Man of Honour
Fiji Radeon 390X comes with 8GB frame buffer
http://www.fudzilla.com/news/graphics/37258-fiji-radeon-390x-comes-with-8gb
drunkenmaster's view on the article
http://www.fudzilla.com/news/graphics/37258-fiji-radeon-390x-comes-with-8gb
drunkenmaster's view on the article
Article is bogus for many reasons, could the 390x have 8GB or an 8GB version, very possible I've said as much for a long time. But could it happen how this article says, no, it's impossible, Faud is an idiot and anyone believing this has a basic lack of knowledge or logical thinking.
It's this simple, a card will take at least 12 months to make, newer architectures/bigger new cards generally take a lot longer with the derivatives (same architecture but different shader/rop/bus counts) being quicker but still taking a decent length of time to design, lay out, tape out and get to market, mostly concurrent work but it takes time regardless.
You can not, with absolute certainty, go 2-3 months from launch and double the memory bus size... this is a literal impossibility. The article is twaddle because it's proposing something that is literally impossible. You can't take a core designed for four external connections to memory chips and just add four more without redesigning the entire chip, relaying the chip out on a transistor by transistor basis, re-taping it out, etc. This would add die size and at minimum a 12 month delay for a huge cost. This is NOT possible.
What is possible is it was always going to be an 8GB chip with 8 memory connections, what is also possible is the wait till summer is because that is when 2GB stacks will be available, letting them offer 4/8GB versions both using 4 memory stacks, just different density stacks.
On a more sensible note, 6 stacks would cost a lot less than 8, give you 6GB at a lower cost to the user yet adding 50% more memory for 4k users which will be more than enough in 99% of games for the next few years.
There are many possible specs for the 390x, what is impossible is what Faud is pushing in this article, that because Nvidia has semi announced a 12GB card that AMD at the last second reworked the card to completely change it.
Not true for power reasons, if you had a 300W Titan X with 12GB gddr5 and one with 8GB HBM, there would be likely around the 30-50W extra power available for the core on the HBM model. So the core could have a higher default clock, a much higher boost clock AND it would have the bandwidth required to feed the core running faster.
HBM's advantage isn't the raw bandwidth, it's the power saving it offers, which can then be used for core clock speeds.
Will the 390x have more room for power usage for the core compared to Titan X because of HBM, absolutely, will it be faster... depends on die size, actual performance, architecture improvements and how much money both dumped into another 28nm design. Effectively everything is somewhat compromised at this point. THe cards being released in the past year, likely including the 290x, were not what Nvidia or AMD had planned for 3 years ago. When you plan an architecture for the next node you can plan to double transistor count, some hardware feature that just simply cost too much die space on 28nm might have made sense at 20/14nm, so was included in the next architecture. When you try and move back up to 28nm with an architecture designed for a lower node, you make compromises.
HBM will allow any individual core to be noticeably faster than the same core equipped with GDDR5, the 390x will be significantly faster with HBM than had it used GDDR5.... but that doesn't translate to it being a brilliant core automatically. Depends on many more things.
Last edited: