x86-64

AMD64 was created as an alternative to the radically different IA-64 architecture, which was designed by Intel and Hewlett Packard. Originally announced in 1999[14] while a full specification became available in August 2000,[15] the AMD64 architecture was positioned by AMD from the beginning as an evolutionary way to add 64-bit computing capabilities to the existing x86 architecture, as opposed to Intel's approach of creating an entirely new 64-bit architecture with IA-64.

The first AMD64-based processor, the Opteron, was released in April 2003.

AMD's processors implementing the AMD64 architecture include Opteron, Athlon 64, Athlon 64 X2, Athlon 64 FX, Athlon II (followed by "X1", "X2", "X3", or "X4" to indicate the number of cores, and XLT models), Turion 64, Turion 64 X2, Sempron ("Palermo" E6 stepping and all "Manila" models), Phenom (followed by "X3" or "X4" to indicate the number of cores), Phenom II (followed by "X2", "X3", "X4", or "X6" to indicate the number of cores), FX, Fusion/APU and Ryzen/Epyc.

The primary defining characteristic of AMD64 is the availability of 64-bit general-purpose processor registers (for example, rax and rbx), 64-bit integer arithmetic and logical operations, and 64-bit virtual addresses. The designers took the opportunity to make other improvements as well.

Some of the most significant changes are described below:

64-bit integer capability

All general-purpose registers (GPRs) are expanded from 32 bits to 64 bits, and all arithmetic and logical operations, memory-to-register and register-to-memory operations, etc., can operate directly on 64-bit integers. Pushes and pops on the stack default to 8-byte strides, and pointers are 8 bytes wide.

Additional registers

In addition to increasing the size of the general-purpose registers, the number of named general-purpose registers is increased from eight (i.e. eax, ecx, edx, ebx, esp, ebp, esi, edi) in x86 to 16 (i.e. rax, rcx, rdx, rbx, rsp, rbp, rsi, rdi, r8, r9, r10, r11, r12, r13, r14, r15). It is therefore possible to keep more local variables in registers rather than on the stack, and to let registers hold frequently accessed constants; arguments for small and fast subroutines may also be passed in registers to a greater extent.

AMD64 still has fewer registers than many RISC instruction sets (e.g. PA-RISC, Power ISA, and MIPS have 32 GPRs; Alpha, 64-bit ARM, and SPARC have 31) or VLIW-like machines such as the IA-64 (which has 128 registers). However, an AMD64 implementation may have far more internal registers than the number of architectural registers exposed by the instruction set (see register renaming). (For example, AMD Zen cores have 168 64-bit integer and 160 128-bit vector floating point physical internal registers.)

Additional XMM (SSE) registers

Similarly, the number of 128-bit XMM registers (used for Streaming SIMD instructions) is also increased from 8 to 16.

The traditional x87 FPU register stack is not included in the register file size extension in 64-bit mode, compared with the XMM registers used by SSE2, which did get extended. The x87 register stack is not a simple register file although it does allow direct access to individual registers by low cost exchange operations.

Larger virtual address space

The AMD64 architecture defines a 64-bit virtual address format, of which the low-order 48 bits are used in current implementations.[11]^(p120) This allows up to 256 TiB (2⁴⁸ bytes) of virtual address space. The architecture definition allows this limit to be raised in future implementations to the full 64 bits,[11]^{(p2)(p3)(p13)(p117)(p120)} extending the virtual address space to 16 EiB (2⁶⁴ bytes).[16] This is compared to just 4 GiB (2³² bytes) for the x86.[17]

This means that very large files can be operated on by mapping the entire file into the process' address space (which is often much faster than working with file read/write calls), rather than having to map regions of the file into and out of the address space.

Larger physical address space

The original implementation of the AMD64 architecture implemented 40-bit physical addresses and so could address up to 1 TiB (2⁴⁰ bytes) of RAM.[11]^(p24) Current implementations of the AMD64 architecture (starting from AMD 10h microarchitecture) extend this to 48-bit physical addresses[18] and therefore can address up to 256 TiB of RAM. The architecture permits extending this to 52 bits in the future[11]^(p24)[19] (limited by the page table entry format);[11]^(p131) this would allow addressing of up to 4 PiB of RAM. For comparison, 32-bit x86 processors are limited to 64 GiB of RAM in Physical Address Extension (PAE) mode,[20] or 4 GiB of RAM without PAE mode.[11]^(p4)

Larger physical address space in legacy mode

When operating in legacy mode the AMD64 architecture supports Physical Address Extension (PAE) mode, as do most current x86 processors, but AMD64 extends PAE from 36 bits to an architectural limit of 52 bits of physical address. Any implementation, therefore, allows the same physical address limit as under long mode.[11]^(p24)

Instruction pointer relative data access

Instructions can now reference data relative to the instruction pointer (RIP register). This makes position independent code, as is often used in shared libraries and code loaded at run time, more efficient.

