I think it's going to depend on how much they managed to bump up the IPC. Trinity desktop parts clock between 3ghz>low 4ghz, thus if they managed to make up that ~10% IPC gap that was created by the Bulldozer slip-up it's reasonable to conclude that any clock speed increase will translate into a direct performance increase. Llano tops out at 3ghz while Trinity tops out at 4.2ghz (desktop parts, the both of them) so the performance may in fact be that 20-30% claim -- the 10%+/- difference attributed to the IPC term encompassing all general workloads. The question here would be whether they managed to get that 10% IPC back from the Phenom II/Thuban > Bulldozer decrease. The leaked benchmarks showed a mixed bag, with an equal IPC, slightly higher or slightly lower and they only represent a very narrow workload (I believe it was Folding?). Needless to say it's difficult to estimate, but we've got the clock speeds so all we're missing is how much they've improved per-clock.
Bulldozer was initially designed to maintain IPC level while boosting clock speeds up 30%. Neither of those goals were attained and we received a Faildozer instead. This time around, though, AMD has managed to clock an absolutely massive APU (monstrous GPU inside it. That's taking up probably more than half the die space) and still get it at 4.2ghz at an equal 100W TDP to Llano. Clearly they've done quite a bit of improvement in perf-per-watt when compared to Bulldozer. It's not easy getting 384 VLIW4 shaders clocked at 800mhz and a 2 modules at 4.2ghz under 100W, so that's incredibly impressive.
Ultimately it boils down to IPC gains. What I'd like to see more in the Trinity APU is some significant IMC gains as well because relying on DRAM frequency is an empty goal if you're unable to saturate that 128-bit bus the APU shares between the GPU+CPU. Some of the other improvements should help the CPU though I'm not sure how significantly that matters as far as GPU performance goes.
Improved prefetch, branch prediction and L2 efficiency are huge and means AMD went right to the heart of the problems at Bulldozer; and that's a laundry list of changes included on those slides and there's likely far more that we're not seeing. Clearly they realized they fucked up big time and they absolutely had to address the issues immediately. This bodes well for Vishera which AMD has already hinted further improvements and not just the addition of L3 cache, so there's a good chance Vishera looks even better than what we see in the above Piledriver cores in Trinity. Ultimately the real question will be how much they've improved upon Bulldozer and how close they've chipped away at that -10% IPC. At the moment that's something we don't know :/ The clock speeds for Vishera are also an unknown at this point. It's quite clear that Trinity is clock speed limited by those absolutely massive VLIW4 shaders and that it's clock speeds could likely have exceeded the mid 4ghz range had it not been TDP limited and forced to lug around those extra transistors. There's also the issue of the diminishing returns past 4ghz on the licensed RCM tech that AMD applied in Piledriver. So how high can the Piledriver cores really clock? and how high can they clock with that 8MB L3 cache? What's the OC ceiling? Who knows, but what seems certain is that the mobile parts will clock quite close to their desktop brethren at lower TDP, aggressive turbo included. Good news for laptops but conceivably not-as-impressive news for Vishera.
I wouldn't expect Sandy levels of performance but maybe what Bulldozer was gunning for originally, somewhere in that 10-15% better than Nehalem range, mainly attributed to clock speed differences.
Facebook
Twitter