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

With the extra cost of HBM, and everything that surrounds it, and the expectation that it will be faster than a Titan X, it makes sense that AMD tries to get as much money out of it as they can. Especially if there is going to be limited supply.
And with no new faster silicon expected from Nvidia in the near future, it may even work.
 
1st gen HBM is limited to 4GB Graphics memory

These stacked memory packages have limitations, in the first generation you are looking at four stacks per package with two 128MB chunks in each layer, so that is 256MB per layer. Four times 256MB = 1024MB (or 1 Gigabyte) per accumulated DRAM stack/package. Currently chip designs allow for four stacks per IC (GPU). So that is 4 packages x (4x256MB) = 4096 MB. Hence the one limitation (if you can call it that) is HBA memory being limit at a maximum 4GB of graphics memory for the graphics card.

Now I am thinking broadly here, thus not just GPUs, think along the lines of APUs, processors and SoC here where HBM can make a substantial difference. This in effect is what HBA is all about. Please do keep in mind that this is 1st generation HBM, hence the limitation of 4GB graphics memory. Upcoming generations can and will have more layered stacks like 8, 16 or 32 - this all is possible, and thus along thee lines increasing the density and volume of DRAM is a viable future alternative as we move to smaller fabrication processes.

The primary factor to remember is that HBM achieves higher bandwidth with less power compared to DDR4 and GDDR5 thanks to the stacking of several memory chips. Whether or not the difference will be huge in terms of performance on this 1st generation remains to be seen. Will AMD be the only one using HBM ? No, Nvidia has already indicated they’ll use it for Pascal in 2016.
http://www.guru3d.com/articles_pages/an_introduction_to_hbm_high_bandwidth_memory,1.html
 
Sweclockers.com reports that AMD might be considering breaking Fiji out of the radeon 300 series and creating a limited, premium brand like the geforce titan series, presumably with a correspondingly titanic price tag as well.

That would be ludicrous if so. Especially considering every other card in the entire lineup would be a simple rebrand using older (in some cases much older) technology. At least Titan has the same feaure set as it's liitle brothers.

I really can't see AMD being insane enough to do that.
 
That would be ludicrous if so. Especially considering every other card in the entire lineup would be a simple rebrand using older (in some cases much older) technology. At least Titan has the same feaure set as it's liitle brothers.

I really can't see AMD being insane enough to do that.

Are you expecting Fiji to have less features than Hawaii?
 
The new MacBook Pro with retina display just launched today with the R9 M370X, with support for an external 5K display at 60Hz.

Does this breath fire to the old rumour that R9 MX370X is based on Litho XT and not Oland? Well, could just be Cape Verde ...
 
Last edited:
Ugh. This video reminds me why I skipped most courses in college and studied from course notes instead.
Is there a place where they have just the slides?

Edit: never mind, the guru3d article has them, "para-dime" and all. And soon the rest of the web as well.
 
Are you expecting Fiji to have less features than Hawaii?

The opposite, it will certainly be more advanced. And seperating your most advanced (and only new) GPU out from the rest of the "new" family of GPU's which is composed entirely of rebranded "old" chips sounds like suicide to me. What would even be the point of a 3xx series from AMD if every single GPU in that series is a re-branded 2xx series GPU with a slight overclock?
 
The new MacBook Pro with retina display just launched today with the R9 M370X, with support for an external 5K display at 60Hz.

Does this breath fire to the old rumour that R9 MX370X is based on Litho XT and not Oland? Well, could just be Cape Verde ...
Note that it's 5120 x 2160, not 2880. So it's not quite support for 5K, it's support for a 21:9 stretch 4K monitor, like what LG has been doing.
 
AMD Fuji (Single GPU) has only 4 GB Vram (Dual GPU version has 8 GB (4 GB x 2)).

http://www.anandtech.com/show/9266/amd-hbm-deep-dive

Expected cost is $849 for single GPU version with 4 GB.

http://www.fudzilla.com/news/graphics/37790-amd-fiji-aims-at-849-retail-price

If true then things arent looking good on the Fiji front. Not unless performance is insanely high anyway (as in easily beating the Titan X). Nvidia will be in a position to launch a 6GB 980Ti clocked higher than the Tian X and priced however they want. If Fiji is "only" the same speed as the 980Ti, then if Nvidia prices it the same and has 50% more memory, AMD won't be in a good position at all.
 
It is interesting to see UMC is the fab for the interposer. Is this part the result of work the consortium that includes IME? ASTAR is listed in one of the AMD slides. (late edit: never mind, perhaps I squinted too much on the ASE line?)
I'm curious how the interposer is implemented physically. The slide presents a package below 70mm on both dimensions, so does that imply more expensive methods to produce 65nm features at such a large size, <<70mm dimensions, or multiple subsections? The DRAMs don't need to communicate, so could the interposer be internally subdivided into more manageable halves or quadrants for the 65nm features?

I guess the next question is where to from here. At least technically, the tech seems a decent enough start aside from the capacity question, which has swung the discussion back down from the earlier speculation of an intermediate capacity expansion scheme.
Is there margin to bump up the number of stacks? Will the pitch be made finer in a later iteration?

Perhaps when the actual product is announced, more details on what can be done with that bandwidth will come out.
 
Last edited:
AMD Fuji (Single GPU) has only 4 GB Vram (Dual GPU version has 8 GB (4 GB x 2)).

That's not the message I get from Anandtech's article at all:

At the same time this also calls into question memory capacity – 4 1GB stacks is only 4GB of VRAM – though AMD seems to be saving that matter for the final product introduction later this quarter. Launching a new, high-end GPU with 4GB could be a big problem for AMD, but we'll see just what they have up their sleeves in due time.

Ryan doesn't even hide the fact that AMD does have something up their sleeves. Otherwise, he'd write "we'll see if they have something".
Maybe I'm reading too much into it, but he's around to put me in my place if that's the case :)
 
4GB is going to hurt.

http://techreport.com/review/28294/amd-high-bandwidth-memory-explained/2

When I asked Macri about this issue, he expressed confidence in AMD's ability to work around this capacity constraint. In fact, he said that current GPUs aren't terribly efficient with their memory capacity simply because GDDR5's architecture required ever-larger memory capacities in order to extract more bandwidth. As a result, AMD "never bothered to put a single engineer on using frame buffer memory better," because memory capacities kept growing. Essentially, that capacity was free, while engineers were not. Macri classified the utilization of memory capacity in current Radeon operation as "exceedingly poor" and said the "amount of data that gets touched sitting in there is embarrassing."

Strong words, indeed.

With HBM, he said, "we threw a couple of engineers at that problem," which will be addressed solely via the operating system and Radeon driver software. "We're not asking anybody to change their games."
So, erm, AMD's promising it won't hurt. Driver technology for the win :rolleyes:
 
It might not have been disappointing if AMD had launched it in dec-jan against the 980 4GB, it doesn't look good now.
As for driver technology, turbo-cache FTW!:runaway:
 
The 4x 1GB stacks are used for illustration purposes for all I can see. 1GB/1024bit was originally limitation on 1st gen HBM, but the latest HBM slides clearly show they allow 2GB/1024bit too
TechReport also mentioned in their most recent podcast, as straight as it can be said, that Fiji will be 8GB (too)
 
Wow so 4GB confirmed. Even if it's true that it won't matter, it's still going to look bad against a 6GB 980Ti. Especially if the 980Ti is faster, but that remains to be seen.
 
Back
Top