No one knew how a Northwood shrink on 90nm would do compared Prescott, but we DO have a 32nm K10 and its called Llano. Llano should be slighty faster than 45nm K10 because it had some minor architectural improvements, but is hard to directly compare to them because besides that you would have to isolate the CPU part from the GPU for any serious comparision (Maybe very easily done with a discrete Video Card, but we don't know if the IMC servicing the CPU only is as good as previous ones) and the features differs from all the others. You can't directly put it against Denebs or Thubans: You have twice the Cache L2, but no Cache L3, and are limited to Quad Core. However, you could make interesing IPC based comparisions if you picked an Athlon II X2 Regor (That got 1 MB Cache L2 per Core) against a Llano with two Cores disabled.Originally Posted by Piledriver
We also don't know the true headroom potential of Llano because there is no way to overclock it without hitting a Base Clock wall as every other Bus derives its Frequency from it, so anything could be holding you back (Including Llano very own GPU). If the supposed model with the Unlocked Multiplier shows up, interesing comparisions on Llano true CPU scaling could be made. If Deneb C3 was 3.8 GHz capable (Forget going beyond it, power consumption gets ugly), I don't see why Llano couldn't reach at least the same values, and maybe 200 MHz more, with better power consumption. Basically, Llano could put Bulldozer even to more shame.
It would be a hard pick though. I don't see enthusiasts adopting Socket FM1 even if Llano has potential as a Bulldozer alternative. You would be losing Thuban 2 extra Cores, the Cache L3, and maybe even if you don't miss Bulldozer, chances are you want to stick Piledriver in your current AM3+ Motherboard.
Besides that because the platform wasn't designed for overclocking as you can't currently isolate Llano CPU potential without messing with everything else, under the limited TDP you have both a CPU and a strong GPU, that's the reason for the conservative Frequencies. And considering that AMD is segmenting FM1 as a value/mainstream platform and AM3+ as the enthusiast one, they don't have any real reason to crank up Llano Frequency or putting it too close to Bulldozer.You haven't played with Llano have you?( a 4 Core processor that can't get past 2.9Ghz )
I'm dissapointed about Bulldozer as a whole. After hearing JF-AMD insisting on Bulldozer having higher IPC (Something that we could consider have considered set on stone coming directly from AMD), I was expecting something consistently superior to K10 and instead we got something that is noticeabily slower. The only way that IPC is higher is if they're comparing an entire Bulldozer Module against a single K10 Core. However, there are some interesing things: Bulldozer got 2000 Millions Transistors at 4 GHz and at nominal Frequency boast a respectable power consumption. However, Bulldozer is pretty much at the top of the Frequency/Voltage curve, this is also the reason why the power consumption gets ridiculous crazy with just a moderate overclock. Not only so, the 2000M Transistors are very densely packed:
Bulldozer 8C 2000M? / 315 mm^2 = 6,35
Llano 4C 1450M / 228 mm^2 = 6,36
Gulftown 6C 1170M / 240 mm^2 = 4,88
Clarkdale 2C 384M / 81 mm^2 = 4,74
Sandy Bridge 4C 995M / 216 mm^2 = 4,61
Sandy Bridge 2C (GT2) 624M / 149 mm^2 = 4,19
Sandy Bridge 2C (GT1) 504M / 131 mm^2 = 3,85
How much does the Transistor density potentially hurts yields or Frequency scaling? I suppose that such 35% higher density compared to Sandy Bridge could be a pain to handle by the fresh process at Global Foundries, and this applies to both Llano and Bulldozer.
Anyways. What dissapoints me is that no matter how faster it was supposed to be compared to a Core i5 or i7, the point is that Bulldozer can't consistently beat what it was mean to replace and you don't need Intel competing when AMD older Processors put it to shame. Not only that, but I doubt many people got enough knowledge on the particularities of Bulldozer to determine the sum of things that it is lacking before it can bring real competitive performance. Maybe one Stepping or two, like for Barcelona TLB bug? It could improve Frequency headroom and maybe fix anything that could cause subpar performance on a subsystem, like the Cache performance. Maybe an architectural revision, requiering us to wait for Piledriver so it does what Stars did for Barcelona? Or is the design flawed and unworkable and maybe we could see a K10 variant at 22nm making a comeback P6-style?
Reply With Quote
Bookmarks