AMD: Pirate Islands (R* 3** series) Speculation/Rumor Thread

If the source is Fudzilla, I would be sceptical as hell until there exists a more reliable site offering independent confirmation...
 
If the source is Fudzilla, I would be sceptical as hell until there exists a more reliable site offering independent confirmation...
The source would only matter if it reported that Fiji didn't support HBM2, IMO. That would be too unbelievable to be true.
 
The source would only matter if it reported that Fiji didn't support HBM2, IMO. That would be too unbelievable to be true.
I really don't see why Fiji should support HBM2, HBM1 allows the 8GB/4096bit now and by the time HBM2 is ready AMD should think about getting Fiji's succesor out already
 
It would be cheaper to build an interposer for the same quantity of memory using half the memory devices and to also get the same bandwidth, with an HBM2 configuration.

Is HBM1 versus HBM2 sort of a repeat of GDDR4 versus GDDR5? The former in each case being rejected pretty widely, at least partly because the latter arrived so soon? Wasn't there about 1 year between GDDR4 and GDDR5?

Is HBM1 a necessary evil on the way to HBM2? Was GDDR4 a necessary evil on the way to GDDR5? Or did both GDDR4 and HBM1 projects end-up being delayed, thus simply running in to their successors?
 
It would be cheaper to build an interposer for the same quantity of memory using half the memory devices and to also get the same bandwidth, with an HBM2 configuration.

Assuming that power consumption is (roughly) constant, it's a no-brainer. But is that a reasonable assumption?
 
Not sure what you're asking, since HBM2 is more power-efficient per pin per bit per second than HBM1.
 
Erm, without reading the article, how exactly does Fuad suggest the memory controllers could handle twice the speed per pin without changes?
And even though both GPU and HBM would be on an interposer, AMD might be able to support HBM2 with a "simple PCB revision."
Good stuff Fuad.
 
If AMD has compute/professional aspirations with Fiji (they should), HBM2 is a very good nice-to-have for increase memory capacity above 8GB. The increased BW is probably a second order bonus.
And, yes, HBM1 seems to be similar to GDDR4.
 
I don't see how it matters or why anyone would care. Even if AMDs upcoming flagship tripled the performance of a 290x
If AMD has compute/professional aspirations with Fiji (they should), HBM2 is a very good nice-to-have for increase memory capacity above 8GB. The increased BW is probably a second order bonus.
And, yes, HBM1 seems to be similar to GDDR4.

This would be the primary use. I can imagine a straight die shrink and trade up to HBM2 would more than bring enough benefits to release such as a successor to whatever upcoming flagship they have. Or just a straight up move to HBM2 for some sort of "Firepro" pro oriented card. A 16gb card, along with the large straight compute advantage the 390x is rumored to have over the Titan X, released late this year would fairly quickly grab a lot of the compute market.
 
No info apart from the pictures, but still can't be that far out now...
http://wccftech.com/xfx-radeon-r9-390-double-dissipation-allegedly-pictured/
XFX-R9-390-DD-2.jpg
 
PCB length seems to be between a R9 280 and a 270X, so somewhere between 260 and 280mm.
 
I got quite exactly 260mm with photoshop if I didn't do mistakes (PCIE connector length (the gold part) should be 85mm, in the picture it's about 48mm. The PCB on the picture is about 147mm, so it should end up at about 260mm)
 
My 290X's PCB is a lot longer than that (it's a custom version from MSI) so I thought they all were too.
 
Is HBM1 versus HBM2 sort of a repeat of GDDR4 versus GDDR5? The former in each case being rejected pretty widely, at least partly because the latter arrived so soon? Wasn't there about 1 year between GDDR4 and GDDR5?
The slide on HBM gen2's pseudo-channel mode makes note of a "legacy mode", which shows more interest in greater continuity between the generations. This did not seem to happen for GDDR4 and GDDR5. I recall there are GPUs with DDR3/GDDR5 controllers, which may show how much GDDR4 mattered once GDDR5 came out.
Since GDDR5 was a project started with AMD and Qimonda, it's unclear how much was coordinated with the efforts that were evolving from GDDR3 to GDDR4, but the transition seemed to be rather unceremonious for the latter.

Is HBM1 a necessary evil on the way to HBM2? Was GDDR4 a necessary evil on the way to GDDR5? Or did both GDDR4 and HBM1 projects end-up being delayed, thus simply running in to their successors?
GDDR4 had its life span cut short by GDDR5, as its next round of speed grades got preempted by the newer standard.

I think HBM1 and HBM2 are part of the same lineage, and HBM1 did not come out as early as was hoped. GDDR5 was not succeeded by a GDDR6, and it has hit speed grades that were not originally projected and lasted longer than most graphics memory standards did. HBM2 seems to make some effort to recognize what HBM1 does.
Another data point is that AMD and Hynix started talking about HBM in 2010, when the other memory types had 1-2 years between initial disclosures and production.

That GDDR5 was an AMD+memory maker project just as HBM1 is could have contributed somewhat to the delay on top of the obvious manufacturing changes. HBM2 might be the more polished version of the idea, and one that may be more palatable for DRAM makers and AMD's competitors. Qimonda's fate is not an auspicious one for how much benefit an AMD joint effort is, and HBM2 might have included more input and received more buy-in from outside stakeholders.
Lessened enthusiasm and the continued polishing of GDDR5 could have tamped down the upside of the first gen, so we may need to see who else but AMD may adopt it.
 
Back
Top