HBM memory is due a chunky upgrade with HBM5 and HBM6 already in the works, with fresh TC Bonders.
Nvidia and AMD are expected to roll out next-gen AI accelerators this year using HBM4, including the Vera Rubin and Instinct MI450 lines.
The pace is silly, so the industry is already lining up what comes after HBM4, with HBM5 and HBM6 being developed alongside new bonding equipment.
A report from the Korean outlet Heraldcorp states that the first Wide TC Bonder gear for next-gen standards will be unveiled at Semicon Korea 2026.
Wide TC Bonder is being pitched as an alternative to hybrid bonding for mass production of HBM memory, after hybrid bonding kit hit technical snags and slipped.
The promise is higher yield for current and future standards like HBM4, HBM4E, HBM5 and HBM6, plus better quality by leaning on more precise bonding.
One headline trick is fluxless bonding, which removes the oxide layer on the chip surface without flux, thereby boosting bond strength and reducing HBM thickness.
HBM5 is expected to maintain an 8 Gbps data rate for the non-e variant while increasing the number of IO lanes to 4096 and boosting bandwidth to 4 TB/s per stack.
It is described as moving to 16-Hi stacks as a baseline, using 40 Gb DRAM dies to scale to 80 GB per stack, with per-stack power pegged at 100W.
Packaging is listed as microbump MR-MUF, with cooling options including immersion, thermal via TTV, and thermal bonding.
It bakes in a dedicated decoupling capacitor die stack, a custom HBM base die with 3D NMC-HBM and stacked cache, plus LPDDR and CXL in the base die.
The write-up links HBM5 to Nvidia’s Feynman and AMD’s Instinct MI500 platforms, with an on-shelf target of 2029.
HBM6 is framed as a post-Feynman jump, doubling bandwidth to 8 TB/s and pushing 48 Gb capacity per DRAM die.
Stacking is expected to stretch beyond 16-Hi to 20-Hi, lifting capacity to 96 to 120 GB per stack, while per-stack power climbs to 120W.
HBM6 is described as moving to bumpless copper-to-copper direct bonding, maintaining immersion cooling, and experimenting with custom multi-tower HBMs.
The research bucket includes active or hybrid silicon-on-glass interposers, along with extras such as an onboard network switch, bridge die, and asymmetric TSV.
HBM4 is set for mass production this year, and the industry is already preparing HBM5 and HBM6, so the next bottleneck is expected to arrive on schedule.