SSE instructions

The original AMD64 architecture adopted Intel's SSE and SSE2 as core instructions. These instruction sets provide a vector supplement to the scalar x87 FPU, for the single-precision and double-precision data types. SSE2 also offers integer vector operations, for data types ranging from 8bit to 64bit precision. This makes the vector capabilities of the architecture on par with those of the most advanced x86 processors of its time. These instructions can also be used in 32-bit mode. The proliferation of 64-bit processors has made these vector capabilities ubiquitous in home computers, allowing the improvement of the standards of 32-bit applications. The 32-bit edition of Windows 8, for example, requires the presence of SSE2 instructions.[21] SSE3 instructions and later Streaming SIMD Extensions instruction sets are not standard features of the architecture.

No-Execute bit

The No-Execute bit or NX bit (bit 63 of the page table entry) allows the operating system to specify which pages of virtual address space can contain executable code and which cannot. An attempt to execute code from a page tagged "no execute" will result in a memory access violation, similar to an attempt to write to a read-only page. This should make it more difficult for malicious code to take control of the system via "buffer overrun" or "unchecked buffer" attacks. A similar feature has been available on x86 processors since the 80286 as an attribute of segment descriptors; however, this works only on an entire segment at a time.

Segmented addressing has long been considered an obsolete mode of operation, and all current PC operating systems in effect bypass it, setting all segments to a base address of zero and (in their 32-bit implementation) a size of 4 GB. AMD was the first x86-family vendor to implement no-execute in linear addressing mode. The feature is also available in legacy mode on AMD64 processors, and recent Intel x86 processors, when PAE is used.

Removal of older features

A few "system programming" features of the x86 architecture were either unused or underused in modern operating systems and are either not available on AMD64 in long (64-bit and compatibility) mode, or exist only in limited form. These include segmented addressing (although the FS and GS segments are retained in vestigial form for use as extra-base pointers to operating system structures),[11]^(p70) the task state switch mechanism, and virtual 8086 mode. These features remain fully implemented in "legacy mode", allowing these processors to run 32-bit and 16-bit operating systems without modifications. Some instructions that proved to be rarely useful are not supported in 64-bit mode, including saving/restoring of segment registers on the stack, saving/restoring of all registers (PUSHA/POPA), decimal arithmetic, BOUND and INTO instructions, and "far" jumps and calls with immediate operands.

Canonical address space implementations (diagrams not to scale)

Current 48-bit implementation

56-bit implementation

64-bit implementation

Although virtual addresses are 64 bits wide in 64-bit mode, current implementations (and all chips that are known to be in the planning stages) do not allow the entire virtual address space of 2⁶⁴ bytes (16 EiB) to be used. This would be approximately four billion times the size of the virtual address space on 32-bit machines. Most operating systems and applications will not need such a large address space for the foreseeable future, so implementing such wide virtual addresses would simply increase the complexity and cost of address translation with no real benefit. AMD, therefore, decided that, in the first implementations of the architecture, only the least significant 48 bits of a virtual address would actually be used in address translation (page table lookup).[11]^(p120)

In addition, the AMD specification requires that the most significant 16 bits of any virtual address, bits 48 through 63, must be copies of bit 47 (in a manner akin to sign extension). If this requirement is not met, the processor will raise an exception.[11]^(p131) Addresses complying with this rule are referred to as "canonical form."[11]^(p130) Canonical form addresses run from 0 through 00007FFF'FFFFFFFF, and from FFFF8000'00000000 through FFFFFFFF'FFFFFFFF, for a total of 256 TiB of usable virtual address space. This is still 65,536 times larger than the virtual 4 GiB address space of 32-bit machines.

