News - AMD Tonga expected in R9 280 and R9 280X refresh next month

[HEXUS]

So say "reliable sources," to Chinese VR Zone. But will it be a 20nm part?

Read more.

[abychristy]

Oh man if you had said this 2 months back I could have got this. But hey I am already saving for an i7 so no gpu for me till 2 years.

[The Hand]

Moving to 20nm is an interesting prospect. It implies AMD maybe ahead of Nvidia on that front at least.

[Jowsey]

Hopefully it is a 280x replacement part.

[Denis_iii]

Hope they support freesync, mantle and AMD's audio thang.

[DanceswithUnix]

20nm is supposed to be a while off yet.

http://www.fudzilla.com/home/item/34...hics-this-year

[Xlucine]

I hope this encourages nvidia to hurry up the 28nm maxwells

[kompukare]

20nm is supposed to be a while off yet.

http://www.fudzilla.com/home/item/34...hics-this-year

Indeed. And furthermore, the cost/transistor of 20nm is not expected to be much below 28nm. So while, a small 20nm GPU might have better perf/watt, any large 20nm will most likely be rather expensive. So while a theoretical 400mm² GPU at 20nm would cramp close to twice the transistors (say about 8 billion transistors), if each wafer costs, for example, x1.7 the cost of a 28nm wafer (plus initial yields will be worse than current 28nm) then such a 8 billion transistor GPU could easily cost twice as much as a current R9-290.

Also, all the hype about GM107 (GTX750Ti) misses one important factor: GM107 vs GK107 the die size went up from 118mm² to 148mm². That is +25%. So while GM107 is more efficient that did not come out of nowhere. So hypothetical GM104 on 28nm might well be 370mm² while a hypothetical 28nm GM110 would an unmanufacturable 700mm² or more. Point being, while GM107 is an interesting design, it is not something Nvidia are likely to be able to apply to bigger dies on 28nm. Remember, GPUs are very parallel so one way to solve power consumption problems is to run wide and slow vs narrow and fast (Nvidia already dropped their clock-doubled parts of their design going from Ferni to Kepler as those parts were drawing a lot of power).

[Terbinator]

I hope this encourages nvidia to hurry up the 28nm maxwells

28nm Maxwell already is out.

[watercooled]

Moving to 20nm is an interesting prospect. It implies AMD maybe ahead of Nvidia on that front at least.

AMD have been first to every node so far, as far as I can remember, followed by Nvidia a few months later complaining about yield and blaming everyone but themselves. If the noise about TSMC 20nm is true though, this time might be more about which company thinks they can make use of a node which seems more geared to mobile parts; it's not necessarily a no-brainer to switch like it's been historically.

As kompukare says though, people are assuming the same 750Ti treatment could be applied to the bigger GPUs, but that seems unlikely, both technically and economically. As well as the die size issues (IIRC, Nvidia aren't far off the TSMC reticle size @ 28nm anyway, but don't quote me on that), a lot of the fat-trimming applied to the small 107 part may not be applicable to the higher performance parts. E.g: http://forums.hexus.net/cpus/241925-...ml#post3205597

Plus, it kind of defeats the object that people are going off the stock 750Ti numbers when looking at efficiency, but a great deal of cards you can actually buy are pre-overclocked, harming said efficiency, and quite considerably in some cases. At the end of the day, the 750Ti is frankly not brilliant value, and is being strongly hyped due to this power efficiency.

[HalloweenJack]

from what I`ve read - 20nm in BULK is in production NOW @ TSMC but , tonga is from GloFO on 28nm

[Dottorrent]

The 28nm process is still good for now. When Pirate Islands comes by, it should use the 20nm process, hopefully giving a much needed jolt of power over Nvidia.

[watercooled]

from what I`ve read - 20nm in BULK is in production NOW @ TSMC but , tonga is from GloFO on 28nm

I've read a few rumours about GPU production at GloFo, which could prove interesting. I wonder if Tonga will be a sort of pipecleaner part to get the ball rolling at the fab?

A lot of the work for porting GCN to GloFo is already done, in theory; they're already producing the GCN GPU on Kaveri.

