I have brought two Nvidia RTX 4000 ada SFF Chips to work on a compute-intensive project. Is it possible to connect both to work as a single unit or something near, I have heard that both SLI and NVLink are stopped for RTX 4000 series, Is there any alternative or is there any way to connect to the mother board? Also Is it possible to chip in two cards as both are quite big in size?
Hi, no, this is not possible. IIRC 4000 claass GPUs never supported NVlink. SLi was an outdated, Scanoutbuffer-only connection between 2 GPUs, NVlink is a much faster, and generally usable link between 2 GPUs. we feature that on our compute focussd highest end chips, and used to on the high end gfx chips like RTX6000/8000 or RTX A6000. but no more on RTX 6000 ada. And explicitly not on your RTX 4000SFF Ada.
It would not make 2 GPUs just look like one, but only allow for fast transfers of any data between 2 GPUs, so could help overcome limits of when a single GPUs framebuffer was too small for the job.
But there always is PCIe, and its gotten faster since we invented SLi as well as NVlink. so you can always copy data between 2 GPUs via PCIe, gen4 with the RTX 4000SFF ada. If you are not limited by the size of framebuffer of a single GPU (in which case going for a single bigger GPU would have been the way to go really), you should seek to parallelize your compute job, so both GPUs can contribute to the compute, with full dataset in each framebuffer, rather than thinking NVlink would make it ‘a single GPU’.
Hope this clarifies your questions? regards
-Frank