=ios
@zephyr_z9 750 per CPU for general workloads is realistic. the dual config move makes the waste optional which is actually smart