[Xlucine]

Indeed. And furthermore, the cost/transistor of 20nm is not expected to be much below 28nm. So while, a small 20nm GPU might have better perf/watt, any large 20nm will most likely be rather expensive. So while a theoretical 400mm² GPU at 20nm would cramp close to twice the transistors (say about 8 billion transistors), if each wafer costs, for example, x1.7 the cost of a 28nm wafer (plus initial yields will be worse than current 28nm) then such a 8 billion transistor GPU could easily cost twice as much as a current R9-290.

Also, all the hype about GM107 (GTX750Ti) misses one important factor: GM107 vs GK107 the die size went up from 118mm² to 148mm². That is +25%. So while GM107 is more efficient that did not come out of nowhere. So hypothetical GM104 on 28nm might well be 370mm² while a hypothetical 28nm GM110 would an unmanufacturable 700mm² or more. Point being, while GM107 is an interesting design, it is not something Nvidia are likely to be able to apply to bigger dies on 28nm. Remember, GPUs are very parallel so one way to solve power consumption problems is to run wide and slow vs narrow and fast (Nvidia already dropped their clock-doubled parts of their design going from Ferni to Kepler as those parts were drawing a lot of power).

And if you look at the 750ti Vs the 650ti, the die has shrunk by 50%. Previously that section of the market was split between GK107 and GK106, now GM107 is higher in the lineup than GK107 was so it's a bigger part - this is not something that matters at all, especially with the large gulf in performance between the 650 and 750.

Also, the die size per processor (by the rough metric of dividing die area by processor count) has dropped by 30% in GM107 compared to GK107 - 0.31mm^2 for GK107 Vs 0.23mm^2 for GM107. This occurred at the same time as an 8-fold increase in cache and an increase in texture units, so reductions in the number of ancillary parts doesn't explain it.

28nm Maxwell already is out.

Not at the level I would like - I want cards that can compete with AMD's in terms of performance as well as energy efficiency.

Plus, it kind of defeats the object that people are going off the stock 750Ti numbers when looking at efficiency, but a great deal of cards you can actually buy are pre-overclocked, harming said efficiency, and quite considerably in some cases. At the end of the day, the 750Ti is frankly not brilliant value, and is being strongly hyped due to this power efficiency.

The massive overclock on the palit model only adds 16w, which is still comfortably below equivalent AMD parts

[sgtfluffybutt]

If it keeps the same performance or better this will be great

[watercooled]

And if you look at the 750ti Vs the 650ti, the die has shrunk by 50%. Previously that section of the market was split between GK107 and GK106, now GM107 is higher in the lineup than GK107 was so it's a bigger part - this is not something that matters at all, especially with the large gulf in performance between the 650 and 750.

Comparing 750Ti to 650Ti isn't useful when comparing die size - the 750Ti is fully enabled, while the 650Ti was both the least-enabled and lowest-clocked GK106 part.

Also, the die size per processor (by the rough metric of dividing die area by processor count) has dropped by 30% in GM107 compared to GK107 - 0.31mm^2 for GK107 Vs 0.23mm^2 for GM107. This occurred at the same time as an 8-fold increase in cache and an increase in texture units, so reductions in the number of ancillary parts doesn't explain it.

GK106 - by a strange co-incidence - 0.23mm^2. As die size increases, the ratio of core:uncore changes. There's no magic. The transistor density is also fairly similar between the dies, if that was what you had in mind; they're still using the same node at the end of the day. (GM107 is slightly higher density, but it also has a higher ratio of cache on board, which is inherently higher density then core logic anyway)

The massive overclock on the palit model only adds 16w, which is still comfortably below equivalent AMD parts

I was being slightly ironic in a roundabout way with that line. My point wasn't that it makes the card hugely inefficient, merely that, as I said, people are making out how the efficiency is of such huge importance, then going out and buying a pre-OC card which adds little to performance, but reduces this "crucial" efficiency number...
Add another 26W, which is fairly negligible in real terms, and get a 265; a considerably faster and cheaper card. Considering the price difference at launch, we worked out it would take a few years of fairly heavy gaming use, and with UK electricity prices, for the 750Ti to work out cheaper in terms of TCO.