AMD Implements xGMI for "Vega 20" as Competition to NVLink

[btarunr]

xGMI (inter-chip global memory interconnect) is a cable-capable version of AMD's Infinity Fabric interconnect. A line of code in the latest version of AMDGPU Linux drivers reveals that "Vega 20" will support xGMI. This line tells the driver to check the state of xGMI link. A practical implementation of this could be inter-card high-bandwidth bridge connectivity that would otherwise saturate the PCI-Express host bus; similar to NVIDIA's usage of the new NVLink bridge for Quadro and Tesla products based on its "Volta" and "Turing" GPU architectures.

By no means should xGMI and NVLink implementations be interpreted as a coming back of multi-GPU to the gaming space. There are still no takers for DirectX 12 multi-GPU, and fewer AAA games support SLI or CrossFire. Even at higher resolutions/refresh-rates, existing SLI/CrossFire physical-layer standards have sufficient bandwidth for multi-GPU. The upcoming GeForce RTX 2000 graphics cards feature a new multi-GPU connector that's physically NVLink, but this is probably an attempt by NVIDIA to discard the legacy SLI bus and minimize redundant interfaces on its silicon. The TU102 and TU104 chips are implemented in the enterprise segment with the Quadro RTX family. The main application of xGMI/NVLink is to make multi-GPU hardware setups abstract to deep-learning software, so hardware can scale in the background with memory access spanning multiple GPUs. "Vega 20" will be launched in Radeon Pro and Radeon Instinct avatars late-2018.



View at TechPowerUp Main Site

 

[londiste]

I really hope xGMI is not the official/public name for it. AMD really needs to get their shit together and use consistent naming. If it is Infinity Fabric, just say it is.

 

[notb]

I really hope xGMI is not the official/public name for it. AMD really needs to get their shit together and use consistent naming. If it is Infinity Fabric, just say it is.

This might be a sign of stepping back from the funky naming and going into more serious marketing. xGMI is an enterprise solution after all.
They tried their luck with "EPYC" and, while it certainly sticks, it just doesn't go well with suits.
You know... not everything that sticks is welcome..
It's like with Game of Thrones, when viewers are just overwhelmed by the number of short-lived supporting characters, but everyone remember Dickon. ;-)

 

[eidairaman1]

I really hope xGMI is not the official/public name for it. AMD really needs to get their shit together and use consistent naming. If it is Infinity Fabric, just say it is.

Why do you care what they call it?

 

[londiste]

Why do you care what they call it?

They have generated a bunch of hype around Infinity Fabric. This is something AMD should leverage.

 

[ZeDestructor]

Why do you care what they call it?

Cause XGMII is already a thing that's got a very, very similar acronym.

EDIT: Also: why aren't they supporting something open like OpenCAPI? or something already deployed like NVLink?

 

[R0H1T]

Cause XGMII is already a thing that's got a very, very similar acronym.

EDIT: Also: why aren't they supporting something open like OpenCAPI? or something already deployed like NVLink?

Why ~ because probably GMI & IF is better than OpenCAPI & Nvlink is proprietary, in case you didn't know?
https://en.wikichip.org/wiki/amd/microarchitectures/zen

 

[ZeDestructor]

Why ~ because probably GMI & IF is better than OpenCAPI & Nvlink is proprietary, in case you didn't know?
https://en.wikichip.org/wiki/amd/microarchitectures/zen

There's also Gen-Z which is being built open from the ground up, and aims to replace the lot.

PS: IF is proprietary too.

 

[R0H1T]

There's also Gen-Z which is being built open from the ground up, and aims to replace the lot.

PS: IF is proprietary too.

Yes, proprietary to AMD. If we're talking about GenZ then AMD is behind it but it'll take Intel & Nvidia's backing, before it becomes any sort of a standard.

 

[ZeDestructor]

Yes, proprietary to AMD. If we're talking about GenZ then AMD is behind it but it'll take Intel & Nvidia's backing, before it becomes any sort of a standard.

Not just AMD. You also have relatively major in the field players like Mellanox, Xilinx, IBM and Google. IBM in particular is interesting because their CPUs support CAPI, OpenCAPI, NVLink using the same physical IP on the CPUs. That means you can have the same ports supporting any of the above 3 protocols. Gen-Z will likely be added with POWER10, and depnding on how the rest of the ecosystem, everyone else will get in line.

If anything, I reckon NVLink 3 or 4 will likely support Gen-Z or be built on Gen-Z because of pressure from Googles, even though nV right now has fairly heavy investmenets into NVLink.