This feature eases later scalability to true 64-bit addressing. Many operating systems (including, but not limited to, the Windows NT family) take the higher-addressed half of the address space (named kernel space) for themselves and leave the lower-addressed half (user space) for application code, user mode stacks, heaps, and other data regions.[22] The "canonical address" design ensures that every AMD64 compliant implementation has, in effect, two memory halves: the lower half starts at 00000000'00000000 and "grows upwards" as more virtual address bits become available, while the higher half is "docked" to the top of the address space and grows downwards. Also, enforcing the "canonical form" of addresses by checking the unused address bits prevents their use by the operating system in tagged pointers as flags, privilege markers, etc., as such use could become problematic when the architecture is extended to implement more virtual address bits.

The first versions of Windows for x64 did not even use the full 256 TiB; they were restricted to just 8 TiB of user space and 8 TiB of kernel space.[22] Windows did not support the entire 48-bit address space until Windows 8.1, which was released in October 2013.[22]

The 64-bit addressing mode ("long mode") is a superset of Physical Address Extensions (PAE); because of this, page sizes may be 4 KiB (2¹² bytes) or 2 MiB (2²¹ bytes).[11]^(p120) Long mode also supports page sizes of 1 GiB (1073741824 (2³⁰) bytes).[11]^(p120) Rather than the three-level page table system used by systems in PAE mode, systems running in long mode use four levels of page table: PAE's Page-Directory Pointer Table is extended from four entries to 512, and an additional Page-Map Level 4 (PML4) Table is added, containing 512 entries in 48-bit implementations.[11]^(p131) A full mapping hierarchy of 4 KiB pages for the whole 48-bit space would take a bit more than 512 GiB of memory (about 0.195% of the 256 TiB virtual space).

Intel has implemented a scheme with a 5-level page table, which allows Intel 64 processors to support a 57-bit virtual address space.[23] Further extensions may allow full 64-bit virtual address space and physical memory by expanding the page table entry size to 128-bit, and reduce page walks in the 5-level hierarchy by using a larger 64 KiB page allocation size that still supports 4 KiB page operations for backward compatibility.[24]

The operating system can also limit the virtual address space. Details, where applicable, are given in the "Operating system compatibility and characteristics" section.

Current AMD64 processors support a physical address space of up to 2⁴⁸ bytes of RAM, or 256 TiB.[18] However, as of June 2010, there were no known x86-64 motherboards that support 256 TiB of RAM.[25][26][27][28] The operating system may place additional limits on the amount of RAM that is usable or supported. Details on this point are given in the "Operating system compatibility and characteristics" section of this article.

The architecture has three primary modes of operation, long mode, legacy mode, and real mode.

State diagram of the x86-64 operating modes

Long mode is the architecture's intended primary mode of operation; it is a combination of the processor's native 64-bit mode and a combined 32-bit and 16-bit compatibility mode. It is used by 64-bit operating systems. Under a 64-bit operating system, 64-bit programs run under 64-bit mode, and 32-bit and 16-bit protected mode applications (that do not need to use either real mode or virtual 8086 mode in order to execute at any time) run under compatibility mode. Real-mode programs and programs that use virtual 8086 mode at any time cannot be run in long mode unless those modes are emulated in software.[11]^:11 However, such programs may be started from an operating system running in long mode on processors supporting VT-x or AMD-V by creating a virtual processor running in the desired mode.

Since the basic instruction set is the same, there is almost no performance penalty for executing protected mode x86 code. This is unlike Intel's IA-64, where differences in the underlying instruction set mean that running 32-bit code must be done either in emulation of x86 (making the process slower) or with a dedicated x86 coprocessor. However, on the x86-64 platform, many x86 applications could benefit from a 64-bit recompile, due to the additional registers in 64-bit code and guaranteed SSE2-based FPU support, which a compiler can use for optimization. However, applications that regularly handle integers wider than 32 bits, such as cryptographic algorithms, will need a rewrite of the code handling the huge integers in order to take advantage of the 64-bit registers.

Legacy mode is the mode used by 32-bit or 16-bit "protected mode" operating systems.[29] In this mode, the processor acts like an older x86 processor, and only 16-bit and 32-bit code can be executed. Legacy mode allows for a maximum of 32 bit virtual addressing which limits the virtual address space to 4 GiB.[11]^{(p14)(p24)(p118)} 64-bit programs cannot be run from legacy mode.

