seronx and I are very different people, for one he already made up his mind. My interest in AMD is not in the FX parts, only in the APU's: Either Intel Haswell vs. AMD Kaveri, no doubt Intel will be better on CPU-intensive, I'm only curious about the IGP. I'm planning on Maxwell/Sea Islands graphics card, which means I'll be stuck on IGP for a while and it needs to be decent enough from my GTX260 and as RMA backup. AMD is so far lucky with Llano and Trinity, but if they drop the ball on Kaveri, I'm done. The rumor that Intel plans to put 64MB of integrated flash memory as an L4 dedicated to the Haswell IGP is intriguing; unless they drop the ball too, I'll be a sad panda.
Steamroller is mainly a CPU enhancement* rather than a GPU enhancement. Kaveri has:
4 SR Cores vs 4 PD Cores
8 GCN Clusters vs 6 VLIW4 Clusters
Dual Channel DDR4 vs Dual Channel DDR3
FM2+ vs FM2
As for the leak, why does it say 4C/8T?
Windows 7 patches
I read that for some reason some programs react to AMD's cores like SMT rather than physical cores which has lead to the suspicion that their FX8000's are really just quads with HT, but I don't know enough about the situation to be so certain.
Windows 7 patches
A relatively underperforming 8-core doesn't mean there are actually less cores, Intel bothered to make 1.8GHz Xeon 8-core in LGA2011 which no doubt any i7 would beat; which ironically beats this AMD ES processor.
I'm not sure if they ever updated Cinebench R11.5 but currently R11.5 uses AVX codepaths for Intel and SSE2 codepaths for AMD. This wasn't a problem with Phenom II but is a huge problem for Bulldozer-derived architectures.
Another aspect of curiousity, Cinebench is purely CPU-intensive, so it scales with frequency and cores 100% within the same architecture; just like 3Dmark06's CPU test is also fully CPU-intensive. So far, the reports for AMD's "8-core" are of a stock 4GHz part with Turbo to 4.2GHz. So if this leak shows a 3.32GHz ES part scoring 5.73, then the retail FX8350 should actually score 7.0. In the sense of absolute performance, yes it is pitiful given clock-for clock against Intel hasn't changed in six years. But the only thing AMD has going for them is retail price in the end, it will always be cheaper than the performance difference.
It falls in line with the Trinity improvements both use improved Komodo cores.
*Kaveri improvements to the core:
Improved Front-end with the ability to dispatch to both cores simultaneous. (Decouples dispatch from the VMT front-end)
Improved Floating-point unit with the ability to execute 4 AVX2** FMA3(Intel)/FMA4(AMD) ops per cycle(2 FP+2 Int)
The cache design was also changed but the only changes that are available is L2 size for Kaveri. Which is 2.4 MB per L2.
**Please note: http://software.intel.com...tref_avx2_overview.htm
AVX2 -> 256-bit Floating-point and 256-bit Integer
Ivy Bridge iGPU:
24 EUs * 4 32-bit * 2 FMA * ~1.3 GHz => ~249.6 GFlops
40 EUs * 4 32-bit * 2 FMA * ~1.5+ GHz => ~480 GFlops w/ 64 MB L4
384 ALUs * 1 32-bit * 2 FMA * ~0.76 GHz => ~583.68 GFlops w/ 256-bit bus connected to two independent 64-bit DDR3 channels
512 ALUs * 1 32-bit * 2 FMA * ~0.76 GHz => ~778.24 GFlops w/ 256-bit bus connected to two independent 64-bit DDR4 channels
CPU wise: (Only going to do Kaveri and Haswell)
Kaveri -> 4 Cores * 4 FP units * 4 32-bit * 2 FMA * ~4 GHz => ~512 GFlops
Haswell -> 4 Cores * 2 FP Units * 8 32-bit * 2 FMA * ~4 GHz => ~512 GFlops
^-- The design choices are really which one is efficient:
Do you want to finish your work faster but consume more power? (AMD) or vice versa (Intel)
Edit: Here is Trinity and Ivy Bridge in comparison
Trinity -> 4 Cores * 2 FP units * 4 32-bit * 2 FMA * ~4 GHz => ~256 GFlops
Ivy Bridge -> 4 Cores * 2 FP units * 8 32-bit * ~4 GHz => ~256 GFlops
^--- Design choices:
Does your workloads support FMA or your applications benefit from FMA? (AMD) or vice versa (Intel)
^^--- Remember both Steamroller and Haswell will support FMA. FMA3 Non-destructive and FMA4 Non-destructive.
<message edited by seronx on Wednesday, August 08, 2012 2:36 AM>