Nvidia cancels its four-die Rubin Ultra GPU design due to manufacturing issues, replacing it with a half-performance version

@jukan05 U should probably clarify that they killed this Rubin Ultra as well right?

This 2+2 MCM design was always stupid

@jukan05 @LongGameEquity

@TShirtnJeans2 @jukan05 2+2 and the 4 die is gone I believe

@jukan05 NVIDIA’s original Rubin Ultra planned a single massive 4-die package with ~1 TB HBM4E. Manufacturing challenges (warping, yields in CoWoS-L) led to scaling back to dual-die GPUs (~384–768 GB HBM4E each).They now achieve equivalent performance via a 2+2 board-level configuration:

@jukan05 $AMD to $1,000

@jukan05 Easier production enables higher shipment volumes, faster availability, and comparable (or better) rack-level capability. It’s a pragmatic de-risking, not a downgrade.

@TShirtnJeans2 @jukan05 I believe they wanted it to be single piece of substrate

That failed and then went to MCM style so 2 pcbs with copper trace

And this is also canceled. Atleast that's what implied by Jukan unless I'm wrong

@jukan05 two dual-die packages deliver four dies total per high-end unit. This often results in similar or higher total memory (e.g., 2 × 768 GB = 1.536 TB) while dramatically improving manufacturability and ramp speed result: No major performance cut.

@TShirtnJeans2 @jukan05 I thought this was the MCM 2+2

Why I sent it

@jukan05 Too much compute for these hoes.

@Alex_Intel_ yup

@jukan05 Forget about getting close to $CBRS inference performance

@A431881K @jukan05 How do you know the performance is "equivalent"?

@jukan05 In term of what? Memory? Computer? Token cost?

@jukan05 That ain’t good. Really? What does it mean for HBM per rack?

@jukan05 My strategy analysis !

🔽as follow

@jukan05 @Alex_Intel_ Can you explain what this mean?

@jukan05 Details of my analysis.

👇Next steps

@Alex_Intel_ @jukan05 What does the red line represent? A PCB?

And if no 2+2, what is it replaced with?