CPUID

The CPUID opcode is a processor supplementary instruction (its name derived from CPU IDentification) for the x86 architecture allowing software to discover details of the processor. It was introduced by Intel in 1993 when it introduced the Pentium and SL-enhanced 486 processors.^[1]

By using the CPUID opcode, software can determine processor type and the presence of features (like MMX/SSE). The CPUID opcode is 0Fh, A2h (as two bytes, or A20Fh as a single word) and the value in the EAX register, and in some cases the ECX register, specifies what information to return.

History

Prior to the general availability of the CPUID instruction, programmers would write esoteric machine code which exploited minor differences in CPU behavior in order to determine the processor make and model.^[2][3]

Outside the x86 family, developers are mostly still required to use esoteric processes to determine the variations in CPU design that are present. While the CPUID instruction is specific to the x86 architecture, other architectures (like ARM) often provide on-chip registers which can be read to obtain the same sorts of information provided by this instruction.

Calling CPUID

In assembly language the CPUID instruction takes no parameters as CPUID implicitly uses the EAX register to determine the main category of information returned. In Intel's more recent terminology, this is called the CPUID leaf. CPUID should be called with EAX = 0 first, as this will return in the EAX register the highest EAX calling parameter (leaf) that the CPU supports.

To obtain extended function information CPUID should be called with the most significant bit of EAX set. To determine the highest extended function calling parameter, call CPUID with EAX = 80000000h.

CPUID leaves greater than 3 but less than 80000000 are accessible only when the model-specific registers have IA32_MISC_ENABLE.BOOT_NT4 [bit 22] = 0 (which is so by default). As the name suggests, Windows NT4 did not boot properly unless this bit was set,^[4] but later versions of Windows do not need it, so basic leaves greater than 4 can be assumed visible on current Windows systems. As of July 2014, basic valid leaves go up to 14h, but the information returned by some leaves are not disclosed in publicly available documentation, i.e. they are "reserved".

Some of the more recently added leaves also have sub-leaves, which are selected via the ECX register before calling CPUID.

EAX=0: Highest Function Parameter

Here is a list of processors and the highest function supported.

EAX=0: Get vendor ID

This returns the CPU's manufacturer ID string – a twelve-character ASCII string stored in EBX, EDX, ECX (in that order). The highest basic calling parameter (largest value that EAX can be set to before calling CPUID) is returned in EAX.

The following are known processor manufacturer ID strings:

• "AMDisbetter!" – early engineering samples of AMD K5 processor
• "AuthenticAMD" – AMD
• "CentaurHauls" – Centaur (Including some VIA CPU)
• "CyrixInstead" – Cyrix
• "GenuineIntel" – Intel
• "TransmetaCPU" – Transmeta
• "GenuineTMx86" – Transmeta
• "Geode by NSC" – National Semiconductor
• "NexGenDriven" – NexGen
• "RiseRiseRise" – Rise
• "SiS SiS SiS " – SiS
• "UMC UMC UMC " – UMC
• "VIA VIA VIA " – VIA
• "Vortex86 SoC" – Vortex

The following are known ID strings from virtual machines:

• "KVMKVMKVM" – KVM
• "Microsoft Hv" – Microsoft Hyper-V or Windows Virtual PC
• " lrpepyh vr" – Parallels (it possibly should be "prl hyperv ", but it is encoded as " lrpepyh vr")
• "VMwareVMware" – VMware
• "XenVMMXenVMM" – Xen HVM

For instance, on a GenuineIntel processor values returned in EBX is 0x756e6547, EDX is 0x49656e69 and ECX is 0x6c65746e. The following code is written in GNU Assembler for the x86-64 architecture and displays the vendor ID string as well as the highest calling parameter that the CPU supports.

