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

Not comparing GDDR5 but I wouldn't be surprised if GDDR5 at the same performance level is more expensive.
Maybe you're right, but it's an entirely academic issue, because there's not a chip in existence large enough that would permit a set of GDDR5 buses wide enough to match the performance of four HBM stacks. It's enough of a struggle as it is to squeeze in the 512 bits of the R290 and GF 280 series cards.
 
Maybe you're right, but it's an entirely academic issue, because there's not a chip in existence large enough that would permit a set of GDDR5 buses wide enough to match the performance of four HBM stacks. It's enough of a struggle as it is to squeeze in the 512 bits of the R290 and GF 280 series cards.
Not only that: such a GDDR5 solution would provide an amount of memory that HBM1 can only dream of.
 
[strike]
Oh, I think I know what your implication was, but extra points there for being slippery.
I am being slippery when you try and twist the words in my mouth? Sorry, but that's just plain cheap.[/strike]
Nvm, this leads to nothing. I think people can read for themselves.

Maybe you're right, but it's an entirely academic issue, because there's not a chip in existence large enough that would permit a set of GDDR5 buses wide enough to match the performance of four HBM stacks. It's enough of a struggle as it is to squeeze in the 512 bits of the R290 and GF 280 series cards.
Actually, there has been since autumn 2013. It only has to be fitted with 8 gbps GDDR5-modules, which are listed in SK-Hynix's product catalogue. True, Hawaii's PHYs are not really designed for high speed memories in excess of maybe 7 GT/s, but in general a 512 bit inteface with 8gbps GDDR5 chips would be sufficient.
 
Last edited:
True, Hawaii's PHYs are not really designed for high speed memories in excess of maybe 7 GT/s
7 sounds rather generous to me. My card is slightly wonky even at factory-overclocked setting of 5.4GT (gives some artifacting in 3DMark Firestrike if one lets the card heat up), and trying to tune it even slightly higher just makes it much much worse. Maybe it's GPU-dependent, but current R290 series being rated at 5GT stock up to 7 is a very big step; how many cards have you met that function correctly up to 7GT? I can't imagine it's very many, unless AMD silently made a major revision to the chip somehow (my board was bought in the spring of last year.)

Could old Tahiti with its faster GDDR speeds and bigger PHYs do 7GT without liquid nitrogen cooling? I don't remember offhand and too lazy to google it up now. :p

but in general a 512 bit inteface with 8gbps GDDR5 chips would be sufficient.
You'd need to go beyond 8GT to hit 640GB/s...
 
Your assumptions are wrong.
GDDR5 with lower frequencies and more narrow memory buses might be somehow called mainstream, BUT the solution in the competition's top cards is anything but mainstream.

It may not be mainstream in terms of implementation or cost but it's certainly mainstream in performance in comparison to HBM.

If we drop the mainstream definition and just take it back to basics, what we have here is AMD managing to equal Nvidia performance (assuming that's actually the case) with a GPU that has around 50% more bandwidth and 50 watt more TDP to play with (more when you consider the memory element). And it's 3 months newer.

That's a little worrying IMO. What could Nvidia do with HBM + 300w with Maxwell? And the Maxwell architechture is 9 months old now compared to Fiji which is hot off the press.
 
640? Last time i checked rumor mill went back to spec'ced 512....
The memory standard supports up to 640 from what I understand, so that's the premise I was posting from. What the rumours du jour regarding ATI AMD Rage Radeon Fury claim doesn't really concern me. ;)
 
Please read up on the specs then. Preferrably, what SK-Hynix' product catalogue specifies for an HBM stack.
 
I understand that posters here have different brand and product preferences but to jump to conclusions about product performance from leaked 3D Mark score is a bit premature.
One could for instance judge AMDs chip from LuxMark performance perspective and laught other products off.

I rather wait for wider spectrum of tests and then judge how good or bad job AMD did.

Oh, and for me not only gaming performance matters as there is compute, video capabilities, power consumption in idle, playback and gaming modes and many more which I consider when buying new product.

