Ruby_Rapids (@RubyRapids) header
Ruby_Rapids (@RubyRapids) profile picture
Sol-3
Joined December 2021
Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)Photo by Ruby_Rapids (@RubyRapids)
Replying to @hms1193
Maxsun is a brand from 商科. The board should be under 商科 product line.
86
Even for custom HBM, DRAM dies are largely universal.
33
Replying to @LiDaosen
JEDEC HBM has always been a universal product. The current HBM4 is just JEDEC HBM that optimized for specific chip.
1
233
Replying to @jukan05
요청하신 페이지를 찾을 수 없습니다.
403
Replying to @jukan05
排骨年糕
383
Replying to @jukan05
32GB HBM4 for HGX Rubin NVL8 xitter.cf/RubyRapids/status/2114…
HGX Rubin NVL8 (Error corrected) NVFP4 Training: 280PFLOPS → 270PFLOPS FP8/6 Training: 140PFLOPS → 130PFLOPS HBM4: 2.3TB 176TB/s → 2TB 130TB/s NVLink Switch Bandwidth: 28.8TB/s → 24TB/s Bi-directional Networking Bandwidth: 1.6TB/s → 1.8TB/s
7
HGX Rubin NVL8 (Error corrected) NVFP4 Training: 280PFLOPS → 270PFLOPS FP8/6 Training: 140PFLOPS → 130PFLOPS HBM4: 2.3TB 176TB/s → 2TB 130TB/s NVLink Switch Bandwidth: 28.8TB/s → 24TB/s Bi-directional Networking Bandwidth: 1.6TB/s → 1.8TB/s
1
45
NVIDIA is using all the available supply to meet the demand.
17
All Vera: LDDR5X: 1.5TB → Up to 1.5TB
2
78
Replying to @harukaze5719
Networking Bandwidth is bigger now.
1
82
70B$ / GW is significantly higher than the revenue break-even point.
25