	.data

s0:
	.asciz	"Largest basic function number supported: %i\n"
s1:
	.asciz	"Vendor ID: %.12s\n"

	.text

	.align	32
	.globl	main
main:
	pushq	%rbp
	pushq   %rbx
	movq	%rsp,%rbp
	subq	$16,%rsp

	xorl	%eax,%eax
	cpuid

	movl	%ebx,0(%rsp)
	movl	%edx,4(%rsp)
	movl	%ecx,8(%rsp)

	movq	$s0,%rdi
	movl	%eax,%esi
	xorb	%al,%al
	call	printf

	movq	$s1,%rdi
	movq	%rsp,%rsi
	xorb	%al,%al
	call	printf

	movq	%rbp,%rsp
	popq    %rbx
	popq	%rbp
	movl    $1,%eax
	int     $0x80

EAX=1: Processor Info and Feature Bits

This returns the CPU's stepping, model, and family information in EAX (also called the signature of a CPU), feature flags in EDX and ECX, and additional feature info in EBX.

The format of the information in EAX is as follows:

• 3:0 – Stepping
• 7:4 – Model
• 11:8 – Family
• 13:12 – Processor Type
• 19:16 – Extended Model
• 27:20 – Extended Family

Intel and AMD have suggested applications to display the family of a CPU as the sum of the "Family" and the "Extended Family" fields shown above, and the model as the sum of the "Model" and the 4-bit left-shifted "Extended Model" fields.^[5] If "Family" is different than 6 or 15, only the "Family" and "Model" fields should be used while the "Extended Family" and "Extended Model" bits are reserved. If "Family" is set to 15, then "Extended Family" and the 4-bit left-shifted "Extended Model" should be added to the respective base values, and if "Family" is set to 6, then only the 4-bit left-shifted "Extended Model" should be added to "Model".^[6][7]

The initial APIC-ID (EBX[bits 31:24]) is used to identify the executing logical processor. It can also be identified via the cpuid 0BH leaf (CPUID.0BH.EDX[x2APIC-ID]).

The processor info and feature flags are manufacturer specific but usually the Intel values are used by other manufacturers for the sake of compatibility.

As of January 2011, the standard Intel feature flags are as follows^[8][9]

EAX=2: Cache and TLB Descriptor information

This returns a list of descriptors indicating cache and TLB capabilities in EAX, EBX, ECX and EDX registers.

EAX=3: Processor Serial Number

This returns the processor's serial number. The processor serial number was introduced on Intel Pentium III, but due to privacy concerns, this feature is no longer implemented on later models (PSN feature bit is always cleared). Transmeta's Efficeon and Crusoe processors also provide this feature. AMD CPUs however, do not implement this feature in any CPU models.

For Intel Pentium III CPUs, the serial number is returned in EDX:ECX registers. For Transmeta Efficeon CPUs, it is returned in EBX:EAX registers. And for Transmeta Crusoe CPUs, it is returned in EBX register only.

Note that the processor serial number feature must be enabled in the BIOS setting in order to function.

EAX=4 and EAX=Bh: Intel thread/core and cache topology

These two leaves are used for processor topology (thread, core, package) and cache hierarchy enumeration in Intel multi-core (and hyperthreaded) processors.^[12] As of 2013 AMD does not use these leaves but has alternate ways of doing the core enumeration.^[13]

Unlike most other CPUID leaves, leaf Bh will return different values in EDX depending on which logical processor the CPUID instruction runs; the value returned in EDX is actually the x2APIC id of the logical processor. The x2APIC id space is not continuously mapped to logical processors however; there can be gaps in the mapping, meaning that some intermediate x2APIC ids don't necessarily correspond to any logical processor. Additional information for mapping the x2APIC ids to cores is provided in the other registers. Although the leaf Bh has sub-leaves (selected by ECX as described further below), the value returned in EDX is only affected by the logical processor on which the instruction is running but not by the subleaf.

The processor(s) topology exposed by leaf Bh is a hierarchical one, but with the strange caveat that the order of (logical) levels in this hierarchy doesn't necessarily correspond the order in the physical hierarchy (SMT/core/package). However, every logical level can be queried as an ECX subleaf (of the Bh leaf) for its correspondence to a "level type", which can be either SMT, core, or "invalid". The level id space starts at 0 and is continuous, meaning that if a level id is invalid, all higher level ids will also be invalid. The level type is returned in bits 15:08 of ECX, while the number of logical processors at the level queried is returned in EBX. Finally, the connection between these levels and x2APIC ids is returned in EAX[4:0] as the number of bits that the x2APIC id must be shifted in order to obtain a unique id at the next level.

As an example, a dual-core Westmere processor capable of hyperthreading (thus having two cores and four threads in total) could have x2APIC ids 0, 1, 4 and 5 for its four logical processors. Leaf Bh (=EAX), subleaf 0 (=ECX) of CPUID could for instance return 100h in ECX, meaning that level 0 describes the SMT (hyperthreading) layer, and return 2 in EBX because there are two logical processors (SMT units) per physical core. The value returned in EAX for this 0-subleaf should be 1 in this case, because shifting the aforementioned x2APIC ids to the right by one bit gives a unique core number (at the next level of the level id hierarchy) and erases the SMT id bit inside each core. A simpler way to interpret this information is that the last bit (bit number 0) of the x2APIC id identifies the SMT/hyperthreading unit inside each core in our example. Advancing to subleaf 1 (by making another call to CPUID with EAX=Bh and ECX=1) could for instance return 201h in ECX, meaning that this is a core-type level, and 4 in EBX because there are 4 logical processors in the package; EAX returned could be any value greater than 3, because it so happens that bit number 2 is used to identify the core in the x2APIC id. Note that bit number 1 of the x2APIC id is not used in this example. However EAX returned at this level could well be 4 (and it happens to be so on a Clarkdale Core i3 5x0) because that also gives a unique id at the package level (=0 obviusly) when shifting the x2APIC id by 4 bits. Finally, you may wonder what the EAX=4 leaf can tell us that we didn't find out already. In EAX[31:26] it returns the APIC mask bits reserved for a package; that would be 111b in our example because bits 0 to 2 are used for identifying logical processors inside this package, but bit 1 is also reserved although not used as part of the logical processor identification scheme. In other words, APIC ids 0 to 7 are reserved for the package, even though half of these values don't map to a logical processor.

The cache hierarchy of the processor is explored by looking at the sub-leaves of leaf 4. The APIC ids are also used in this hierarchy to convey information about how the different levels of cache are shared by the SMT units and cores. To continue our example, the L2 cache, which is shared by SMT units of the same core but not between physical cores on the Westmere is indicated by EAX[26:14] being set to 1, while the information that the L3 cache is shared by the whole package is indicated by setting those bits to (at least) 111b. The cache details, including cache type, size, and associativity are communicated via the other registers on leaf 4.

Beware that older versions of the Intel app note 485 contain some misleading information, particularly with respect to identifying and counting cores in a multi-core processor;^[14] errors from misinterpreting this information have even been incorporated in the Microsoft sample code for using cpuid, even for the 2013 edition of Visual Studio,^[15] and also in the sandpile.org page for CPUID,^[16] but the Intel code sample for identifying processor topology^[12] has the correct interpretation, and the current Intel Software Developer’s Manual has more clear language. The (open source) cross-platform production code^[17] from Wildfire Games also implements the correct interpretation of the Intel documentation.

Topology detection examples involving older (pre-2010) Intel processors that lack x2APIC (thus don't support the EAX=Bh leaf) are given in a 2010 Intel presentation.^[18] Beware that using that older detection method on 2010 and newer Intel processors may overestimate the number of cores and logical processors because the old detection method assumes there are no gaps in the APIC id space, and this assumption is violated by some newer processors (starting with the Core i3 5x0 series), but these newer processors also come with an x2APIC, so their topology can be correctly determined using the EAX=Bh leaf method.

EAX=7, ECX=0: Extended Features

This returns extended feature flags in EBX, ECX, and EDX.

EAX=80000000h: Get Highest Extended Function Supported

The highest calling parameter is returned in EAX.

EAX=80000001h: Extended Processor Info and Feature Bits

This returns extended feature flags in EDX and ECX.

AMD feature flags are as follows^[19][20]

EAX=80000002h,80000003h,80000004h: Processor Brand String

These return the processor brand string in EAX, EBX, ECX and EDX. CPUID must be issued with each parameter in sequence to get the entire 48-byte null-terminated ASCII processor brand string.^[5] It is necessary to check whether the feature is supported by the CPU by issuing CPUID with EAX = 80000000h first and checking if the returned value is greater or equal to 80000004h.

.section .data

s0 : .asciz "Processor Brand String: %.48s\n"
err : .asciz "Feature unsupported.\n"

.section .text

.global main
.type main,@function
.align 32
main:
	pushq	%rbp
	movq	%rsp,	%rbp
	subq	$48,	%rsp
	pushq	%rbx
	
	movl	$0x80000000,	%eax
	cpuid
	
	cmpl	$0x80000004,	%eax
	jl	error
	
	movl	$0x80000002,	%esi
	movq	%rsp,	%rdi

.align 16
get_brand:
	movl	%esi,	%eax
	cpuid
	
	movl	%eax,	(%rdi)
	movl	%ebx,	4(%rdi)
	movl	%ecx,	8(%rdi)
	movl	%edx,	12(%rdi)
	
	addl	$1,	%esi
	addq	$16,	%rdi
	cmpl	$0x80000004,	%esi
	jle	get_brand

print_brand:
	movq	$s0,	%rdi
	movq	%rsp,	%rsi
	xorb	%al,	%al
	call	printf
	
	jmp	end

.align 16
error:
	movq	$err,	%rdi
	xorb	%al,	%al
	call	printf

.align 16
end:
	popq	%rbx
	movq	%rbp,	%rsp
	popq	%rbp
	xorl	%eax,	%eax
	ret

EAX=80000005h: L1 Cache and TLB Identifiers

This function contains the processor’s L1 cache and TLB characteristics.

EAX=80000006h: Extended L2 Cache Features

Returns details of the L2 cache in ECX, including the line size in bytes, type of associativity (encoded by a 4 bits) and the cache size.

.section .data

info : .ascii "L2 Cache Size : %u KB\nLine size : %u bytes\n"
.asciz "Associativity : %02xh\n"
err : .asciz "Feature unsupported.\n"

.section .text

.global main
.type main,@function
.align 32
main:
	pushq	%rbp
	movq	%rsp,	%rbp
	pushq	%rbx
	
	movl	$0x80000000,	%eax
	cpuid
	
	cmpl	$0x80000006,	%eax
	jl	error
	
	movl	$0x80000006,	%eax
	cpuid
	
	movl	%ecx,	%eax
	
	movl	%eax,	%edx
	andl	$0xff,	%edx
	
	movl	%eax,	%ecx
	shrl	$12,	%ecx
	andl	$0xf,	%ecx
	
	movl	%eax,	%esi
	shrl	$16,	%esi
	andl	$0xffff,%esi
	
	movq	$info,	%rdi
	xorb	%al,	%al
	call	printf
	
	jmp end
	
.align 16
error:
	movq	$err,	%rdi
	xorb	%al,	%al
	call	printf

.align 16
end:
	popq	%rbx
	movq	%rbp,	%rsp
	popq	%rbp
	xorl	%eax,	%eax
	ret

EAX=80000007h: Advanced Power Management Information

This function provides advanced power management feature identifiers.

EAX=80000008h: Virtual and Physical address Sizes

Returns largest virtual and physical address sizes in EAX.

CPUID usage from high-level languages

This information is easy to access from other languages as well. For instance, the C++ code for gcc below prints the first five values, returned by the cpuid:

#include <iostream>

int main()
{
  int a, b;

  for (a = 0; a < 5; a++)
  {
    __asm__("cpuid;"
            :"=a"(b)                 // EAX into b (output)
            :"0"(a)                  // a into EAX (input)
            :"%ebx","%ecx","%edx");  // clobbered registers

    std::cout << "The code " << a << " gives " << b << std::endl;
  }

  return 0;
}

The equivalent code in C is:

#include <stdio.h>

int main()
{
  int a, b;

  for (a = 0; a < 5; a++)
  {
    __asm__("cpuid"
            :"=a"(b)                 // EAX into b (output)
            :"0"(a)                  // a into EAX (input)
            :"%ebx","%ecx","%edx");  // clobbered registers

    printf("The code %i gives %i\n", a, b);
  }

  return 0;
}

Or, a generally useful C implementation that works on 32- and 64-bit systems:

