AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread

Titan gets around 4550-4600 in the graphics score going on a quick check of 4 or 5 reviews, this is 4511 for the 290X. There's not much in it anyway, gotta wonder how easily AMD would have had it beaten with a similar sized die. Instead they've let them off the hook, again.
 
Titan gets around 4550-4600 in the graphics score going on a quick check of 4 or 5 reviews, this is 4511 for the 290X. There's not much in it anyway, gotta wonder how easily AMD would have had it beaten with a similar sized die. Instead they've let them off the hook, again.

Is there really much point in adding CUs when you're already power-constrained?

Besides, FireStrike is just one benchmark.
 
NVIDIA themselves could probably at least match the Titan with a 4xx mm^2 Kepler die, if they clock it higher within the same power constraints (which should be doable given fewer SMXs) and don't disable any SMXs (unlike Titan/780).
 
Is there really much point in adding CUs when you're already power-constrained?

Besides, FireStrike is just one benchmark.
More units don't necessarily mean more power. If that were the case, GPUs would be few-core and run at a high frequency.

Really, at least at the frequency ranges GPUs operate at, more CUs = less power at a given performance level.

I'm really confused why you're asserting the opposite, when the most obvious high level change in Kepler utilized this rule -- more cores at a lower frequency resulted in lower power usage. The tradeoff is higher cost.
 
More units don't necessarily mean more power. If that were the case, GPUs would be few-core and run at a high frequency.

Really, at least at the frequency ranges GPUs operate at, more CUs = less power at a given performance level.

I'm really confused why you're asserting the opposite, when the most obvious high level change in Kepler utilized this rule -- more cores at a lower frequency resulted in lower power usage. The tradeoff is higher cost.

Right, but Hawaii is supposedly a 44CU chip, probably at 250W. Adding CUs would mean reducing clock speeds, which:

a) would result in a sub-linear increase in performance, especially if performance doesn't scale very well with CU count,
b) would increase static power (because of the bigger chip) unless voltage could be reduced by a sufficient margin,
c) would reduce the performance of the front-end unless it were widened as well, and might, either way, produce a poorly balanced GPU,
d) would increase cost, obviously, and possibly time to market.

Considering all of the above, I'm really not convinced it would have been worth it.
 
Last edited by a moderator:
c) would reduce the performance of the front-end unless it were widened as well, and might, either way, produce a poorly balanced GPU,
4 tris/clock pretty much implies they did that.
 
AMD-Radeon-R9-290-specifications.jpg
 
The 290 might be enough to compete against the 780, it's likely that some of the leaked 290X benchmarks are actually 290 benches.

Also i have to say that i'm already sick of the naming scheme, the difference between XT and Pro models only being an "X" is annoying.
 
Last edited by a moderator:
Also i have to say that i'm already sick of the naming scheme, the difference between XT and Pro models only being an "X" is annoying.

I don't think the "X" is intended as a XT model. Memory only runs @5GHz, that's quite a bit slower than the state of the art. I'm guessing AMD intend to bin highend dies for a future XT product with higher clocks and faster memory.

Cheers
 
with the 290 having 90% of the ALU's the 290x does. Aftermarket cooled 290 we come for sure! hope they dont take long to appear i dont know how long my 6950 (6970) will last in BF4!!!!!
 
Memory only runs @5GHz, that's quite a bit slower than the state of the art. I'm guessing AMD intend to bin highend dies for a future XT product with higher clocks and faster memory.

Or having wide enough bus they chose to save die space by using slow memory controller.
 
I have my doubts.

Hawaii-300x227.png


GL40 is almost certainly the 40 CU Firepro version of the 290 Hawaii Pro, leaving the XTGL and XT at 44 CU's, the GL Gemini as the 40 CU Firepro X2, and the LE coming later.
 
Or having wide enough bus they chose to save die space by using slow memory controller.
This was discussed earlier, right? Was the conclusion that fast 386 (Tahiti) IO blocks are larger than slow 512 (based on area of slow IO block of some lower level chip)?

It feels counter-intuitive. They are definitely shifting some cost from the die area then to the package (?) and external components and PCB. But maybe a 5Gbps/pin GDDR5 is a lot cheaper as well.

And it should be easier to create over-clockable board by populating the PCB with faster spec'ed GDDR5 RAMs, which board partners will love.

Interesting choice.
 
Back
Top