A new graphics engine is now integrated on-die. It has its own voltage plane, and takes full advantage of the power management built into Sandy Bridge. As with past Intel graphics efforts, it’s aimed at providing a “mainstream” gaming experience, but it may finally achieve parity with integrated (or entry level discrete) solutions from AMD and Nvidia.
- The execution units (“shader units” have been improved so that each EU is now 2x faster than existing EUs in current Intel graphics cores (the company claims). A maximum of 12 EUs will be built into the first release, but these numbers may be scaled back for lower-end products.
- MPEG-2, H.264, and VC-1 decodes are now offloaded into a dedicated fixed-function unit inside the media processing engine.
- The graphics core is “DX10+”, supporting DirectX 10 graphics, but allowing some elements of DirectCompute to run on the GPU cores (which is allowed by DirectX 11). Other DirectX 11 features, particularly hardware tessellation, are not supported.
- Turbo Boost will also work with the GPU side of the processor, scaling up frequency for short periods of time under heavy loads.
Perhaps the most significant design decision was to move the GPU to the same side of the highest-level cache as the CPU. This means that the GPU and CPU now share the last level cache (what Intel calls the “LLC”). This is typically an L3 cache on most modern CPUs. What data is cached by the GPU is determined by the graphics driver.
The System Agent (formerly “Uncore”)
What Intel is now calling the “System Agent” integrates the internal interconnects, memory controller, and power control unit (PCU).
- The interconnect consists of a high speed, modular ring topology. Each connection can move data at peak throughputs of up to 96 GB/sec. Bandwidth scales with cores: a quad core CPU will move up to 384GB/sec, while a dual core CPU supports “only” 192 GB/sec.
- What Intel is calling the “last level cache” is also modular. While some LLC is required, the sizes are flexible.
- The PCU (power control unit) is smarter than the one built into Westmere and Nehalem, and also can power-manage the graphics part of the chip.
- The memory controller is a dual channel DDR3 controller, clearly defining this as a mainstream CPU, and not one oriented towards performance enthusiasts.
- Further reinforcing the mainstream/mobile emphasis, only 16 PCI Express lanes originate from the PCI Express controller on the CPU, so only a single x16 or dual x8 discrete graphics cards will be supported.
Sandy Bridge will require new motherboards using an 1155-pin interface. The supporting chipsets will be dubbed H67 and P67. If you drop in a Sandy Bridge into a P67 motherboard, the on-die graphics engine is disabled completely, including the media decode engine. So, all media decode will either be handled by the discrete GPU or the CPU itself.
Overall, Sandy Bridge looks to be a solid mainstream offering. Performance enthusiasts should note that LGA 1366 is not going away, and Westmere-based hexa-core CPUs will continue to be offered. Intel even suggested that future LGA 1366 offerings may become available, but wasn’t prepared to make any definitive announcements.
On the other hand, Sandy Bridge-based quad-core systems should make excellent mainstream gaming systems. The only downside is the low number of PCI Express lanes available for discrete graphics. But a quad-core system built on P67 with an Nvidia GeForce GTX 480 or Radeon HD 5970 should still be a pretty hardcore gaming rig.