byronyi

5 ms·

1. 1.▲
   by byronyi 5y ago

   The analysis is bullshit.
2. 2.▲
   by byronyi 5y ago

   I'd bet on SPIR-V.
3. 3.▲
   by byronyi 5y ago

   +1 please put some info in your profile.
4. 4.▲
   by byronyi 7y ago

   I don't get it. How is it different from "Wide & Deep Learning for Recommender Systems (2016)" by Google? The whole project, up til now, contains less than 2,500 lines of Python, and 271 lines of Markdown as doc, 159 line
5. 5.▲
   by byronyi 8y ago

   Unlike TPU that is only made available through Google Cloud Platform, Ascend 910 will be available on market in 2019 Q2, in form of PCIe line cards and OEM servers.
6. 6.▲
   256 TFlops on a Single 7nm Chip: Huawei Ascend 910 Unveiled(mybroadband.co.za)

   5 points by byronyi 8y ago|1 comments
7. 7.▲
   Show HN: GPU Direct RDMA Out-Of-Band Tensor Transport for TensorFlow(github.com)

   3 points by byronyi 9y ago|0 comments
8. 8.▲
   by byronyi 10y ago

   Indeed. A more interesting comparison would be between C/C++ based shared library and Rust based one.
9. 9.▲
   by byronyi 10y ago

   Why does it have to do with scaling in the first place? Performance has nothing to do with scalability whatsoever, by the way.