Real mode is the initial mode of operation when the processor is initialized. It is backwards compatible with the original 8086 and 8088 processors. Real mode is primarily used today by operating system bootloaders, which are required by the architecture to configure virtual memory details before transitioning to higher modes.

Historically, AMD has developed and produced processors with instruction sets patterned after Intel's original designs, but with x86-64, roles were reversed: Intel found itself in the position of adopting the ISA which AMD had created as an extension to Intel's own x86 processor line.

Intel's project was originally codenamed Yamhill (after the Yamhill River in Oregon's Willamette Valley). After several years of denying its existence, Intel announced at the February 2004 IDF that the project was indeed underway. Intel's chairman at the time, Craig Barrett, admitted that this was one of their worst-kept secrets.[30][31]

Intel's name for this instruction set has changed several times. The name used at the IDF was CT (presumably for Clackamas Technology, another codename from an Oregon river); within weeks they began referring to it as IA-32e (for IA-32 extensions) and in March 2004 unveiled the "official" name EM64T (Extended Memory 64 Technology). In late 2006 Intel began instead using the name Intel 64 for its implementation, paralleling AMD's use of the name AMD64.[32]

The first processor to implement Intel 64 was the multi-socket processor Xeon code-named Nocona in June 2004. In contrast, the initial Prescott chips (February 2004) did not enable this feature. Intel subsequently began selling Intel 64-enabled Pentium 4s using the E0 revision of the Prescott core, being sold on the OEM market as the Pentium 4, model F. The E0 revision also adds eXecute Disable (XD) (Intel's name for the NX bit) to Intel 64, and has been included in then current Xeon code-named Irwindale. Intel's official launch of Intel 64 (under the name EM64T at that time) in mainstream desktop processors was the N0 stepping Prescott-2M.

The first Intel mobile processor implementing Intel 64 is the Merom version of the Core 2 processor, which was released on July 27, 2006. None of Intel's earlier notebook CPUs (Core Duo, Pentium M, Celeron M, Mobile Pentium 4) implement Intel 64.

Intel's processors implementing the Intel64 architecture include the Pentium 4 F-series/5x1 series, 506, and 516, Celeron D models 3x1, 3x6, 355, 347, 352, 360, and 365 and all later Celerons, all models of Xeon since "Nocona", all models of Pentium Dual-Core processors since "Merom-2M", the Atom 230, 330, D410, D425, D510, D525, N450, N455, N470, N475, N550, N570, N2600 and N2800, all versions of the Pentium D, Pentium Extreme Edition, Core 2, Core i9, Core i7, Core i5, and Core i3 processors, and the Xeon Phi 7200 series processors.

• Intel 64's BSF and BSR instructions act differently than AMD64's when the source is zero and the operand size is 32 bits. The processor sets the zero flag and leaves the upper 32 bits of the destination undefined.
• AMD64 requires a different microcode update format and control MSRs (model-specific registers) while Intel 64 implements microcode update unchanged from their 32-bit only processors.
• Intel 64 lacks some MSRs that are considered architectural in AMD64. These include SYSCFG, TOP_MEM, and TOP_MEM2.
• Intel 64 allows SYSCALL/SYSRET only in 64-bit mode (not in compatibility mode),[38] and allows SYSENTER/SYSEXIT in both modes.[39] AMD64 lacks SYSENTER/SYSEXIT in both sub-modes of long mode.[11]^:33
• In 64-bit mode, near branches with the 66H (operand size override) prefix behave differently. Intel 64 ignores this prefix: the instruction has 32-bit sign extended offset, and instruction pointer is not truncated. AMD64 uses 16-bit offset field in the instruction, and clears the top 48 bits of instruction pointer.
• AMD processors raise a floating point Invalid Exception when performing an FLD or FSTP of an 80-bit signalling NaN, while Intel processors do not.
• Intel 64 lacks the ability to save and restore a reduced (and thus faster) version of the floating-point state (involving the FXSAVE and FXRSTOR instructions).
• AMD processors ever since Opteron Rev. E and Athlon 64 Rev. D have reintroduced limited support for segmentation, via the Long Mode Segment Limit Enable (LMSLE) bit, to ease virtualization of 64-bit guests.[40][41]
• When returning to a non-canonical address using SYSRET, AMD64 processors execute the general protection fault handler in privilege level 3,[42] while on Intel 64 processors it is executed in privilege level 0.[43]

• Early AMD64 processors (typically on Socket 939 and 940) lacked the CMPXCHG16B instruction, which is an extension of the CMPXCHG8B instruction present on most post-80486 processors. Similar to CMPXCHG8B, CMPXCHG16B allows for atomic operations on octa-words (128-bit values). This is useful for parallel algorithms that use compare and swap on data larger than the size of a pointer, common in lock-free and wait-free algorithms. Without CMPXCHG16B one must use workarounds, such as a critical section or alternative lock-free approaches.[44] Its absence also prevents 64-bit Windows prior to Windows 8.1 from having a user-mode address space larger than 8 terabytes.[45] The 64-bit version of Windows 8.1 requires the instruction.[46]
• Early AMD64 and Intel 64 CPUs lacked LAHF and SAHF instructions in 64-bit mode. AMD introduced these instructions (also in 64-bit mode) with their Athlon 64, Opteron and Turion 64 revision D processors in March 2005[47][48][49] while Intel introduced the instructions with the Pentium 4 G1 stepping in December 2005. The 64-bit version of Windows 8.1 requires this feature.[46]
• Early Intel CPUs with Intel 64 also lack the NX bit of the AMD64 architecture. This feature is required by all versions of Windows 8.x.
• Early Intel 64 implementations (Prescott and Cedar Mill) only allowed access to 64 GiB of physical memory while original AMD64 implementations allowed access to 1 TiB of physical memory. Recent AMD64 implementations provide 256 TiB of physical address space (and AMD plans an expansion to 4 PiB), while some Intel 64 implementations could address up to 64 TiB.[50] Physical memory capacities of this size are appropriate for large-scale applications (such as large databases), and high-performance computing (centrally oriented applications and scientific computing).

Preliminary infrastructure work was started in February 2004 for a x86-64 port.[55] This development later stalled. Development started again during July 2007[56] and continued during Google Summer of Code 2008 and SoC 2009.[57][58] The first official release to contain x86-64 support was version 2.4.[59]

FreeBSD first added x86-64 support under the name "amd64" as an experimental architecture in 5.1-RELEASE in June 2003. It was included as a standard distribution architecture as of 5.2-RELEASE in January 2004. Since then, FreeBSD has designated it as a Tier 1 platform. The 6.0-RELEASE version cleaned up some quirks with running x86 executables under amd64, and most drivers work just as they do on the x86 architecture. Work is currently being done to integrate more fully the x86 application binary interface (ABI), in the same manner as the Linux 32-bit ABI compatibility currently works.

x86-64 architecture support was first committed to the NetBSD source tree on June 19, 2001. As of NetBSD 2.0, released on December 9, 2004, NetBSD/amd64 is a fully integrated and supported port. 32-bit code is still supported in 64-bit mode, with a netbsd-32 kernel compatibility layer for 32-bit syscalls. The NX bit is used to provide non-executable stack and heap with per-page granularity (segment granularity being used on 32-bit x86).

OpenBSD has supported AMD64 since OpenBSD 3.5, released on May 1, 2004. Complete in-tree implementation of AMD64 support was achieved prior to the hardware's initial release because AMD had loaned several machines for the project's hackathon that year. OpenBSD developers have taken to the platform because of its support for the NX bit, which allowed for an easy implementation of the W^X feature.

The code for the AMD64 port of OpenBSD also runs on Intel 64 processors which contains cloned use of the AMD64 extensions, but since Intel left out the page table NX bit in early Intel 64 processors, there is no W^X capability on those Intel CPUs; later Intel 64 processors added the NX bit under the name "XD bit". Symmetric multiprocessing (SMP) works on OpenBSD's AMD64 port, starting with release 3.6 on November 1, 2004.

It is possible to enter long mode under DOS without a DOS extender,[60] but the user must return to real mode in order to call BIOS or DOS interrupts.

It may also be possible to enter long mode with a DOS extender similar to DOS/4GW, but more complex since x86-64 lacks virtual 8086 mode. DOS itself is not aware of that, and no benefits should be expected unless running DOS in an emulation with an adequate virtualization driver backend, for example: the mass storage interface.

