The 16GB card on sale runs at 30 Gbps on a 256-bit bus, yielding 960 GB/s. The Chiphell packaging leak confirms the 24GB variant keeps the same 256-bit bus. The 1024 GB/s figure traces to an older kopite7kimi post that the checked sheet treats as a separate card, rated at 32 Gbps on the same bus width. The packaging leak states capacity and bus width but no pin rate.

At 4-bit weights, 32 billion parameters consume 16GB. On the 16GB card that leaves zero for context. On the 24GB card, 8GB remains. Models at 8B and 14B fit in both; 70B fits in neither.

The bandwidth gap, if the 32 Gbps rate holds, is 64 GB/s, about 7 percent. Token speed follows bandwidth. But the structural change is capacity: it moves the 32B class from does-not-fit to fits-with-room.

Sources: Chiphell packaging leak via NotebookCheck, OC3D, kopite7kimi on X. The hole: Nvidia has not announced either variant. The 24GB figure rests on a packaging leak and a single rumour. The 32 Gbps pin rate belongs to an older, separate leak. Nobody has measured either card.

The claim conflates two leaks. The Chiphell packaging leak confirms 24GB and 256-bit bus but states no pin rate. The 32 Gbps / 1024 GB/s figure comes from an older kopite7kimi post that the checked sheet explicitly treats as a separate card. The 7% bandwidth gain is therefore unconfirmed for the 24GB variant. What is confirmed by the packaging leak is the capacity jump, and that is where the real structural change lives: 32B at 4-bit weights needs 16GB, which the 16GB card cannot accommodate with any context headroom, while the 24GB card leaves 8GB. The bandwidth question remains open until Nvidia announces a pin rate for the 24GB variant.