If Fiji is amazing I will buy it, otherwise I will wait for 14/16nm products from both camps before upgrading.:LOL:
 
It may not be mainstream in terms of implementation or cost but it's certainly mainstream in performance in comparison to HBM.

If we drop the mainstream definition and just take it back to basics, what we have here is AMD managing to equal Nvidia performance (assuming that's actually the case) with a GPU that has around 50% more bandwidth and 50 watt more TDP to play with (more when you consider the memory element). And it's 3 months newer.

That's a little worrying IMO. What could Nvidia do with HBM + 300w with Maxwell? And the Maxwell architechture is 9 months old now compared to Fiji which is hot off the press.

Fiji itself is brand new but it appears to be a scaled-up Tonga (except for HBM, of course) and the latter was released in September 2014. At the time, it was already rumored to have been delayed a while due to inventory glut, which I'm inclined to believe since it still hasn't been released in fully enabled form. So Fiji is probably mostly based on IP that is about a year old.

If AMD chose to focus its micro-architectural work on the FinFET node and will introduce Maxwell-like improvements at that time, they'll probably be just fine. But if these are the fruits of arduous micro-architectural labor, then yes, they might be in serious trouble next year.
 
So Fiji is probably mostly based on IP that is about a year old.

I hope you're wrong on this - I'm still holding out for FL12_1 support. But I suspect you're right.

If AMD chose to focus its micro-architectural work on the FinFET node and will introduce Maxwell-like improvements at that time, they'll probably be just fine.

That's an optimistic view but I really hope it's correct.
 
If AMD chose to focus its micro-architectural work on the FinFET node and will introduce Maxwell-like improvements at that time, they'll probably be just fine. But if these are the fruits of arduous micro-architectural labor, then yes, they might be in serious trouble next year.

Yeah it's too early to jump to conclusions. We don't know if Fiji is a parting shot for the current GCN line or the start of something new.
 
Actually, there has been since autumn 2013. It only has to be fitted with 8 gbps GDDR5-modules, which are listed in SK-Hynix's product catalogue. True, Hawaii's PHYs are not really designed for high speed memories in excess of maybe 7 GT/s, but in general a 512 bit inteface with 8gbps GDDR5 chips would be sufficient.
An AMD PHY suited for 6 Gbps or higher would probably be Tahiti's, with the corresponding area and power penalties inherent to scaling it to 512 bits and significantly faster than Tahiti ever reached. I think the specs for 8 Gbps GDDR5 include a voltage bump. This assumes that a Tahiti-style interface can reach that speed. A new high-speed interface above Tahiti seems unlikely since signs point to AMD having given up on that direction some time ago.

A third more interface operating at two thirds higher speeds would give the bandwidth, but the GPU as a whole would have difficulty offering much more beyond that.


If AMD chose to focus its micro-architectural work on the FinFET node and will introduce Maxwell-like improvements at that time, they'll probably be just fine. But if these are the fruits of arduous micro-architectural labor, then yes, they might be in serious trouble next year.
AMD gave the range for their next-gen jump to FinFET, and it is line with the process transition alone. Perhaps if these are the first chance AMD has to tweak the architecture, it might be for 12_1, which does not clearly translate into a perf/W bullet point.
 
Last edited:
So Tweaktown reported today that Fury X will only see 30K units for 2015. My personal opinion is that, with a new technology like HBM, low volume due to various issues is exactly the kind of thing you'd expect. Especially when there are so many partners involved. Whether that's true or not, we'll probably never know, because it's impossible to judge whether a lack of retail availability is due high demand or low supply.