Linux was the first operating system kernel to run the x86-64 architecture in long mode, starting with the 2.4 version in 2001 (preceding the hardware's availability).[61][62] Linux also provides backward compatibility for running 32-bit executables. This permits programs to be recompiled into long mode while retaining the use of 32-bit programs. Several Linux distributions currently ship with x86-64-native kernels and userlands. Some, such as Arch Linux,[63] SUSE, Mandriva, and Debian allow users to install a set of 32-bit components and libraries when installing off a 64-bit DVD, thus allowing most existing 32-bit applications to run alongside the 64-bit OS. Other distributions, such as Fedora, Slackware and Ubuntu, are available in one version compiled for a 32-bit architecture and another compiled for a 64-bit architecture. Fedora and Red Hat Enterprise Linux allow concurrent installation of all userland components in both 32 and 64-bit versions on a 64-bit system.

x32 ABI (Application Binary Interface), introduced in Linux 3.4, allows programs compiled for the x32 ABI to run in the 64-bit mode of x86-64 while only using 32-bit pointers and data fields.[64][65][66] Though this limits the program to a virtual address space of 4 GB it also decreases the memory footprint of the program and in some cases can allow it to run faster.[64][65][66]

64-bit Linux allows up to 128 TiB of virtual address space for individual processes, and can address approximately 64 TiB of physical memory, subject to processor and system limitations.[67]

Mac OS X 10.4.7 and higher versions of Mac OS X 10.4 run 64-bit command-line tools using the POSIX and math libraries on 64-bit Intel-based machines, just as all versions of Mac OS X 10.4 and 10.5 run them on 64-bit PowerPC machines. No other libraries or frameworks work with 64-bit applications in Mac OS X 10.4.[68] The kernel, and all kernel extensions, are 32-bit only.

Mac OS X 10.5 supports 64-bit GUI applications using Cocoa, Quartz, OpenGL, and X11 on 64-bit Intel-based machines, as well as on 64-bit PowerPC machines.[69] All non-GUI libraries and frameworks also support 64-bit applications on those platforms. The kernel, and all kernel extensions, are 32-bit only.

Mac OS X 10.6 is the first version of macOS that supports a 64-bit kernel. However, not all 64-bit computers can run the 64-bit kernel, and not all 64-bit computers that can run the 64-bit kernel will do so by default.[70] The 64-bit kernel, like the 32-bit kernel, supports 32-bit applications; both kernels also support 64-bit applications. 32-bit applications have a virtual address space limit of 4 GiB under either kernel.[71][72]

OS X 10.8 includes only the 64-bit kernel, but continues to support 32-bit applications.

macOS 10.15 includes only the 64-bit kernel and no longer supports 32-bit applications.

The 64-bit kernel does not support 32-bit kernel extensions, and the 32-bit kernel does not support 64-bit kernel extensions.

macOS uses the universal binary format to package 32- and 64-bit versions of application and library code into a single file; the most appropriate version is automatically selected at load time. In Mac OS X 10.6, the universal binary format is also used for the kernel and for those kernel extensions that support both 32-bit and 64-bit kernels.

Solaris 10 and later releases support the x86-64 architecture.

For Solaris 10, just as with the SPARC architecture, there is only one operating system image, which contains a 32-bit kernel and a 64-bit kernel; this is labeled as the "x64/x86" DVD-ROM image. The default behavior is to boot a 64-bit kernel, allowing both 64-bit and existing or new 32-bit executables to be run. A 32-bit kernel can also be manually selected, in which case only 32-bit executables will run. The isainfo command can be used to determine if a system is running a 64-bit kernel.

For Solaris 11, only the 64-bit kernel is provided. However, the 64-bit kernel supports both 32- and 64-bit executables, libraries, and system calls.

x64 editions of Microsoft Windows client and server—Windows XP Professional x64 Edition and Windows Server 2003 x64 Edition—were released in March 2005.[73] Internally they are actually the same build (5.2.3790.1830 SP1),[74][75] as they share the same source base and operating system binaries, so even system updates are released in unified packages, much in the manner as Windows 2000 Professional and Server editions for x86. Windows Vista, which also has many different editions, was released in January 2007. Windows 7 was released in July 2009. Windows Server 2008 R2 was sold in only x64 and Itanium editions; later versions of Windows Server only offer an x64 edition.

Versions of Windows for x64 prior to Windows 8.1 and Windows Server 2012 R2 offer the following:

• 8 TiB of virtual address space per process, accessible from both user mode and kernel mode, referred to as the user mode address space. An x64 program can use all of this, subject to backing store limits on the system, and provided it is linked with the "large address aware" option.[76] This is a 4096-fold increase over the default 2 GiB user-mode virtual address space offered by 32-bit Windows.[77][78]
• 8 TiB of kernel mode virtual address space for the operating system.[77] As with the user mode address space, this is a 4096-fold increase over 32-bit Windows versions. The increased space primarily benefits the file system cache and kernel mode "heaps" (non-paged pool and paged pool). Windows only uses a total of 16 TiB out of the 256 TiB implemented by the processors because early AMD64 processors lacked a CMPXCHG16B instruction.[79]

Under Windows 8.1 and Windows Server 2012 R2, both user mode and kernel mode virtual address spaces have been extended to 128 TiB.[22] These versions of Windows will not install on processors that lack the CMPXCHG16B instruction.

The following additional characteristics apply to all x64 versions of Windows:

• Ability to run existing 32-bit applications (.exe programs) and dynamic link libraries (.dlls) using WoW64 if WoW64 is supported on that version. Furthermore, a 32-bit program, if it was linked with the "large address aware" option,[76] can use up to 4 GiB of virtual address space in 64-bit Windows, instead of the default 2 GiB (optional 3 GiB with /3GB boot option and "large address aware" link option) offered by 32-bit Windows.[80] Unlike the use of the /3GB boot option on x86, this does not reduce the kernel mode virtual address space available to the operating system. 32-bit applications can, therefore, benefit from running on x64 Windows even if they are not recompiled for x86-64.
• Both 32- and 64-bit applications, if not linked with "large address aware," are limited to 2 GiB of virtual address space.
• Ability to use up to 128 GiB (Windows XP/Vista), 192 GiB (Windows 7), 512 GiB (Windows 8), 1 TiB (Windows Server 2003), 2 TiB (Windows Server 2008/Windows 10), 4 TiB (Windows Server 2012), or 24 TiB (Windows Server 2016/2019) of physical random access memory (RAM).[81]
• LLP64 data model: "int" and "long" types are 32 bits wide, long long is 64 bits, while pointers and types derived from pointers are 64 bits wide.
• Kernel mode device drivers must be 64-bit versions; there is no way to run 32-bit kernel mode executables within the 64-bit operating system. User mode device drivers can be either 32-bit or 64-bit.
• 16-bit Windows (Win16) and DOS applications will not run on x86-64 versions of Windows due to the removal of the virtual DOS machine subsystem (NTVDM) which relied upon the ability to use virtual 8086 mode. Virtual 8086 mode cannot be entered while running in long mode.
• Full implementation of the NX (No Execute) page protection feature. This is also implemented on recent 32-bit versions of Windows when they are started in PAE mode.
• Instead of FS segment descriptor on x86 versions of the Windows NT family, GS segment descriptor is used to point to two operating system defined structures: Thread Information Block (NT_TIB) in user mode and Processor Control Region (KPCR) in kernel mode. Thus, for example, in user mode GS:0 is the address of the first member of the Thread Information Block. Maintaining this convention made the x86-64 port easier, but required AMD to retain the function of the FS and GS segments in long mode – even though segmented addressing per se is not really used by any modern operating system.[77]
• Early reports claimed that the operating system scheduler would not save and restore the x87 FPU machine state across thread context switches. Observed behavior shows that this is not the case: the x87 state is saved and restored, except for kernel mode-only threads (a limitation that exists in the 32-bit version as well). The most recent documentation available from Microsoft states that the x87/MMX/3DNow! instructions may be used in long mode, but that they are deprecated and may cause compatibility problems in the future.[80]
• Some components like Microsoft Jet Database Engine and Data Access Objects will not be ported to 64-bit architectures such as x86-64 and IA-64.[82][83]
• Microsoft Visual Studio can compile native applications to target either the x86-64 architecture, which can run only on 64-bit Microsoft Windows, or the IA-32 architecture, which can run as a 32-bit application on 32-bit Microsoft Windows or 64-bit Microsoft Windows in WoW64 emulation mode. Managed applications can be compiled either in IA-32, x86-64 or AnyCPU modes. Software created in the first two modes behave like their IA-32 or x86-64 native code counterparts respectively; When using the AnyCPU mode, however, applications in 32-bit versions of Microsoft Windows run as 32-bit applications, while they run as a 64-bit application in 64-bit editions of Microsoft Windows.