#include <stdio.h>

int main() {
    int i;
    unsigned int index = 0;
    unsigned int regs[4];
    int sum;
    __asm__ __volatile__(
#if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
        "pushq %%rbx     \n\t" /* save %rbx */
#else
        "pushl %%ebx     \n\t" /* save %ebx */
#endif
        "cpuid            \n\t"
        "movl %%ebx ,%[ebx]  \n\t" /* write the result into output var */
#if defined(__x86_64__) || defined(_M_AMD64) || defined (_M_X64)
        "popq %%rbx \n\t"
#else
        "popl %%ebx \n\t"
#endif
        : "=a"(regs[0]), [ebx] "=r"(regs[1]), "=c"(regs[2]), "=d"(regs[3])
        : "a"(index));
    for (i=4; i<8; i++) {
        printf("%c" ,((char *)regs)[i]);
    }
    for (i=12; i<16; i++) {
        printf("%c" ,((char *)regs)[i]);
    }
    for (i=8; i<12; i++) {
        printf("%c" ,((char *)regs)[i]);
    }
    printf("\n");
}

Microsoft Visual C compiler has builtin function __cpuid() so the cpuid instruction may be embedded without using inline assembly, which is handy since the x86-64 version of MSVC does not allow inline assembly at all. The same program for MSVC would be:

#include <iostream>
#include <intrin.h>

int main()
{
  int b[4];

  for (int a = 0; a < 5; a++)
  {
    __cpuid(b, a);
    std::cout << "The code " << a << " gives " << b[0] << ", " << b[1] << ", " << b[2] << ", " << b[3] << std::endl;
  }

  return 0;
}

For Borland/Embarcadero C compilers (bcc32), native asm function calls are necessary, as there is no asm() implementation. The pseudo code:

  unsigned int a, b, c, d;
  unsigned int InfoType = 0;
  __asm xor EBX, EBX;
  __asm xor ECX, ECX;
  __asm xor EDX, EDX;
  __asm mov EAX, InfoType;
  __asm cpuid;
  __asm mov a, EAX;
  __asm mov b, EBX;
  __asm mov c, ECX;
  __asm mov d, EDX;

Many interpreted or compiled scripting languages are capable of using CPUID via an FFI library. One such implementation shows usage of the Ruby FFI module to execute assembly language that includes the CPUID opcode.

CPU-specific information outside x86

Some of the non-x86 CPU architectures also provide certain forms of structured information about the processor's abilities, commonly as a set of special registers:

• ARM architectures have a CPUID coprocessor register.^[22]
• The IBM System z mainframe processors support a Store CPU ID (STIDP) instruction since the 1983 IBM 4381^[23] for querying the processor ID.^[24]
• The MIPS32 architecture defines a mandatory Processor Identification (PrId) and a series of daisy-chained Configuration Registers.^[25]
• The PowerPC processor has the 32-bit read-only PVR register identifying the processor model in use.^[26]

DSP and transputer-like chip families have not taken up the instruction in any noticeable way, in spite of having (in relative terms) as many variations in design. Alternate ways of silicon identification might be present; for example, DSPs from Texas Instruments contain a memory-based register set for each functional unit that starts with identifiers determining the unit type and model, its ASIC design revision and features selected at the design phase, and continues with unit-specific control and data registers. Access to these areas is performed by simply using the existing load and store instructions; thus, for such devices there is no need for extending the register set for the device identification purposes.

See also

• CPU-Z, a Windows utility that uses CPUID to identify various system settings

References

External links

According to this note, the former Intel app note 485, which was specifically about CPUID, is now incorporated in the Intel® 64 and IA-32 Architectures Software Developer’s Manual. As of July 2014 the manual however still directs the reader to the app note 485 for further information. The latest published version of the app note 485, dating to May 2012, is available via archive.org. App note 485 contains some information that can be and was easily misinterpreted though, particularly with respect to processor topology identification.

The big Intel manuals tend to lag behind the Intel ISA document, available at the top of this page, which is updated even for processors not yet publicly available, and thus usually contains more CPUID bits. For example, as of this writing the ISA book (at revision 19, dated May 2014) documents the CLFLUSHOPT bit in leaf 7, but the big manuals although apparently more up-to-date (at revision 51, dated June 2014) don't mention it.