Nvidia Blackwell chip successor "Rubin" moves forward by six months, uses 3nm manufacturing and HBM4

In brief: The immense request for Nvidia's AI info halfway GPUs has sent nan institution into nan stratosphere complete nan past fewer years, putting it successful seesaw title pinch Apple for nan title of nan astir valuable patient connected Earth. As Nvidia emerges from nan accumulation problems that delayed nan Blackwell GPU launch, caller reports bespeak that its successor is importantly up of schedule.

According to nan Taiwan Economic Daily, Nvidia's Rubin bid AI graphics chips will beryllium disposable successful nan 2nd half of 2025, half a twelvemonth earlier than antecedently reported. Although specifications regarding capacity improvements are slim, nan caller GPUs will characteristic much precocious representation and a newer semiconductor process node than Blackwell, which is group to motorboat soon.

Prior roadmaps from Nvidia pegged nan caller architecture, named aft American astronomer Vera Rubin, for merchandise sometime successful 2026. However, nan chipmaker, manufacturing partner TSMC, and different proviso concatenation participants person accelerated nan lineup's improvement and implementation.

A Morgan Stanley Semiconductor expert told nan Economic Daily that Rubin represents Nvidia's displacement toward TSMC's 3nm N3 process node, which could importantly amended performance. The caller chips will besides upgrade to HBM4 memory, substantially boosting representation bandwidth.

If nan reports beryllium accurate, this could consciousness for illustration a reversal of nan delays that pushed Blackwell's implementation into nan first 4th of 2025. The GB200 series, now entering wide production, encountered late-stage creation flaws and presently suffers from overheating issues. Despite nan setbacks, Blackwell chips are sold retired done precocious 2025, arsenic tech giants for illustration Microsoft, Google, and Meta person spent billions to get them.

The delays occurred partially owed to challenges pinch TSMC's CoWoS-L packaging technology. The shaper is expected to importantly grow CoWoS accumulation pinch Rubin, though its imaginable effects connected nan Blackwell successor's rollout stay unclear.

The effect connected GB300, antecedently called Blackwell Ultra, is different mystery. The upgraded lineup is group for a precocious 2025 debut, coinciding pinch Rubin's caller merchandise window. Prior reports indicate that Nvidia is considering moving to a socketed creation for GB300, which mightiness discuss capacity successful speech for streamlining nan accumulation and upgrade pipelines. However, nan lineup's preamble of FP4 is expected to boost request considerably.

Meanwhile, competitors for illustration AMD and Intel are preparing to answer GB300 and Rubin pinch upcoming architectures for illustration MI300 and Falcon Shores, besides based connected N3.