Nvidia Rubin Claims Overstate Blackwell Throughput Gains
Posts examine Nvidia claims on Rubin versus Blackwell token throughput under real lab conditions.
Reply by @yacineMTB states that Nvidia's assertion of 10x token throughput for Rubin over Blackwell applies only under extreme per-user interactivity. Typical AI labs operate at 50-60 tokens per second where the claimed gains do not materialize. Quote by @teortaxesTex discusses shifts in usage profile at 150 TPS and notes Blackwell efficiency at 30M TPS per MW, equating to 108M tokens per 1 kWh. Both posts reference generated headlines on the topic without independent confirmation of the performance figures.
Combined views
7.1K
2 posts, first seen 1d ago
Nvidia Rubin Claims Overstate Blackwell Throughput Gains
Posts examine Nvidia claims on Rubin versus Blackwell token throughput under real lab conditions.
Reply by @yacineMTB states that Nvidia's assertion of 10x token throughput for Rubin over Blackwell applies only under extreme per-user interactivity. Typical AI labs operate at 50-60 tokens per second where the claimed gains do not materialize. Quote by @teortaxesTex discusses shifts in usage profile at 150 TPS and notes Blackwell efficiency at 30M TPS per MW, equating to 108M tokens per 1 kWh. Both posts reference generated headlines on the topic without independent confirmation of the performance figures.
Reactions
2 postsKA 
kache@yacineMTB@zephyr_z9 The more you sell the more you save
- likes: 5
- replies: 0
- bookmarks: 0
- reposts: 0
T( Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTexthis changes the dominant usage profile. 150 TPS is just, what, 50% of peak throughput at 50? This is a great deal, people will pay 2x for 3x speed in a heartbeat, especially for something as cheap as V4-Pro. …wait. 30M tps/1MW is… 108M tokens per 1 kWh?
- likes: 51
- replies: 4
- bookmarks: 6
- reposts: 0
