AMD's Navi Rumored to Launch at E3 2019

[Guest]

Sources report that AMD may be launching Navi as early as E3 2019 with a budget GPU angle.

AMD's Navi Rumored to Launch at E3 2019 : Read more

 

[Supahos]

Shaping up to be an interesting "middle of 2019" for AMD.

If nothing else the landscape of gaming pcs will be drastically different overall by Christmas between ryzen 3000/Vega 7/Navi/ Intel 10nm cpus

 

[TCA_ChinChin]

My hype for AMD's new CPUs and GPUs is immeasurable right now, but I'm afraid for that exact same reason.

 

[shrapnel_indie]

Interesting times are ahead...

 

[redgarl]

One thing for sure, we will know if the leaks from AdoredTV are true at some point this year.

If AMD can match Vega 64 + 5-10% performances for 250-275$, than the whole RTX series will be in deep trouble.

 

[tim.hotze]

AMD is really becoming a "tale of two companies." On one hand, they've executed on their Zen/Ryzen roadmap like clockwork since 2017, and they were touting their performance/watt efficiency advantage over Intel at CES.

On the other, their GPU releases look NOTHING like that chart: They never released a "14nm+" (probably TSMC 12nm) Vega GPU, but they DID release a 12nm Polaris GPU (not on the chart) and a 7nm Vega GPU (the Radeon VII and the Instinct that it's based on, also not on that chart). We'll see what Navi brings, but given that there basically no rumors on Radeon VII, I wonder how accurate the Navi rumors are. It's been a very, very long time since anyone's released a GPU that targeted the middle before going for the high end.

If I had to make a bet, I'd guess that Navi is 7nm Vega with a GDDR memory controller to keep prices down. AMD has tons of experience mixing and matching GPU features with its semi-custom business (Kaby Lake-G's Radeon, for example), so it seems like something they could do without too much effort. They MIGHT add ASICs for raytracing as Nvidia did, or they might not.

That means the 2020 release would be a new (post-GCN) architecture that would probably start at the high end before working its way down (just as we've traditionally seen from Nvidia and AMD).

 

[JamesSneed]

AMD has been very tight lipped with Navi so I really expect they have something special with Navi. I really hope they surprise us as we desperately need competition back in the GPU space.

 

[InvalidError]

AMD has been very tight lipped with Navi so I really expect they have something special with Navi.

Nothing too special: you have ~3X the transistor density on 7nm that you had at 14nm, so you can cram twice the processing power in 40-50% less space depending on the frequency bump and architectural efficiency gains, which means you should be able to provide roughly twice the performance at every price point if you wanted to.

Assuming Navi is a decent improvement over Polaris and Vega on performance and efficiency, how much of a success Navi will be will hinge largely on how much of the savings from increased performance per die area AMD and its AIB partners decide to pass on to the end-users.

 

[AlistairAB]

You don't get 3x the transistor density, it doesn't work that way.

 

[InvalidError]

You don't get 3x the transistor density, it doesn't work that way.

Intel's 10nm in its original form had ~2.7X the density of its 14nm+++ so yes, ~3X for TSMC's 7nm vs AMD's other fabs-for-hire's less mature 12/14/16nm should be a reasonable approximation.

 

[bulldogg18]

Don't they normally do big reveals at Computex? I don't remember the last time AMD debuted something at E3, if you don't count consoles as an AMD release.

 

[Olle P]

July seems about right...
A planned presentation of Navi at CES (with launch in March/April?), but launch delayed a bit over three months due to some unforeseen problem.

... It's been a very, very long time since anyone's released a GPU that targeted the middle before going for the high end.

It's not that long since Polaris was released before Vega.

If I had to make a bet, I'd guess that Navi is 7nm Vega with a GDDR memory controller to keep prices down. ...

My bet is that Navi is a chiplet design like Zen2, and nothing like Vega.

 

[InvalidError]

My bet is that Navi is a chiplet design like Zen2, and nothing like Vega.

I'm going to bet one step behind that and say Navi may be a precursor to chiplet-style architecture, similar to CCX in Zen: a fundamental building block designed with future scalability in mind but stand-alone in its first implementation much like Zeppelin/Ryzen 1.

Given how much memory bandwidth GPUs require, having every chiplet go through an IO die to access VRAM doesn't sound practical, so I'm going to bet that "Navi II" will give each chiplet its own local memory and primarily use the IO die for PCIe and coordination between chiplets. Since AMD has full control over what happens on the GPU, having a NUMA architecture shouldn't be anywhere near as problematic as it was for TR1/2 and would make the chiplet's interface bandwidth requirements far more manageable.

 

[JamesSneed]

My bet is that Navi is a chiplet design like Zen2, and nothing like Vega.

I'm going to bet one step behind that and say Navi may be a precursor to chiplet-style architecture, similar to CCX in Zen: a fundamental building block designed with future scalability in mind but stand-alone in its first implementation much like Zeppelin/Ryzen 1.

Given how much memory bandwidth GPUs require, having every chiplet go through an IO die to access VRAM doesn't sound practical, so I'm going to bet that "Navi II" will give each chiplet its own local memory and primarily use the IO die for PCIe and coordination between chiplets. Since AMD has full control over what happens on the GPU, having a NUMA architecture shouldn't be anywhere near as problematic as it was for TR1/2 and would make the chiplet's interface bandwidth requirements far more manageable.

I lean this way as well. If they do make a more than one chiplet design(not sure they will) I expect it to be setup with infinity fabric in between the chips so minimal latency issues synching memory between the two chips but it would function mostly like a really fast Crossfire setup. So no propper shared memory but I would expect if you can do this with cheaper GDDR6 memory it would be feasible.

Not having shared memory in the professional space doesn't matter anyhow since most things people use profesional cards for go highly parallel.

 

[Olle P]

... If they do make a more than one chiplet design (not sure they will)...

With each chiplet (by rumor) having 20 CUs I'd say at least two chiplets are required for a mid-range GPU.

 

[JamesSneed]

... If they do make a more than one chiplet design (not sure they will)...

With each chiplet (by rumor) having 20 CUs I'd say at least two chiplets are required for a mid-range GPU.

This is going to be really interesting to follow.

 

[Olle P]

Given that the "20 CUs per chiplet" rumor is true I expect to see a quad chiplet Navi GPU released within a year from now, seriously outperforming the Vega VII at a lower price point...