AMD: Navi Speculation, Rumours and Discussion [2019-2020]

Not different.

 

wow, doesn't this forum have its priorities messed up.

 

Really, instead of technical discussion we see a lot PR people making advertising. And no, I am not referring only to the banned user - nor at his "side".

 

Navi - Leadership Performance - In"Dr Lisa Su" Class.

But anyway, as suggested, best not read too much into it.

 

With HBM3 ramping up I would think HBM2(e) has reached it's peak price point and on it's way down in price, if not cheaper than GDDR6(x).

Can we see a scenario where there is a $499+ dGPU with HBM (in early 2021)?

 

16GB 256-bit card?:

re-post via:

https://videocardz.com/newz/alleged-amd-radeon-rx-6000-engineering-sample-spotted

 

Jesus, that pc is dirty.

 

I hope it doesn't come with that cooler lol

 

Why the CPU cooler?

 

that's so dirty. Can't discern how many PCIe pins it needs, it could give us a pretty good idea of the power consumption to some extent.

 

This looks like complete BS.

>MCM
Not likely from linux drivers pointing to a single unit. All signs show AMD will probably not make the move to multi die design this gen but at a later gen. From the XSX design, nothing seem to point to moving in this direction.

>3x L1 cache
Very unlikely since increasing a cache by 3x in size will generally kill latency most of the time. Also why would the L1 need to be 3x the size. From the XSX presentation it seem that L2 sizes per compute units are pretty much inline with Navi 10, no mention of L1 size. Would be a strange detail to leave out if this was a thing.

>2x ROPs
64 ROPs are assumed from Linux drivers. Which is the same as Navi10. AMD probably thinks 64ROPs are good enough for 4k. With higher clocks, they probably do have enough ROP performance with only 64. Although more could be a possibility and wouldn't be too surprising.

>7x performance in ML uses
From what? It's not like it has tensor cores. Could be int4 but thats very specific cases and would probably only be in inference. Also since RDNA is supposed to focus on gaming, it will be unlikely to really focus on ML performance improvements since CDNA is now a thing. Claiming 7x improvement as average of multiple use cases isn't likely at all even if the thing has a lot better performance.

>HBM
Not likely, also it shows 2 blocks? 2 stacks? which wouldn't be enough bandwidth.


Also why would anyone at AMD send anyone a picture like this? Both terrible as a technical document and looks terrible as marketing material too.

 

^Agreed. Thought I'd share it anyway.

 

At the median 2.1x 5700xt performance, that would put it 45% faster than a 2080ti.

 

Gave me a good chukle ngl.

 

Because it's a bring-up, not full production sample. You don't know actual power consumptio, actual scalability on the v/f curve or anything really. So you need to be prepared for the worst.

 

Well, "Navi 21" with 256-bit bus only? It seems suspicious to me.

 

@rogame tweeted some time ago that there were four variants of Navi21, XTX. XT, XL, XE. XE suffix was used years ago for cards with severely cut-down memory bus.

 

Midrange GPUs (Polaris, Navi 10, etc.) are also using the same XT, XL, ... nomenclature.

 

Hmm.. no, the suffix was "SE". But it should be the same type of card.