Besides the new Haswell design and DDR4 SDRAM, the LGA2011-v3 platform features the new Intel X99 chipset (codenamed Wellsburg). Replacing the outdated X79, it brings a full set of modern interfaces into Intel’s top-end platform. The X99 surpasses the recently released Z97 chipset for LGA1150 processors in terms of functionality.
We’ll discuss X99-based mainboards in our upcoming reviews, so today we’ll just briefly describe the chipset. Here’s its diagram:
The diagram is correct for the CPU models with 40 PCIe 3.0 lanes which support PCIe x16 slots with a speed formula of 16x/16x/8x or 8x/8x/8x/8x/8x (the mainboard needs an additional clock generator for the latter). The Core i7-5820K has 28 PCIe 3.0 lanes, so it will supports speed formulas of 16x/8x/4x or 8x/8x/8x. The specific implementation will vary on different mainboards, so if you’re planning to build a multi-GPU configuration, you should check out beforehand how your mainboard distributes PCIe lanes among its slots.
The LGA2011-v3 platform uses DDR4 SDRAM. According to the specs, the top memory clock rate supported by the new CPUs is 2133 MHz. But in practice Haswell-E CPUs can clock memory at up to DDR4-2666 in their default mode and up to DDR4-3200 when overclocked. DDR4 SDRAM not only features higher bandwidth but also allows building large-capacity memory subsystems. Most LGA2011-v3 mainboards for desktop computers have two DIMM slots for each of the four memory channels and can work with up to 64 gigabytes of system memory. In the near future, unbuffered DDR4 modules with capacities of 16 GB are going to arrive, so the maximum amount of memory supported by Haswell-E based computers will go up to 128 gigabytes.
The Intel X99 has 10 SATA 6 Gbit/s ports via two independent controllers. Only one controller lets you build RAID arrays, so you can unite up to six drives into a RAID, just as with the previous chipset. The X99 also features the Flex I/O technology for combining SATA ports and chipset-based PCIe lanes into M.2 and SATA Express interfaces, so these two interfaces are fully supported.
A high-speed USB controller is implemented in the X99, so six out of the chipset’s 14 USB ports can work as USB 3.0 with a peak bandwidth of 5 Gbps.
We can note one weak spot in the X99 design. The DMI 2.0 bus which connects the chipset and the processor still uses four PCI Express 2.0 lanes. Its bandwidth of 20 Gbps in each direction may turn out to be insufficient for all the numerous high-speed interfaces implemented in the chipset.
That’s why the fastest interfaces are supposed to be implemented with CPU-based PCIe lanes on the new platform. Talking about the X99, Intel promises that there will be X99-based mainboards with the Thunderbolt 2 interface featuring a peak bandwidth up to 20 Gbps. Four times the bandwidth of USB 3.0, it can only be implemented using high-speed CPU-based PCI Express 3.0 lanes. Some mainboard makers implement the M.2 interface in the same way.
The LGA2011-v3 socket you can find on X99-based mainboards is much alike to Socket LGA2011 but there’s no electrical or mechanical compatibility between the older Ivy Bridge-E and the newer Haswell-E processors. You cannot put a wrong CPU in. The socket’s fastening levers have changed somewhat, too.
Mainboard makers have prepared a lot of X99-based LGA2011-v3 products, some of which are in the micro-ATX form-factor. They are priced at $210 to over $500. The least expensive offers come from ASRock and MSI (e.g. ASRock X99 Extreme3 and MSI X99S SLI Plus) whereas ASUS ships the most expensive ones (ASUS X99-E WS and ASUS Rampage V Extreme).
Haswell-E series processors can only work with DDR4 SDRAM. It means the DIMM slots on LGA2011-v3 mainboards have 48 more contacts and a differently located key. DDR3 SDRAM is not an option with the Haswell-E. DDR4 modules are visually similar to DDR3 modules except that the middle contacts are longer than the outermost ones to make it easier to plug them into slots.
DDR4 SDRAM has a revised internal design to make it more appropriate for high-speed multi-core CPUs. A typical 8-gigabit DDR4 SDRAM device with a 4-bit interface consists of four groups of four banks each. Inside each bank there are 217 (131072) rows, each 512 bytes long. A same-capacity DDR3 device consists of only eight banks with 216 (65536) rows, each 2 KB long. Thus, DDR4 devices have more banks but smaller rows, which helps execute incoming requests in parallel.
It is the faster parallel execution on the bank level that enables DDR4 SDRAM to work at high clock rates. As a matter of fact, any SDRAM, including DDR4, uses an internal clock rate of 100 to 266 MHz, and this clock rate has remained the same since the previous century. The higher clock rate of the external interface, which can vary from 2133 to (in a long-term perspective) 4266 MHz, is achieved by multiplexing data requests to different memory banks. Coupled with the 8x data prefetch introduced in DDR3 SDRAM, each outside access to DDR4 SDRAM generates 16 internal parallel data transfers. That’s why DDR4 SDRAM can theoretically have twice the clock rate of DDR3 SDRAM, although with higher latencies.
Currently, Haswell-E processors officially support DDR4-2133 with timings of 15-15-15 as specified by the JEDEC standard. And this seems to be worse than the speed offered by modern DDR3 SDRAM. Intel has never been pedantic in implementing memory standards, though. Haswell-E processors have unlocked multipliers, also for the memory controller, and there are faster memory modules available, up to DDR4-3200 CL16. You should only keep it in mind that the top guaranteed DDR4 frequency is limited to 2666 MHz unless you overclock your CPU. Higher clock rates are only possible if you change the base clock rate from the default 100 MHz to 125 MHz.
Thus, the single key advantage of DDR4 SDRAM so far is its low heat dissipation. It typically works at a voltage of 1.2 volts whereas overclocker-friendly modules may be designed for 1.35 volts. The new memory is 30-40% more energy efficient than DDR3 SDRAM.