Navi to use the Infinity Fabric?

[Vanguardas]

this is pure SPECULATION based on the information i've read/heard.

 

Will Navi actually use the Infinity Fabric? Based on their roadmap, I have a feeling it really is the I.F. It would really be cool to have 2 smaller GPUs work as one. Should be cheaper to produce and should allow AMD to catch up to Nvidia. Hopefully Navi can fix the imbalance in the GPU market.

 

What do you guys think?

 

forgive me for my paint skills 

 

 

[VerticalDiscussions]

Paint god.

[Valentyn]

Infinity Fabric was already listed for APUs, and GPUs for it's bandwidth scaling.

 

http://segmentnext.com/2017/01/17/amd-infinity-fabric-details/

 

https://community.amd.com/thread/211126

 

Quote

Infinity Fabric is a coherent implementation which means that cache coherency is maintained across multiple processors externally and scaling up cores, in a CPU or a GPU, is not a problem and only limited by the bandwidth of the transport itself (which we have mentioned above). This philosophy is in contrast to Intel’s stricter vision of a tailor-made design. It also allows AMD to scale up and down designs within a matter of hours rather than months without spending additional human capital or resources. This flexibility will allow it to serve a larger number of custom clients than anyone else.

 

[Vanguardas]

4 minutes ago, VerticalDiscussions said:

Paint god.

Thank you 

 

2 minutes ago, Valentyn said:

Infinity Fabric was already listed for APUs, and GPUs for it's bandwidth scaling.

 

https://community.amd.com/thread/211126

Cool. Thanks! Oh right the new APUs. But it would be cool to see it in an actual video card with 2 fused GPUs using the I.F.

[Valentyn]

2 minutes ago, Vanguardas said:

Thank you 

 

Cool. Thanks! Oh right the new APUs. But it would be cool to see it in an actual video card with 2 fused GPUs using the I.F.

It's highly likely that's Navi. Raja already spoke in 2016 about making multi GPU as common as Multi-Core CPUs; since multiple smaller cores are cheaper and easier to make compared to very large ones as well.

 

[Vanguardas]

2 minutes ago, Valentyn said:

It's highly likely that's Navi. Raja already spoke in 2016 about making multi GPU as common as Multi-Core CPUs; since multiple smaller cores are cheaper and easier to make compared to very large ones as well.

 

Nice. I can sleep comfy tonight (knowing graphics cards might deliver stellar price/perf on 2019).Hope RTG doesn't fuck this up. Nvidia is already a massive cash monster.

[captain cactus]

Note this isn't your typical multi-GPU design we're talking about here. This is about lowering cost of a GPU die by a bazillion % by having multiple smaller dies connected via Infinity Fabric rather than one massive die size and then connecting two massive dies together.

 

Take the Fury X. 4096 Stream Processors. You can either make one massive die of 4096 stream processors with all of the other logic attached (display stuff, memory controller etc) or make 4 1024-SP dies connected via Infinity Fabric to the separately manufactured display stuff, memory controller etc. You end up with way higher yields on your silicon wafers by having more dies per wafer and thus having a higher succes rate per wafer which drives die costs into the ground. Connecting them via Infinity Fabric with the other logic is the final step in the process, and now you have a high-performance GPU for like 2/3rd the cost. Like, 1080Ti performance for 1070 cost.

 

Or AMD can go absolutely bonkers and make like 4 2048 (so RX570-sized minus display/memory logic etc) dies and connect them for a 8192-SP monster of a GPU that eats everything in its path alive at 1080Ti costs. It's entirely possible to do that.

 

Not to mention that to software it's seen as one GPU rather than multiple in CrossFire eliminating the need for game devs to implement CF/SLI scaling. 

 

 

tldr This is AMD going "Hi, we're AMD, and welcome to Jackass".

[Vanguardas]

3 minutes ago, lots of unexplainable lag said:

Note this isn't your typical multi-GPU design we're talking about here. This is about lowering cost of a GPU die by a bazillion % by having multiple smaller dies connected via Infinity Fabric rather than one massive die size and then connecting two massive dies together.

 

Take the Fury X. 4096 Stream Processors. You can either make one massive die of 4096 stream processors with all of the other logic attached (display stuff, memory controller etc) or make 4 1024-SP dies connected via Infinity Fabric to the separately manufactured display stuff, memory controller etc. You end up with way higher yields on your silicon wafers by having more dies per wafer and thus having a higher succes rate per wafer which drives die costs into the ground. Connecting them via Infinity Fabric with the other logic is the final step in the process, and now you have a high-performance GPU for like 2/3rd the cost. Like, 1080Ti performance for 1070 cost.

 

Or AMD can go absolutely bonkers and make like 4 2048 (so RX570-sized minus display/memory logic etc) dies and connect them for a 8192-SP monster of a GPU that eats everything in its path alive at 1080Ti costs. It's entirely possible to do that.

 

Not to mention that to software it's seen as one GPU rather than multiple in CrossFire eliminating the need for game devs to implement CF/SLI scaling. 

 

 

tldr This is AMD going "Hi, we're AMD, and welcome to Jackass".

Yup. I'm also hyped for that. If they can manage to connect more GPUs that would be insane. I.F. would also work insanely well with the high speeds of HBM2? and GGDR6?

 

Hopefully thermals and power draw are kept in check. Well Ryzen thermals and power draw seems fine to me. Fingers crossed, 2019 is my upgrade year. 7nm Ryzen + Navi with I.F.

[Valentyn]

4 minutes ago, Vanguardas said:

Yup. I'm also hyped for that. If they can manage to connect more GPUs that would be insane. I.F. would also work insanely well with the high speeds of HBM2? and GGDR6?

 

Hopefully thermals and power draw are kept in check. Well Ryzen thermals and power draw seems fine to me. Fingers crossed, 2019 is my upgrade year. 7nm Ryzen + Navi with I.F.

 

They'll most likely use HBM2 due to lower latency, and wider channel, to use it all as High Bandwidth cache.

[captain cactus]

22 minutes ago, Vanguardas said:

Yup. I'm also hyped for that. If they can manage to connect more GPUs that would be insane. I.F. would also work insanely well with the high speeds of HBM2? and GGDR6?

 

Hopefully thermals and power draw are kept in check. Well Ryzen thermals and power draw seems fine to me. Fingers crossed, 2019 is my upgrade year. 7nm Ryzen + Navi with I.F.

 

17 minutes ago, Valentyn said:

 

They'll most likely use HBM2 due to lower latency, and wider channel, to use it all as High Bandwidth cache.

Yeah, I think it's save to say AMD wants to step away form traditional VRAM and use the word VCache for it or something since it's not really used as traditional VRAM.

 

AFAIK this first gen of I.F. can already scale up to like 512GB/sec or something so raw data throughput won't be an issue.

[Valentyn]

2 minutes ago, lots of unexplainable lag said:

 

Yeah, I think it's save to say AMD wants to step away form traditional VRAM and use the word VCache for it or something since it's not really used as traditional VRAM.

 

AFAIK this first gen of I.F. can already scale up to like 512GB/sec or something so raw data throughput won't be an issue.

Yes it can, which is awesome; and why Naples/Epyc can handle so much.
I can only imagine what I.F 2.0 will be able to do.

 

They've always brought out some very interesting tech.