After some wait after Conroe hit the shelves, Intel struck again with a new architecture. While Lynnfield was a mild update of the previous generation, the new Sandy Bridge is more than just evolution.
The first thing Intel did was to change the Level 0 Cache which now stores µOPs instead of the full instructions. While it was common to execute whole instructions, Intel changed that a while ago due to the speed advantage of RISC (Reduced Instruction Set computiing). Any complicated instruction is split in more than one µOP which is in fact a RISC instruction. While the CPU fetches a new instruction, the decoder checks if the cache already has this instruction cached and avoids decompiling the instruction again. The cache-size of about 1.5k µOPs is big enough to contain pretty much any instruction which will increase performance. The L1 caches did not change, the instructions and data feature 32kB each. The µOP cache is also included in the L1 cache which gives an approximatly 80% hit-chance according to Intel.
The next improvement is the redesigned branch prediction. It is especially improved with long branches, although it has its shortcomings with shorter ones such as loops and elseif constructions. This is quite complicated, so let us just say that the troughput has been improved and enables the CPU for the new AVX instructions. They are an improvement over SSE4 with a bitwidth of 256bits per instruction. Apart from new instructions, the important news is that a two-operand form of a:=a+b can now use a non-destructive three-operand form of c:=a+b, preserving both source operands, which will reduce the code significantly.
Also notable is that turbo has also been improved. The new CPU always clocks 100MHz higher than specified, so a i5-2500 will always clocks at 3.4GHz because the thermal budget is never used as specified. In single threaded applications the CPU increases the clock by 400MHz, with two cores by 300MHz and with three cores under load by 200MHz.
With the higher throughput of the CPU, Intel also needed to improve the memory controller. Lynnfield could access two memory requests per cylce which resulted in 16 bytes load and 16 bytes store per cycle. The new memory controller can now read two memory requests and one write memory request per clock. Faster memory will not improve your performance as in previous generations making faster modules nearly useless. The standard support has been increased to 1333MHz, anything beyond that is offically not supported, except for some notebook CPUs.
Besides that, Intel improved many things inside the CPU, improved power managment and also accelerated the GPU architecture. The GPU is now twice as fast a the previous generation and in case of the HD3000, it has twice as many execution units, which should give you four times the performance of its predecessor. It's still too slow for any serious gaming but should be fine for HD playback and everyday work.
However, the new architecture has two downsides to it. First, Intel changed the socket to force any customer to buy a new motherboard. The second one is that overclocking is very restricted with any normal CPU. CPUs without Turbo can't be overclocked at all. The clock generator is now within the CPU and can only be overclocked to about 106MHz which also increases the clocks for PCIe and PCI. If you have a P67 or Z68 chipset, turbo can overclock by another 400MHz as long as your CPU stays inside the TDP budget. H65/Q6x/B65 don't allow for any overclocking besides GPU. Of course the chipset Intel sells to you is always the same chip, just with some fuses blown or not.
For overclockers Intel offers the "K" CPUs which are slightly pricier than the non-K counterparts. Of course there is a downside to that, because Intel refuses to provide their customers with VT-d, which is important if you want to have faster virtual machines. You do get the fastest GPU core but it's quite useless for gamers.
The naming is still a mess with meaningless numbers and without structure. The i5-2390T especially confuses because it isn't a quad core CPU which you would suspect, but just a dual-core with Hyperthreading enabled. They should have named it i3-2390T, but luckily it's still not in stores yet.
For our tests we have three CPUs. The i3-2100T which is a dual-core CPU clocked at 2.5GHz with Hyperthreading but no Turbo, 3MB L3 cache and only 35W TDP but fused off AES support. The second one is the i5-2500K which is a quad core clocked at 3.3GHz, 6MB L3 Cache, multiplicator free without Hyperthreading and a TDP of 95W. The last one is the i7-2600K which is basically the same as the i7-2500K but clocked at 3.4GHz, Hyperthreading enabled and 8MB L3 cache.
The new 60 Series chipset is just an upgrade to the existing 50 Series. The only notable point is the dual 6Gbps SATA controller while the rest stayed the same.
The Z68 is the high-end chipset that supports graphics and CPU overclocking as well as 2x PCIe 2.0 x8 support. The P67 is the same without graphics, while the H67 can just overclock the GPU but is limited to 1x PCIe 2.0 x16.
Because Intel wants to get rid of the PCI bus, they have blown the fuse for that part for all higher-end SKUs, just leaving the Q67 and B65 with native PCI support. If you own professional video or audio-cards you may run into trouble due to increased lag of the addtional PCIe-PCI bridge-chips, which even Intel mounts on its boards. It's more than silly to force vendors to use a seperate PCIe-to-PCI bridge because this increases costs for the end customer. The H61 and B65 can't overclock and also loose RAID support, and while the B65 supports at least one 6Gb SATA connection, the H61 has just four 3Gb connectors and looses even the dual channel memory support. We think that cutting features is easy to overdo but the company that makes 24 billion turnover and 6 billion profit, does tend to screw its customers on a constant basis.