This is similar to NVLink, but with the added benefit that Zen CPUs will likely be able to talk over a PCIe slot to a GPU at much lower latencies and appear to the system as an APU-like coprocesser more than a peripheral.
AMD Vega 20 introducing xGMI (inter-socket Infinity Fabric) links
Noah Garcia
Austin Green
Nathaniel Barnes
HBMeme again, AMD never learn do they?
Nathaniel Hall
Bentley Collins
the other thing of note is that they could be hinting at a much earlier than expected release of 7nm Epyc.
Ethan Turner
So it will barely compete with the 1080TI, let alone the 2080TI that will be launched months before? FFS get it together AMD.
Lincoln Walker
are we finally getting HSA?
Caleb Robinson
the reason hbm1 was so bad was the 4gb limitation
what makes hbm2 so bad?
Chase Reyes
Vega 20 is enterprise GPGPU only, user.
AMD had to choose between Zen and Radeon for R&D shekels, and they probably made the right choice.
Navi in 2019 will be the last rehash of GCN (in 7nm at least) for consumers, and the unnamed "next-gen" ca. 2020 will be their next attempt to do something serious in the consumer GPU segment.
Vega could actually have been pretty damn great if the primitive discard pipeline had been usable, since it boosts geometry throughput 2x-3x, which has always been high-end GCN's bottleneck (never >4 triangle fragments per clock), but they shat the bed from corner cutting driver dev work it seems.
Owen Collins
>Vega 20 is enterprise
does this mean I can run it in an ITX/MATX case and forgo a second GPU but still be able to play the vidya in a virtual machine?
Benjamin Reyes
At least Navi should possibly have it working on launch.
Nolan Perez
The price and supply...
Easton Green
It won't be a consumer card. xGMI will be critical for coherent memory across accelerators; this interconnect will be critical for HSA which has always been AMD's goal and is a milestone for employing a different standard in HPC GPGPU that isn't a CUDA ecosystem.
As for performance itself, the power efficiency will be doubled and performance will be elevated 40% likely due to node advantage. If they do release a prosumer variant like the Vega FE, with the front end clocked that much higher, faster HBM2 as well as a front end it will be only slightly slower than the Titan V, and will be a 17.5TFLOPS FP32 accelerator compared to nvidia's 13TFLOPS. Essentially Vega 7nm will be competitive but it is still just a stepping stone for entering HPC. It will still be handily outclassed in matrix multiplication though (will require 64 clocks per 4x4 matrix as opposed to 1 clock in Tensor Cores).
Aaron Gonzalez
>NVLink
the real joke is that NVLink is only supported by IBM Power8/9.
How does Intel and AMD expect to support future machine learning workloads with the crap CPU-GPU interconnect?
Nicholas Turner
Considereding AMDs poor GPUGPU software ecosystem their hardware would have to be 5x as good for the given price or else why even bother with it. I would welcome something like CUDA or better support for openCL 2.0
Gavin Gomez
I'd buy it if it were competitively priced and at the high end of the market.
Joshua Sanchez
artificial scarcity and inflated prices.
Blake Garcia
Holy shit its literally the same gpu
Amd rebrand confirmed
AMDrones on suicide watch
Wyatt Reed
I'm calling it now: in the next thread the brazilian shill will present this as a fact
Julian Lee
SR-IOV is workstation GCN cards is the eternal cocktease, user.
I don't think even Vega Pros ever allowed this last gen, and the hardware support has existed since at least Fiji/Fury.
Henry Campbell
Absolutely this. No one is going to abandon established CUDA pipelines without a very good reason.
AMD has their work cut out for them.
Joshua Smith
>amd hype
Amdfags will never learn