But luckily, ExtremeTech came to the rescue to at least put some HBM related rumors to rest:
One persistent rumor we’ve heard about HBM yields and costs, we can thankfully debunk. It’s been bandied about that AMD would face a crippling cost structure in the wake of the GTX 980 Ti’s debut a few weeks ago, due to poor yields or costs for High Bandwidth Memory (HBM). In reality, HBM has been on sale for quite some time, as evidenced in this memory databook from Hynix, which was released in Q3 2014. It shows two types of HBM as already available — a 128GB/s, 1.0Gbps component and a 128GB/s, 0.8Gbps component, both already available in the same 4Hi stack that AMD is using for Fiji.
Now, obviously we don’t know what kind of premium Hynix is charging for HBM, but the fact that the memory has been available for nine months now points to strong yields.
So there we have it: when a component has been published in a databook, that's sufficient proof to debunk claims that said component has been sold for that long, that yields are great, and that cost is fine.

Components get put in databooks to prod potential customers into using them in a future design. This needs to happen as early as possible so that designs can be ready as soon as the component is ready for mass production. The typical time to announce a component when first silicon has just come back from the fab and has shown early signs of life. That doesn't mean you will be able to get your hands on it (early samples are reserved to a select few), it doesn't mean yields are great, it doesn't even mean that all functionality will be there, it definitely doesn't mean that costs will fine.

Dear ExtremeTech editor, can you please refrain from spreading this kind of nonsense?
 
So Tweaktown reported today that Fury X will only see 30K units for 2015. My personal opinion is that, with a new technology like HBM, low volume due to various issues is exactly the kind of thing you'd expect. Especially when there are so many partners involved. Whether that's true or not, we'll probably never know, because it's impossible to judge whether a lack of retail availability is due high demand or low supply.

But luckily, ExtremeTech came to the rescue to at least put some HBM related rumors to rest:
So there we have it: when a component has been published in a databook, that's sufficient proof to debunk claims that said component has been sold for that long, that yields are great, and that cost is fine.

Components get put in databooks to prod potential customers into using them in a future design. This needs to happen as early as possible so that designs can be ready as soon as the component is ready for mass production. The typical time to announce a component when first silicon has just come back from the fab and has shown early signs of life. That doesn't mean you will be able to get your hands on it (early samples are reserved to a select few), it doesn't mean yields are great, it doesn't even mean that all functionality will be there, it definitely doesn't mean that costs will fine.

Dear ExtremeTech editor, can you please refrain from spreading this kind of nonsense?

Should go and read this ;) wccftech.com/amd-fury-30k-units-2015-air-liquid-cooling/
 
Yeah, it's ok to peddle in rumors, copy them, confirm them, debunk them by making a phone call whatever. And to mark them as such. I understand that they need clicks. Wccftech supposedly did what journalists should be doing: check their own sources. They didn't come up with BS reason and state it as fact. (Hard to believe that I'm taking sides with wccftech...)
 
So Tweaktown reported today that Fury X will only see 30K units for 2015. My personal opinion is that, with a new technology like HBM, low volume due to various issues is exactly the kind of thing you'd expect. Especially when there are so many partners involved. Whether that's true or not, we'll probably never know, because it's impossible to judge whether a lack of retail availability is due high demand or low supply.

But luckily, ExtremeTech came to the rescue to at least put some HBM related rumors to rest:
So there we have it: when a component has been published in a databook, that's sufficient proof to debunk claims that said component has been sold for that long, that yields are great, and that cost is fine.

Components get put in databooks to prod potential customers into using them in a future design. This needs to happen as early as possible so that designs can be ready as soon as the component is ready for mass production. The typical time to announce a component when first silicon has just come back from the fab and has shown early signs of life. That doesn't mean you will be able to get your hands on it (early samples are reserved to a select few), it doesn't mean yields are great, it doesn't even mean that all functionality will be there, it definitely doesn't mean that costs will fine.

Dear ExtremeTech editor, can you please refrain from spreading this kind of nonsense?
So extremetech is spreading nonsense but tweaktown is completely good on their zero source article.
 
Lots of 8K screenshots spread around by AMD. Are they trying to put the 4GB is not enough to the rest? Is Fiji a monster?
All of the above?

Fun times.
 
Back
Top