CUDA under FreeBSD

[Jan Beich]

Shane Ambler <FreeBSD at ShaneWare.Biz> writes:

> On 04/12/2017 21:19, galtsev at kicp.uchicago.edu wrote:
>
>> 
>> On Mon, December 4, 2017 4:24 am, Carmel NY wrote:
>>> Out of morbid curiosity, I was wondering if anyone could tell me
>>> the real reason that Nvidia does not support CUDA under
>> 
>> Arrogance would be my guess.
>
> The morbid part is that they give us the linux libcuda, so we should be
> able to run linux binaries that use cuda, just not native apps.

Modern CUDA toolkit is 64bit but runtime only works on 32bit (bug 206711).
Building as -m32 is probably still possible but may not fit all workloads
nor run as fast.

>
>>> FreeBSD? Also, what are the realistic expectations for it getting 
>>> supported shortly?
>> 
>> Zero is my estimate. The way to let one's steam about them is just
>> not to buy ther hardware. Their attitude to open sourse and
>> unwillingness to disclose details of their hardware was always much
>> worse than that of their competitors (ATI/AMD, matrox...).
>> 
>> This is just my opinion based on my subjective observations.
>
> I'm sure on an episode of bsdnow, they mentioned asking an nvidia dev
> at one of the conferences and they said there shouldn't be any technical
> reason, it just isn't enabled in the build and they would look into it.
> Still hasn't helped any.

Just like Vulkan, just like KMS, just like encoding/capture acceleration.
NVIDIA always conveniently forgets about FreeBSD. However, the ailment
isn't really specific to NVIDIA but affects most binary blob vendors.
For one, Widevine CDM is maintained by Google but EME itself was pushed
to W3C by Netflix, a FreeBSD vendor which conveniently forgot a browser
can run on FreeBSD.