1

I'm reading the NvLink specs and there's a table which says

Semiconductor|Interconnect| Sublink data-rate per data direction | Total data rate (out+in)
Nvidia V100  |NVLink 2.0  | 200 Gbit/s = 25 GByte/s              | 300 GByte/s

Does this mean that NVLink is able to just output 25 GB/s in one direction (e.g. input to the GPU)? If a GPU (e.g. the newest RTX 2080) has 616 GB/s of stated bandwidth, does that mean it will only be able to use 25 GB in memory transfers out of those 616?

Improve this question
0

1 Answer 1

Reset to default
0

Does that mean it will only be able to use 25 GB in memory transfers out of those 616?

No, The limitations you quoted was applicable for Nvidia's previous generation of GPU (Volta). It does not apply to the current generation (Turing).

Does this mean that NVLink is able to just output 25 GB/s in one direction (e.g. input to the GPU)?

Further down in the same article, the architecture of NVLink is explained better, it would be 25 GB/sec per link. If the Nvidia V100 has the same limitation as Nvidia P100 (Pascal), then it would be a total of 150 GB/sec up and another 150 GB/sec down.

Each NVLink (link interface) offers a bidirectional 20 GB/sec up 20 GB/sec down, with 4 links per GP100 GPU, for an aggregate bandwidth of 80 GB/sec up and another 80 GB/sec down.

Source

Improve this answer
1
  • 1
    I should have read the article without being distracted. The total bandwidth for the V100 is 300 GB/sec vs P100 160 GB/sec so each direction would have 150 GB/sec vs 80 GB/sec. Since your question was specifically about the article in question if I happen to find the bandwidth for the Turning hardware, I will update my question (but don't expect that to happen). As written my answers the question as it was originally written.
    – Ramhound
    Commented Aug 20, 2018 at 21:23

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .