Harpertown Performance Preview

Pages: 1 2 3 4 5 6 7 8 9 10 11 12

The Stoakley Platform

The Seaburg or 5400 chipset is the core of the new Stoakley platform, and is generally targeted at workstation and HPC systems – although there is no reason why it cannot be used as a server. Like Harpertown, the Seaburg chipset is a derivative of the previous generation (Blackford chipset), but with several alterations. The Seaburg chipset inherited the point-to-point bus arrangement used in Blackford, although it runs faster at 1.6GT/s rather than 1.33GT/s. The request buffering is also deeper to utilize the extra bandwidth, and a new out-of-order transactional mode is supported, which improves performance for many HPC workloads. The memory addressing and capacity has been doubled. A single system can now deploy up to 32 DIMMs and address 128GB of physical memory. Figure 1 below shows a system diagram of Seaburg.

Figure 1 – Stoakley Platform

The snoop filter in the memory controller has also been expanded and re-architected for better performance. The snoop filter is physically larger to accommodate the larger caches used in Harpertown (24MB total for a 2 socket system, versus 16MB with Clovertown). The filter has also been organized into four affinity groups, one for each L2 cache in the system to reduce negative interactions. The affinity groups ensure that when processor 0 brings in a new cache line, it will evict one of the older entries in the snoop filter for processor 0. In the previous Blackford design, when processor 0 brought in a new line, it could end up forcing the snoop filter to evict a line from a different cache (remember that when the snoop filter evicts a cache line, the cache must also evict that line as well).

Each processor can also provide ‘way hints’ to ensure that the right cache lines are cast out of the snoop filter. The snoop filter can also migrate cache lines between each affinity group, so that when a cache line moves to a different processor, the snoop filter entry will migrate in a corresponding manner.

The most visible changes from an end-user perspective though are in the I/O system. Seaburg sports a pair of x16 PCI-Express Gen2 slots meant for graphics cards (probably Quadro graphics cards). The previous workstation chipset (Greencreek) was under-equipped and could only use a single graphics card. These lanes can also be configured as 4 x8 PCI-E Gen1 slots or 8 x4 slots for connection with various storage subsystems or high performance networking such as 10 gigabit Ethernet.

Pages: « Prev   1 2 3 4 5 6 7 8 9 10 11 12   Next »

Discuss (35 comments)