Hardware stub

NVIDIA Pascal P100 GPU Enters Volume Production, Ships “Soon”

Posted on April 5, 2016

Pascal is the imminent GPU architecture from nVidia, poised to compete (briefly) with AMD's Polaris, which will later turn into AMD Vega and Navi. Pascal will shift nVidia onto the new memory technologies introduced on AMD's Fury X, but with the updated HBM2 architecture (High Bandwidth Memory architecture version 2); Intel is expected to debut HBM2 on its Xeon Phi HPC CPUs later this year. View previous GTC coverage of Mars 2030 here.

HBM2 operates on a 4096-bit memory bus with a maximum theoretical throughput of 1TB/s. HBM version 1, for reference, operated at 128GB/s per stack on a 1024-bit wide memory bus. On the Fury X – again, for reference – this calculated-out to approximately 512GB/s. HBM2 will double the theoretical memory bandwidth of HBM1.

During the GTC 2016 keynote, nVidia CEO Jen-Hsun Huang noted that the company's new Pascal Hyperscale GPU has entered volume production and will ship (release date) by first quarter, 2017 (1Q17). Huang expects partners Dell, HP, Cray, and IBM Server to announce availability of the P100 within OEM machines by 4Q16, with production availability in 1Q17.

Pascal in the P100 will ship to Cloud first, then OEMs.


The new Pascal-equipped P100 GPU hosts 150B XTORS, 5.3TF of FP64, 10.6TF of FP32, and 21.2TF of FP16. 14MB of SM RF, and 4MB of L2 Cache. The P100 runs on “five miracles” that nVidia has defined as technologically difficult to overcome: a new Pascal architecture, 16nm FinFET process nodes, CoWoS with HBM2, NVLink, and new AI algorithms.

The Tesla P100 is not a consumer-class card, but has already entered volume production and will utilize the architecture that is expected to ship to consumer (GeForce) later this year.

NVidia indicated that, with thanks to Pascal, there's a 12x speed-up year-over-year for super computer processing on Pascal. NVidia's metrics scale the GPU count upward to 8 Pascal GPUs for processing deep learning data, capable of computing in 2 hours what 4 Maxwell GPUs processed in 25 hours.

More information on Pascal within the consumer market as it emerges.

Editorial: Steve "Lelldorianx" Burke
Photography: Jim Vincent.