AMD: Volcanic Islands R1100/1200 (8***/9*** series) Speculation/ Rumour Thread
The more complex voltage regulation is probably one of the saving graces, and it's still primitive relative to what the best CPUs are doing. If power delivery were better, maybe consumption would have been lower.
AMD is able to take things closer to the edge than Nvidia's slower solution, which might be able to more decisively best GCN if it didn't have to be more conservative.
Kudos to AMD to getting a power management loop with tens of microseconds of latency. That's not as fast as the best CPUs have done (granted one of those techs, Foxton, may have been too fast and unpredictable), but it's a different time regime entirely than 100ms.
So, good physical implementation does help compensate for a process that's not going anywhere and an architecture whose eye is not wholly on the (API-based?) gaming ball.
The higher temp setting might be important, however.
In terms of reliability, keeping the package and its components at a consistent temperature instead of cycling very rapidly might be worth the higher leakage that results. It also probably helps to have very fast power control because there is less margin below thermal runaway.
The possibility seems real to me that growing the die to the same size as Titan might improve power efficiency without necessarily increasing the performance by the same amount.
The total number of active transistors isn't really going to change when it's very clear that AMD is doing everything it can to keep under a power ceiling--up to and including weakening the descriptiveness of its product specifications and what it promises gamers with its "up to" clock speeds.
That leaves decreasing density with physically larger power-saving features, different device sizings, or devoting more and slower logic to the same tasks or for more complex power-saving features.
The tenor of the discussion around 20nm planar is that it isn't that great. This is relative to a modern scaling situation where good scaling isn't all that great on an absolute scale, and being planar means the foundries are again one node too late on improving their device properties like they were with HiKMG.
I keep wondering if they could ride out Hawaii until the hybrid FinFET nodes get out there, but betting on the foundries' timeliness sounds risky.
I'd be curious because of leakage concerns if AMD is going to stick with 95C at 20nm.
The hybrid nodes might be to 28nm what it was to 55nm.
I'd love to be able to ask an AMD engineer, "your GPU consistently burns more watts for similar performance, why?"
I guess some of it is down to the more distributed hardware that follows AMD's compute focus.
The graphics-specific portions don't seem to be evolving as much, and for games this seems to translate to years of curious tessellation behavior, the same filtering advantages or deficits versus Nvidia at different formats, improving but generally less consistent frame times.
There is still that impression that the GPUs get CPU limited sooner than the competition, and GCN consistently shows that it needs more work in terms of size N in compute or resolution in gaming to build up a full head of steam--with the power consumption to match.
I've made my uncertainty about whether GCN as it stands is all the way to where it needs to be in the Mantle thread, though.
I do not know if it is a victory. Sure it the fastest card, but it also needs the most energy and the reference cooler is just bad and the rpice is also very low. NV has been selling that performance level for $1000 and AMD starts at nearly half.
Well, it's a victory for us. I think AMD would also consider it a victory, however small and short-lived.
Still think NVIDIA will beat the 290X with a faster GK110 (GK180?) part, but gone are the days of the $1000 single GPU card (for now at least, and you could argue that ended with the GTX780).
TSMC claims 30% less power vs. 28nm HPM, for what it's worth.
http://www.cadence.com/Community/bl...-20nm-16nm-finfet-and-3d-ic-technologies.aspx
Anyone clear on whether "Volcanic Islands" as a term ever came up for the 290 series?
D
Deleted member 13524
Guest
GF110 -> Geforce GTX580: 1:8 DP - 200 DP FLOPS
GK104 -> Geforce GTX680: 1:24 DP - 128 DP FLOPS
Did you also facepalm then?
Tahiti was an all-around flagship for all-around performance, with special commitment to computing performance. Hawaii is a chip made for gaming (at high resolutions) above all.
You can see that through the architectural changes AMD made from Tahiti to Hawaii:
- Doubled the front-end
- Significantly increased memory bandwidth
- Trimmed down on DP performance for area/power/heat savings
- Marginally increased ALU performance
I wonder how much the console connection influenced the decreased focus on DP? I'm guessing it played a big part. We may well see the next iteration of GCN go back to the higher DP ratio.
Zero. As mentioned in the initial GCN architecture introductions back with Tahiti, DP is just a scalable aspect of the architecture that can and will be tuned dependent on product needs and focus.
Dave can you say more about how that scales? Are there physical changes to the ALUs, data paths etc depending on target DP performance?
It's much clearer on the nVidia side - they seem to either scale by the number of units (680) or simply throttle it in software (780).
Physical implementation in logic I have no clue about. But likewise, there is a maximum physical DPFP rate and a register controlled rate, if need be.
So why does Hawaii have reduced DP? It's very surprising for a flagship with more ACEs than Tahiti, a lot of bandwidth, and a 512-bit bus that could accommodate 8GB of RAM. I mean, apart from DP, Hawaii looks like a great GPGPU chip.
Can you say if the 1:8 rate is what the Hawaii silicon is really capable of or if a possible FirePro version could be faster? Some testers said that Mike Mantor indicated that possibility to them.
Good answer! Theory quashed!
I would be notably surprised if the FirePro (or some other SKU / line that would try again where Firestream failed epically) did not expose more DP throughput.
D
Deleted member 13524
Guest
I disagree. The ALU-to-ROP ratio is the lowest we've seen in AMD flagships for quite a while.
Want better than a Tahiti for GPGPU? Get another Tahiti. It's a more efficient GPU for that.
I think the ACEs increase is the real console-inspired modification (to be more precise: the PS4's SoC), probably because there's a good chance that multi-platform ports will make heavier use of GPGPU.
It could also be a strategy to hold out in that market with Tahiti until 20nm arrives and focus on the gaming market with Hawaii sacrificing DP performance (and possibly ECC) in the process. You have to weigh the saved silicon and therefore incrementally higher profit for each sold consumer GPU against the possible additional revenue generated from the relatively low volume (but high margin) FirePro sales. I don't have the numbers to say how this would calculate, of course.
Hawaii offers more FLOPS per square mm than Tahiti (regardless its 64 ROPs).
But at least the R9 290X incarnation offers just a bit more than half the DP flops per mm²
.