News for hackers. Stuff that matters.
Re:Nvidia Driver not yet supported for Linux Kernel 5.9 by bayindirh on Monday October 19, 2020 @02:23PM (#24829255)
Just food for thought:

What if AMD actually acquires Xilinix?

Everybody thought it's about CPU-FPGA integration but, they may pull something very, very different.

Re:Nvidia Driver not yet supported for Linux Kernel 5.9 by dragontamer on Monday October 19, 2020 @02:36PM (#24829408)
I'm not an expert on Xilinx's stuff, but Xilinx's 18-bit x 27-bit multiplier DSP48 architecture still seems quite awkward to me.

https://www.xilinx.com/support/documentation/white_papers/wp...

True, a systolic processor can be built out of these components, but I doubt it'd be anything as fast as a proper matrix-multiplication unit that NVidia is putting out. And that's after the complications associated with FPGA programming.

----------

Xilinx's FPGAs are a combination of LUTs (4-bit or 6-bit look-up tables), Routing, and "DSP Slices" (prefabricated multipliers). The bulk of your math is still going to be performed in the DSP Slice.

Where FPGAs win is that their LUTs and Routing tables allow you to create custom high-speed glue-logic between these DSP slices. But when it comes to something like a regular Convolutional neural net, the routing is very straightforward.

------

I'm not really a deep learning expert. I know that FP16 seems to be pushed by NVidia. This INT8 stuff seems odd to me (there's no way INT8 gets the same dynamic range as FP16), but I'm also not really sure if the dynamic range is needed or useful in deep learning applications.

Re:Nvidia Driver not yet supported for Linux Kernel 5.9 by klyrs on Monday October 19, 2020 @03:36PM (#24830038)
I've done some systolic processors on Xilinx hardware. It's really dang hard, on their biggest, fanciest chips, to pump the DSPs at anything close to their advertised clock speeds. Maybe if I was smarter, I could do an asynchronous design that would somehow avoid clocking issues, but I suspect that doing async between mismatched clock regions would effectively halve the clock (and that might be optimistic).

The only consolation for me is that I/O over PCI ends up being my ultimate bottleneck. If AMD did a full-on integration with their infinity fabric, it could actually be a game changer. For me, in my weirdo niche.

Re:Nvidia Driver not yet supported for Linux Kernel 5.9 by dragontamer on Monday October 19, 2020 @04:35PM (#24830712)
I'm pretty sure I/O over PCIe is a substantial bottleneck in many applications (which is part of the reason for the 300GB/s POWER9 / OpenCAPI / NVLink / GPUs). Its a shame that NVLink is isolated to pretty much the DGX-line of computers and POWER9 (The "big" ones too, I don't think Talos II supports NVLink). But I guess an interconnect that fast will require dedicated, non-standard hardware.

----------

AMD's GPU Infinity Fabric seems to be a different infinity fabric from their CPU. CPU Infinity Fabric is ~50GB/s, but the GPU Infinity Fabric is GPU-to-GPU links of ~90GB/s.

Realistically, PCIe 4.0 and 5.0 will be the path forward for standard PCs. There will always be incredible application-specific solutions (such as NVLink / NVSwitch on the DGX), but the general computer user will wait for the open standards to support such speeds before adopting them.

At the moment, GPU-to-GPU communications probably will scale better than CPU-to-GPU comms. CPUs just aren't built for the bandwidth (outside of POWER9 / POWER10).