Hopper
Nvidia's GPU architecture generation including the H100 and H200, preceding Blackwell.
As defined in Dwarkesh Patel Podcast - Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute
Other definitions
A funnel-shaped container or device used to supply raw materials (such as clay) onto a conveyor belt or into processing equipment in a controlled manner
日本最古の煉瓦工場で赤レンガを製造するプロセス。明治時代から受け継がれるレンガ製造の技術と歴史
Where it comes up · 7
- Nvidia v. Google, Scaling Laws, and the Economics of AI5 claims
- Ideal Episode Length, Monetization, and Production Process1 claim
- How AI Datacenters Eat the World1 claim
- Dwarkesh Patel Podcast - Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute1 claim
- Gavin Baker - AI, Semiconductors, and the Robotic Frontier - [Invest Like the Best, EP.385]1 claim
- Mike Munger on Private and Public Rent-Seeking (and Chilean Buses) 08/23/20101 claim
- 日本最古の煉瓦工場で赤レンガを製造するプロセス。明治時代から受け継がれるレンガ製造の技術と歴史1 claim
Most important claims · 10 of 11Search all claims for “Hopper” →
- A single pre-training cluster cannot exceed roughly 200,000 Hoppers that are coherent, meaning all GPUs share memory and act as one system, so progress requires waiting for the next chip generation to scale further.Nvidia v. Google, Scaling Laws, and the Economics of AI
- The move from Hopper to Blackwell was the most complex product transition in technology history, requiring a shift from air-cooled to liquid-cooled racks, rack weight rising from about 1,000 to 3,000 pounds, and power density rising from 30 to 130 kilowatts per rack.Nvidia v. Google, Scaling Laws, and the Economics of AI
- Even after Blackwell is deployed, it takes six to nine months to match Hopper performance because Hopper is finely tuned and engineers know its quirks; Hopper itself took six to 12 months to outperform its predecessor.Nvidia v. Google, Scaling Laws, and the Economics of AI
- Pre-training scaling depended on the next generation of coherent chips, and since xAI could not get more than about 200,000 Hoppers coherent, pre-training progress would have stalled from mid-2024 until Gemini 3, but reasoning models let AI keep advancing without new hardware.Nvidia v. Google, Scaling Laws, and the Economics of AI
- Nvidia GPU power consumption and performance has skyrocketed across generations: Volta 250W TDP, Ampere 400W, Hopper 700W, Blackwell 1000W, with GB200 superchip reaching 2,700W, and future GPUs approaching 2000W per chip, demonstrating exponential density increases.How AI Datacenters Eat the World
- When Colossus has many produced episodes banked, hosts shift attention away from guest sourcing, which depletes the pipeline and later creates shortages.Ideal Episode Length, Monetization, and Production Process
- A friend of Munger's who worked in the Michigan Legislature in the 1970s reported that at the start of each session about 20 bills were put in the hopper, each proposing new taxes or regulation on a particular industry. Each bill was withdrawn as soon as that industry made a campaign contribution. Munger notes this was relatively 'efficient' rent extraction: because it was sector-specific and paid as direct contributions, it was closer to a pure transfer than in-kind lobbying. Roberts labels it extortion.Mike Munger on Private and Public Rent-Seeking (and Chilean Buses) 08/23/2010
- xAI built a 100,000-GPU coherent training cluster in Memphis by redesigning the data center from first principles, achieving enough density to train on Hopper GPUs without next-generation networking from Nvidia or Broadcom.Gavin Baker - AI, Semiconductors, and the Robotic Frontier - [Invest Like the Best, EP.385]
- The factory sources Mikawa clay from two different excavation sites and blends them together to stabilize the quality of the final product.日本最古の煉瓦工場で赤レンガを製造するプロセス。明治時代から受け継がれるレンガ製造の技術と歴史
- H100 rental prices have inflected upward, with some AI labs signing two-to-three-year deals as high as $2.40/hour for Hopper even though it costs ~$1.40/hour to build over five years—margins far above the original ~35%—because rolling-off short-term contracts let willing buyers crowd out other suppliers.Dwarkesh Patel Podcast - Dylan Patel — Deep dive on the 3 big bottlenecks to scaling AI compute