Check out my first novel, midnight's simulacra!

Nehalem: Difference between revisions

From dankwiki
(Created page with '* Store-forwarding aliasing issue on 4k strides. * Double-pumped FP SSE + integer SSE/x87 + load + store units * Fetch up to 16 bytes of aligned instructions from cache per cycle...')
 
mNo edit summary
 
(13 intermediate revisions by the same user not shown)
Line 1: Line 1:
[[File:Nehalem.svg|thumb|right|Nehalem-EP NUMA arrangement]]
[[File:IAmNehalem.jpg|thumb|right|Ehyeh asher ehyeh]]
[[File:Intel Nehalem arch.png|thumb|right|Nehalem microarchitecture]]
The successor of [[Core 2]], and predecessor of [[Sandy Bridge]].
* Move to 133MHz [[QPI (Quick Path Interconnect)]], against which the CPU clock is multiplied
* Reintroduction of [[SMP on x86#SMT|SMT (HyperThreading)]]
* Store-forwarding aliasing issue on 4k strides.
* Store-forwarding aliasing issue on 4k strides.
* Double-pumped FP SSE + integer SSE/x87 + load + store units
* Double-pumped FP SSE + integer SSE/x87 + load + store units
Line 6: Line 12:
* Forwarding results between integer, integer SIMD, and FP units adds latency compared to forwards within the domain.
* Forwarding results between integer, integer SIMD, and FP units adds latency compared to forwards within the domain.
* One register may be written per cycle.
* One register may be written per cycle.
* 48 load buffers, 32 store buffers, 10 fill buffers.
* 48 load buffers (up from 32), 32 store buffers (up from 20), 10 fill buffers.
* 36 reservation stations, 96 ROB entries.
* 36 reservation stations (up from 32), 128 ROB entries (up from 96).
* Calltrace cache of 16 entries.
* Calltrace cache of 16 entries.
* 2-way loop end BTB for every 16 bytes, 4-way general BTB.
* 2-way loop end BTB for every 16 bytes, 4-way general BTB.
* Loop Stream Detector replays from IDQ if the loop consists of:
* [[Daytripper|Loop Stream Detector]] replays from IDQ if the loop consists of:
** 4 16-byte icache fetches or less
** 4 16-byte icache fetches or less
** 28 total µops or less
** 28 total µops or less
Line 17: Line 23:
* Be sure to use register parameter-passing conventions, not the stack, to avoid stalls on store-forward of high-latency floating point stores.
* Be sure to use register parameter-passing conventions, not the stack, to avoid stalls on store-forward of high-latency floating point stores.
* Peak issue rate of 1 128-bit load and 1 128-bit store per cycle.
* Peak issue rate of 1 128-bit load and 1 128-bit store per cycle.
* [[Turbo Boost]] in 133MHz increments
==See Also==
* The Dark Knight's [http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3448 AnandTech] article, 2008-11-03.
[[Category: x86]]
[[CATEGORY: Hardware]]

Latest revision as of 16:32, 25 February 2012

Nehalem-EP NUMA arrangement
Ehyeh asher ehyeh
Nehalem microarchitecture

The successor of Core 2, and predecessor of Sandy Bridge.

  • Move to 133MHz QPI (Quick Path Interconnect), against which the CPU clock is multiplied
  • Reintroduction of SMT (HyperThreading)
  • Store-forwarding aliasing issue on 4k strides.
  • Double-pumped FP SSE + integer SSE/x87 + load + store units
  • Fetch up to 16 bytes of aligned instructions from cache per cycle.
  • Up to 4 instructions, no more than 1 complex (this does not necessarily mean 1 µop), decoded per cycle. 64-bit macro-fusion
  • Instructions with more than 4 µops are fed from MSROM, and will take more than one cycle in the Instruction Decoder Queue.
  • Forwarding results between integer, integer SIMD, and FP units adds latency compared to forwards within the domain.
  • One register may be written per cycle.
  • 48 load buffers (up from 32), 32 store buffers (up from 20), 10 fill buffers.
  • 36 reservation stations (up from 32), 128 ROB entries (up from 96).
  • Calltrace cache of 16 entries.
  • 2-way loop end BTB for every 16 bytes, 4-way general BTB.
  • Loop Stream Detector replays from IDQ if the loop consists of:
    • 4 16-byte icache fetches or less
    • 28 total µops or less
    • 4 taken branches or less, none of them a RET
    • preferably more than 64 iterations?
  • Be sure to use register parameter-passing conventions, not the stack, to avoid stalls on store-forward of high-latency floating point stores.
  • Peak issue rate of 1 128-bit load and 1 128-bit store per cycle.
  • Turbo Boost in 133MHz increments

See Also

  • The Dark Knight's AnandTech article, 2008-11-03.