Pascal GP104 Architecture Unsurprisingly Deploying GDDR5X at Low-End

This isn't news to anyone who's followed the site through our Pascal and GDDR5X posts, but new leaks by “benchlife.info” indicate nVidia's dedication to use both HBM and GDDR5X. The Chinese language site has previously proven to be reliable in its leaks.

GPU architecture has come to a head with memory. Pascal will host HBM2 on its high-end devices, but the cost makes low-end and mid-range cards (the equivalent of a current GTX 960) impossibly expensive. NVidia plans to deploy Micron's new GDDR5X high-bit-rate memory for a cheaper alternative to HBM2; GDDR5X is more expensive than GDDR5, landing it between the oldest (current) and newest (current) technology in product cost.

We've previously disclosed that GDDR5X is targeting an end-game bit-rate of 14Gbps. Micron's initial low-volume samples have tested at 13Gbps, a 47% delta improvement over GDDR5's 8Gbps throughput. GDDR5X requires less power and outputs greater performance than current GDDR5.

Benchlife points toward nVidia's Pascal GP104 GPU as relying on off-die memory (GDDR5X, specifically) for its VRAM. This has not been officially confirmed by nVidia and remains under the “rumor” classification at this time, but we'd rate this one as highly plausible.

We'll be at GDC this week to hold conversations and interviews with AMD and nVidia both, and hope to continue this dialogue into April. More soon.

- Steve “Lelldorianx” Burke.