Seriously? DO you actaully believe your own BS or have you drunk your own kool-aid?

No Vega 20 will support(1) what is called xGMI(Infinity Fabric Based) for GPUs so that’s what will be similar to Nvidia’s Nvlink IP. Also AMD is a founding member of OpenCAPI so maybe there can be that support also for any OpenPower Power CPU to AMD GPU interfacing. Maybe AMD will offer a dual Vega 20 DIEs on a single PCIe card variant similar to the dual GPU variants that AMD has offered in the past, and currently(Radeon Pro V340 uses 2 Vega 10 DIEs in a Binned Vega 10/3584-SPs-enabled-per-die configuration), to get more Compute onto a single PCIe card.

That TSMC 7nm process node should afford greater power savings at the same clock speeds so maybe AMD can double up the Vega 20 DIEs at 7nm and still make the power budget while doubling the SP counts to 8192 and get double the compute that way. xGMI(infinity fabric) could be used to wire up 2 Vaga 20 dies across the PCIe card’s PCB instead of using PCIe. The same route as AMD does in a similar manner on Epyc with each Zen/Zeppelin die making us of GMI(Infinity Fabric Based) coherency traffic across the sockets on Epyc/SP3 motherboards via taking over some PCIe PHY lanes.

It’s the Infinity Fabric that’s AMD’s answer to Nvidia’s NVLink and IF is similar to IBM’s BlueLink in scope. So xGMI for GPUs just like GMI is for Epyc to bring the Infinity Fabric across the socket for Epyc/SP3.

Vega has always had the Infinity Fabric IP included so Vega 20 will finally bring that support for cross GPU DIEs coherecy via an external IF protocol interface called xGMI. And Infinity Fabric is similar to NVLink in that ability to have 2/or more GPUs interfaced in such a coherent manner as to appear more like one logical larger GPU to the software.

“XGMI is a peer-to-peer high-speed interconnect and is based on Infinity Fabric. XGMI is basically AMD’s alternative to NVIDIA’s NVLink for inter-connecting GPUs. ” (1)

[Note: All of AMD’s Current large Desktop/Accelerator Discrete GPUs are based on that one “Vega 10” Base die Tapeout with 4096 SPs. So that’s the Radeon Pro WX 9100 and Radeon instinct MI25, Vega 64(the full 4096 SPs enabled) and Vega 56(3584 SP enabled/binned Vega 10 die). And there already is a dual Vega 10 die based SKU, the Radeon Pro V340 dual(3584 SPs enable dies) for a total of 7168 SPs on one PCIe card and that SKU is for Datacenter Visualization Workloads. The only reason that AMD has not offered any Full Vega 10(4096 SP enabled) Dual GPU/Single PCIe card variants at 14nm is power delivery and power budget constraints, but that may change at 7nm if TSMC’s process affords AMD the opportunity of creating a dual Vega 20 on one PCIe card Compute/AI variant if the power delivery and power budget constraints allow, thermal constraints also]

(1)

“Radeon Vega 20 Will Have XGMI – Linux Patches Posted For This High-Speed Interface”

https://www.phoronix.com/scan.php?page=news_item&px=AMDGPU-XGMI-Vega20-Patches