AMD Radeon RX 480 (14nm Polaris)

[Xlucine]

If this die shot is legit, it seems like even Polaris 10 isn't fully enabled as was rumoured a while back: http://forums.anandtech.com/showpost...0&postcount=40

It's the 380X wait all over again!

[watercooled]

If it's true, things like that are usually down to yield, but surely a chip of its size shouldn't be yielding so badly they can't release a fully enabled die? I guess with Tonga it might have been a case of not wanting to cannibalise sales of their Tahiti stock, perhaps they're doing that here too if they e.g. have a lot of Hawaii stock left?

[scaryjim]

Don't buy it, the layout is wrong. A GCN CU has 64 shaders, arranged in 4 blocks of 16. We know that Polaris is organised into 4 clusters of 9 CUs each, which the die shot clearly shows (the CUs are the 9 'vertical' rows). What's the other major feature of a GCN compute unit? Texturing. The blocks that have been marked up as "disabled" clusters are the only thing in that die shot that are in the right CU ratio to be texturing units.

Otherwise - if that die shot was labelled correctly - Polaris would be arranged in 36 CUs of 80 shaders each (5 clusters of 16 shaders). That simply isn't how GCN works.

Sorry, move along, nothing to see here. Full Polaris 10 is 2304 shaders.

EDIT: I've just read the full thread watercooled's link was from, and a couple of people on there have also pointed out the other clusters are almost certainly texturing.

[CAT-THE-FIFTH]

So an RX490 8GB looks like it is being launched - also the first comfirmation of the Sapphire RX480 Nitro 4GB.

[kalniel]

And GDDR5, so unlikely to be Vega. Wonder if this is the so called 'beast mode' version of the 480 rather than a new chip? Ie a binned, high voltage part only available from partners rather than a reference design?

[CAT-THE-FIFTH]

And GDDR5, so unlikely to be Vega. Wonder if this is the so called 'beast mode' version of the 480 rather than a new chip? Ie a binned, high voltage part only available from partners rather than a reference design?

Vega is meant to be two chips - so it could be a "small" Vega instead??

But it also looks like the GTX1060 is getting some competition in the form of improved RX480 4GB cards. I suspect if they are true 4GB cards,there might be improvements in performance/watt too.

[kalniel]

Vega is meant to be two chips - so it could be a "small" Vega instead??

Might not be small at all - probably a die shrunk hawaii. Seems all sorts of confusion around product strategy - is this pulled forward from initial placings of Vega in 2017?

[CAT-THE-FIFTH]

Might not be small at all - probably a die shrunk hawaii. Seems all sorts of confusion around product strategy - is this pulled forward from initial placings of Vega in 2017?

IIRC,there are meant to be two chips - Vega 10 and Vega 11.

A "small" Vega would be around 350MM2 IMHO,so a "large" one might be a more compute focussed chip for HPC.

[scaryjim]

Curious that it's listed as a "Radeon 490" rather than a "Radeon RX 490" - which would be the correct nomenclature according AMD's new scheme. According to the naming scheme the 490 is > 256bit and targets 4k gaming, so shouldn't be a "beast mode" 480.

Has anyone counted the memory controllers on the Polaris 10 die shot...?

Second thought - gemini card? RX 480 is a 11260, Radeon 490 is listed as 21260. 2x 4GB RX 480 on one board?

Anyway, doing a little snooping it appears that AMD "accidentally" listed the RX 490 as an eligible card for one of its promotions in the last few days, so there is definitely something coming. Hmmm.....

EDIT: also - Nitro RX 460 4GB

[kalniel]

It is very intriguing. 21260-00 is the part number for the 480, not 21260-01, but the screenshot is correct, I just had a look myself and see exactly the same thing.

[scaryjim]

It is very intriguing. 21260-00 is the part number for the 480, not 21260-01, but the screenshot is correct, I just had a look myself and see exactly the same thing.

Ah, yes, 11260 is the OC version. All very confusing.

[kalniel]

It's annoying - I need something just a little faster than the 480.. but don't want to have to leap all the way to the 1070 if I can help it. Partner cards are the first 'maybe'.. then failing that waiting for this supposed 490.. and failing that proper Vega in January. Unless nVidia actually pull out a nice and forward looking 1060 I guess. Yeah I doubt that too.

[watercooled]

Don't buy it, the layout is wrong. A GCN CU has 64 shaders, arranged in 4 blocks of 16. We know that Polaris is organised into 4 clusters of 9 CUs each, which the die shot clearly shows (the CUs are the 9 'vertical' rows). What's the other major feature of a GCN compute unit? Texturing. The blocks that have been marked up as "disabled" clusters are the only thing in that die shot that are in the right CU ratio to be texturing units.

Otherwise - if that die shot was labelled correctly - Polaris would be arranged in 36 CUs of 80 shaders each (5 clusters of 16 shaders). That simply isn't how GCN works.

Sorry, move along, nothing to see here. Full Polaris 10 is 2304 shaders.

EDIT: I've just read the full thread watercooled's link was from, and a couple of people on there have also pointed out the other clusters are almost certainly texturing.

Yeah now you point that out it makes a lot of sense - besides being a different colour those blocks have that extra chunk on the bottom so I agree they look like something other than unified shaders.

[CAT-THE-FIFTH]

It's annoying - I need something just a little faster than the 480.. but don't want to have to leap all the way to the 1070 if I can help it. Partner cards are the first 'maybe'.. then failing that waiting for this supposed 490.. and failing that proper Vega in January. Unless nVidia actually pull out a nice and forward looking 1060 I guess. Yeah I doubt that too.

The sad thing is my GTX960 is starting to increasingly fall behind the R9 380 and I had the same issue with my GTX660 against the HD7870.

But considering the reference RX480 is both power throttling and clock throttling,I suspect the AIB cards will be reasonably faster.

[watercooled]

But considering the reference RX480 is both power throttling and clock throttling,I suspect the AIB cards will be reasonably faster.

Only AMD are capable of doing that to a card launch. Again...

[CAT-THE-FIFTH]

Only AMD are capable of doing that to a card launch. Again...

It does look like AMD pushed up clockspeeds a bit at the last minute,probably due to the GTX1060 release being pulled forward.