AMD: Navi Speculation, Rumours and Discussion [2017-2018]

As the paper is exascale, those would be the most efficient clocks. Those APUs may be using less energy than a single Vega at stock settings.

 

NAVI rumored to be launching in Q3 2018, https://www.tweaktown.com/news/59428/amds-next-gen-navi-gpu-launching-august-2018/index.html
"Once again I have an exclusive story that AMD will have Navi ready to go sometime in July-August 2018"

IMO, this early date relies on a perfect & on-time 7nm launch by GF. We'll see...

 

Not to mention it heavily conflicts with the professional oriented Vega 20 schedule

 

I for one think July-August sounds way too soon for Navi. Unless there's been a huge breaktrough, GF's 7nm DUV will probably be at risk production at that time.
It could make some sense to release high-margin cards (like such as FirePro Vega 20) during risk production, but not so much with Navi that's supposedly replacing Polaris and Vega on the consumer side.

The leaked/rumored roadmaps already put Vega 20 on 7nm during H2 2018. It also put Vega 20 alone in the double-precision market even after Navi's release.
Only story here is consumer Navi coming up half a year ahead of the Pro solutions.

Moreover, perhaps it doesn't make much sense for Navi's chips/chiplets to have 1:2 FP64 throughput, and Vega 20 will stand alone in that market for a long-ish time. Like Hawaii that has been AMD's only solution for DP workloads since 2014.

 

If ever? Shots were fired...

There's nothing stoping AMD from eventually realeasing 'something' called Navi. Except bankruptcy and being bought out of course

 

The TweakTown claims specifically Pro-Navi, not Navi in general or consumer-Navi

 

I doubt PS5 and XB3 possible contracts are about anything except Navi (except they surprise as going with Nvidia, something quite logical from a perf/watt point of view). I think, at least for these contracts, Navi IP will be out somehow. What seems clear is that AMD wont make any efforts porting the dissapointing Vega architecture to 7nm.

 

Navi for XB3 (surely you mean 4? OG XB; 360; Xbone/XboneX; Nextbox) or PS5, yeah, maybe. There's a lot of lead time in console designs, but PSPro and XB1X should last a while, and by then vega would be old hat by then.

I don't think NV will be a viable candidate tho, as they don't have a x86 license. They can only offer ARM cores for integration, and while fast perhaps, it's a different ecosystem. x86 (x64 really, these days...) is familiar territory for devs, and current consoles are similar to PCs, which makes porting less of a (costly) chore. I don't think many would appreciate moving away from that setup.

 

nVidia needs to learn how to make non-abortion-tier SoCs first though.

 

There was a rumor next PS console would have discrete CPU and GPU. In fact whis way is where things are headed, as next GPUs seem to be multi chip architectures. If that track is really considered, then would be quite feasible a Zen CPU+Nvidia GPU. As we have seen with Nintendo Switch Nvidia is willing to be inside a console, now maybe they want another win to make their architecture more multiplatform friendly with console based developments.

The rumor:

https://www.tweaktown.com/news/57845/playstation-5-feature-full-discrete-gpu-apu/index.html

 

A key part of the MCM concept is the interconnect between the individual chips on the package. AMD and Nvidia have different solutions for this. One of them would have to create a one-off implementation of the other's interface for this to work and there is an engineering cost associated with this. Or they could just go all AMD and use the Infinity Fabric interface that already exists in both their CPUs and GPUs. My money would be on the latter.

 

Doubtful. There's significant cost and performance benefits with single-chip scenario, say those in the know. External buses between CPU and GPU would make GPGPU transactions very costly latency-wise, and single-chip can also feature single memory controller with unified memory, which eases development headaches considerably (as witnessed during PS360 generation.)

Also, mixing different makes of CPU and GPU would require reinventing the wheel whereas communications interface is concerned, seeing as AMD is focusing on infinity fabric, and NV is not (edit: as pointed out by Corbo just above.)

 

Well nowadays, compiling your game for x86 or ARM (by far the biggest game market, ie smartphones) is not a big deal. Moreover, SWITCH popularity means that more and more devs are now very familiar with ARM/CUDA/Nvidia ecosystem.
At the end, it's a matter of how much resource Nvidia want to put into a project and what price they want to offer (green team loves their high margin). Because in terms of technology, they have nothing to fear from AMD...

 

I agree with the first, but no so much with the second. Some of AMD's engineering choices that don't appear to yield more performant PC GPUs, are very much more applicable and beneficial in a console GPU.

 

I just came with a crazy prediction for Navi: the usage of MCM, Inifinity Fabric and Interposer to make Navi perform head-to-head on all metrics, including efficiency, with Nvidia 7nm offerings.

Take Polaris 10(2304 SPs) and Upgrade it at Vega level of features(DX12.1, HBCC, Primitive Shaders, NCU and etc). Now you do some treatment to improve the per CU performance for a new level. Then you chose the SoC version of GF 7nm(Power/Frequency tables for both 7LP flavors here: http://btbmarketing.com/iedm/docs/29...ha_Fig%202.jpg). The right choice would be the SHP choice for such a monolhitic GPU, but that's not "scalability".

You shrink four of this evolved Polaris 10 GPUs, use infinity fabric and MCM to tie it all togheter, like EPYC. Then you have a total of 9216 SPs. Then you use Interposer to match the GPU with 4096-bit(a couple of stacks) 16GB or 32GB of HBM3("NexGen Memory"). Then there it is, with an total Area not much bigger Than Vega64 with its HBM2, ready to take on Nvidia's best offering on TSMC 7nm.


Crazy idea, but it looks plausible, no? I Just wish Navi is true GCN2 and not only GCN1.6. I'm enough of minor incremental updates.

 

Isn't it either MCM or interposer but not both?

I've been waiting for that for a long time, I'm not holding my breath. However this time is different as they don't really have a choice. If they don't improve perf/GFLOP in Navi they will no longer be anywhere near nvidia and won't be able to compete.

 

Given there is so far not a single patent nor a pilot leak of an upcoming AMD's "MCM" GPU, it seems more probable to be just another boring GCN stuff.

battle proven 4096SPs
NNCUs topping at 2GHz
4SEs should be enough for anyone
IPC ~Fiji, because why not
TDP top 500W with a hitech cooler
a dozen of epic new software features which would never be enabled
perf target GTX 1080Ti

Scalability - Raja has already touted the scalability of Vega's TDP. This would be even more scalable! 200-500W
NextGen Mem - Imagine a HBM2 version with lowered voltages and call it HBM2+ or HBM3. It's simple.

 

Why would they go multi-chip if its the same amount of SP's that current chips can handle?

Then how will they improve geometry processing? They'd have to change the way GCN is architected.

Thats pretty salty.