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

Spoiler

So this looks kind of like a gpu from the back, hint at it being a long card? Not sure what the white specks are.

 

No way is the difference in price for 8 GB going to be just $50. DRAM contract prices seem to be hovering in the range of $3/GB so assuming a bit higher for GDDR I'd guess 8 GB is closer to $40. I'd expect the retail price to be a fair bit higher (at least double). And I suspect these "leaked" prices are for N22 not N21. Guessing N22 would be 256 bit GDDR6 at 16 Gbps.

If N21 really is what the rumours/XSX numbers point towards, its going to be priced around the 3080 level.

When is MI100 set to be released btw?

Im curious, is G6X an Nvidia exclusive? or can AMD use it for next gen?

 

Binned one yeah.

It's early-mid Q4 availability in your fairly generic servers from your big 3 vendor of choice so soon.
Actual AMD launch event is ????? for both it and Milan.

Actually yeah I wonder when the fuck are they going to talk Milan given it already early shipped to hyperscalers.

Seems so.
It's like MCDRAM for the most part.

Maybe.
It's not a JEDEC standard so it's all up in the air.

 

you please clarify about fmax? with w/fmax you mean at max clock limit for Navi10 for example? not sure if it was 2100Mhz. If so and navi 21 fmax is say 2400Mhz, thats where navi 21 has 50% better P/W then? In that case, P/W of cards based on navi 21 will depend on how close their final clock is not to fmax but to voltage and consumption sweet spot anyway. I would say for navi 10 sweet spot is around 1750-1800Mhz...but its clear that if Sony clocks PS5 at 2.23Ghz, that should be near sweetspot and far from fmax.

 

They're using hotter, less efficient clocked closer to fmax boards for perf/W comparison so no one can call bullshit on them.

They won't tell us yet.

Kinda.

 

Another piece of random info from igorsLAB. The video is in German and my German is terrible... Navi info starts towards the end.

• Big Navi at 275W is faster than RTX 3070, possibly with over 300W it should to be in RTX 3080 territory.
• AIBs still haven't been sent the Big Navi BOM, so launch is AMD-only.

 

wow they've gained 40%-ish perf by dialing 30watts in, wahoo.
(another AIB goon will be sacked)

 

Faster than RTX 3070 != as fast as RTX 3070
over 300W != exactly 305W

 

Yes, N22 is indeed faster.

No, reference N21 board is hardcapped.
You'll have to wait for AIBs to do a housefire race with nVidia.

 

People want AMD news, but there is no AMD news.
Anything now goes.

I have some very nice screencaps from his patreon-only discord that show how fundamentally dumb he is.
It's funny.

 

Given that they basically introduced Rome 6-9 months before it shipped, the Milan announcement/reveal certainly seems late. But they obviously don't want to Osborne themselves either. I expect we'll hear about it soon enough though.

Would be nice to see it on the next mid-range parts for Navi 3. Let's hope.

 

I don't think they can since those two share the same fab capacity and if anything, Milan would acc mss gains.
It's a nicer all-around product that requires way less validation than Rome (or god forbid Naples) ever did.

Pretty weird that big3 brief wasn't leaked into the whole wide internets.

G6X has some fairly anal PCB design requirements so idk about sticking it into any mainstream parts.

 

Geez that teaser is cringe material..

Either announce the cards or announce the day you're going to announce the cards, AMD.
Fortnite players don't even need new cards.

 

Hopefully that's because AMD is spending more money on engineering than marketing.

 

Argh, ok on Navi GPU with two stacks of HBM was confirmed by linux commits. Maximum bandwidth, assuming the flops-bandwidth ratio stays the same as the Series X (and why wouldn't it) the HBM version could also be "Big Navi", as it'd support up to 24 teraflops with Samsung's 4.2gps spec, 20 at SK Hynix's spec. They could dual source as well.

Still costs more than GDDR6, but for whatever reason they've got a GPU with it anyway, so it's possible.

 

Hm... 16GB for RX Navi 21 and 32GB for Navi 21 Pro (for Apple)?
Side note: I assume Apple will stick to RDNA until CDNA launches?

I mean, if the performance is there (competes/beats RTX 3090), AMD may as well just go with 32GB for RX Navi 21 and charge $999-1200. That's a big if though.

 

I'd expect the 32gb version to bear some ridiculous price if its trye, that's expensive.

PS: The idea that "Big Navi" would "only" be competitive with a 3070 is silly. A 5700xt clocked as high as a PS5, which RDNA2 can obviously do since it is doing it, with 18gbps would be competitive with a 3070. The realistic performance range for Big Navi is definitely somewhere between slightly below a 3080 to somewhere above a 3090.

 

If ps5 is competing with a 3070 they must be doing magic.

 

Thanks to PS5 specs, all points to rdna 2 could carry some impressive clocks, biggest jump since pascal. This GCN-RDNA-RDNA2 reminds 2008-2009 era when amd "simplified" R600 arquitecture pipeline for better clocks and gaming performance with HD 4870 and HD 5870. On the other side and apart from the node, nvidias current gpu seems not so simple or clock friendly as 680, maxwell and pascal were.