From WikiChip
Difference between revisions of "flops"

(Zen and Zen 2's add units are independent of the FMA units. Peak throughput is 2 x 128/256-bit FMA + 2 x 128/256-bit Addition.)
(Did my math wrong. Fixed peak throughput for Zen and Zen 2)
Line 89: Line 89:
 
| rowspan="3" | {{amd|Zen|l=arch}}<br>{{amd|Zen+|l=arch}} || '''EUs''' || colspan="2" | 2 × 128-bit FMA + 2 × 128-bit Addition || rowspan="3" | {{x86|AVX2}} & FMA (256-bit)
 
| rowspan="3" | {{amd|Zen|l=arch}}<br>{{amd|Zen+|l=arch}} || '''EUs''' || colspan="2" | 2 × 128-bit FMA + 2 × 128-bit Addition || rowspan="3" | {{x86|AVX2}} & FMA (256-bit)
 
|-
 
|-
−
| '''DP''' || 16 FLOPs/cycle || 2 x 4 FLOPs + 2 x 4 FLOPs
+
| '''DP''' || 12 FLOPs/cycle || 2 x 4 FLOPs + 2 x 2 FLOPs
 
|-
 
|-
−
| '''SP''' || 32 FLOPs/cycle || 2 x 8 FLOPs + 2 x 8 FLOPs
+
| '''SP''' || 24 FLOPs/cycle || 2 x 8 FLOPs + 2 x 4 FLOPs
 
|-
 
|-
 
| rowspan="3" | {{amd|Zen 2|l=arch}} || '''EUs''' || colspan="2" | 2 × 256-bit FMA + 2 x 256-bit Addition || rowspan="3" | {{x86|AVX2}} & FMA (256-bit)
 
| rowspan="3" | {{amd|Zen 2|l=arch}} || '''EUs''' || colspan="2" | 2 × 256-bit FMA + 2 x 256-bit Addition || rowspan="3" | {{x86|AVX2}} & FMA (256-bit)
 
|-
 
|-
−
| '''DP''' || 32 FLOPs/cycle || 2 x 8 FLOPs + 2 x 8 FLOPs
+
| '''DP''' || 24 FLOPs/cycle || 2 x 8 FLOPs + 2 x 4 FLOPs
 
|-
 
|-
−
| '''SP''' || 64 FLOPs/cycle || 2 x 16 FLOPs + 2 x 16 FLOPs
+
| '''SP''' || 48 FLOPs/cycle || 2 x 16 FLOPs + 2 x 8 FLOPs
 
|-
 
|-
 
! colspan="5" | [[Centaur]] Microarchitectures
 
! colspan="5" | [[Centaur]] Microarchitectures

Revision as of 16:16, 25 March 2020

Floating-point operations per second (FLOPS) is a measure of compute performance used to quantify the number of floating-point operations a core, machine, or system is capable of in a one second.

Overview

FLOPS are a measure of performance used for comparing the peak theoretical performance of a core, microprocessor, or system using floating point operations. This unit is often used in the field of high-performance computing (e.g., supercomputers) in order to evaluate the peak theoretical performance of various scientific workloads. Traditionally, the FLOPS of a microprocessor could be calculated using the following equation:

Equation FLOPS Subscript core Baseline equals StartFraction FLOPs Over cycle EndFraction times StartFraction cycles Over second EndFraction

With the advent of multi-socket and multi-core architectures, additional levels of explicit parallelism have been introduced resulting in the following modified equation:

Equation FLOPS Subscript node Baseline equals StartFraction FLOPs Over cycle EndFraction times StartFraction cycles Over second EndFraction times StartFraction cores Over node EndFraction

and,

Equation FLOPS Subscript system Baseline equals StartFraction FLOPs Over cycle EndFraction times StartFraction cycles Over second EndFraction times StartFraction cores Over node EndFraction times StartFraction nodes Over system EndFraction

Modern microprocessors exploit data parallelism further through the introduction of various vector extensions such as x86's AVX and ARM's SVE. With those extensions, it's possible to perform multiple floating-point operations within a single instruction. For example, a typical fused multiply-accumulate (FMAC) operation can perform two floating-point operations at once. For a single core, this can be expressed as

Equation FLOPS Subscript core Baseline equals StartFraction instructions Over cycle EndFraction times StartFraction operations Over instruction EndFraction times StartFraction FLOPs Over operation EndFraction times StartFraction cycles Over second EndFraction

And for a full system, this can be extended to:

Equation FLOPS Subscript system Baseline equals StartFraction instructions Over cycle EndFraction times StartFraction operations Over instruction EndFraction times StartFraction FLOPs Over operation EndFraction times StartFraction cycles Over second EndFraction times StartFraction cores Over node EndFraction times StartFraction nodes Over system EndFraction

Nomenclature

  • KiloFLOPS / KFLOPS: 103 FLOPS
  • MegaFLOPS / MFLOPS: 106 FLOPS
  • GigaFLOPS / GFLOPS: 109 FLOPS
  • TeraFLOPS / TFLOPS: 1012 FLOPS
  • PetaFLOPS / PFLOPS: 1015 FLOPS
  • ExaFLOPS / EFLOPS: 1018 FLOPS
  • ZettaFLOPS / ZFLOPS: 1021 FLOPS
  • YottaFLOPS / YFLOPS: 1024 FLOPS

FLOPs by microarchitecture

x86

Microarchitecture FLOPs ISA
Intel Microarchitectures
Core
Penryn
Nehalem 		EUs 1 × 128-bit Multiplication + 1 × 128-bit Addition SSE (128-bit)
DP 4 FLOPs/cycle 2 FLOPs + 2 FLOPs
SP 8 FLOPs/cycle 4 FLOPs + 4 FLOPs
Sandy Bridge
Ivy Bridge 		EUs 1 × 256-bit Multiplication + 1 × 256-bit Addition AVX (256-bit)
DP 8 FLOPs/cycle 4 FLOPs + 4 FLOPs
SP 16 FLOPs/cycle 8 FLOPs + 8 FLOPs
Haswell
Broadwell
Skylake
Kaby Lake
Amber Lake
Coffee Lake
Whiskey Lake 		EUs 2 × 256-bit FMA AVX2 & FMA (256-bit)
DP 16 FLOPs/cycle 2 × 8 FLOPs
SP 32 FLOPs/cycle 2 × 16 FLOPs
Skylake (server) EUs 2 × 512-bit FMA (varies by SKU) AVX-512 & FMA (512-bit)
DP 32 FLOPs/cycle 2 × 16 FLOPs
SP 64 FLOPs/cycle 2 × 32 FLOPs
Intel MIC Microarchitectures
Knights Landing EUs 2 × 512-bit FMA (varies by SKU) AVX-512 & FMA (512-bit)
DP 32 FLOPs/cycle 2 × 16 FLOPs
SP 64 FLOPs/cycle 2 × 32 FLOPs
AMD Microarchitectures
K10 EUs 1 × 128-bit Multiplication + 1 × 128-bit Addition SSE (128-bit)
DP 4 FLOPs/cycle 2 FLOPs + 2 FLOPs
SP 8 FLOPs/cycle 4 FLOPs + 4 FLOPs
Bulldozer
Piledriver
Steamroller
Excavator 		EUs 2 × 128-bit FMA (per two cores) AVX & FMA (128-bit)
DP 8 FLOPs/cycle 2 x 4 FLOPs
SP 16 FLOPs/cycle 2 x 8 FLOPs
Zen
Zen+ 		EUs 2 × 128-bit FMA + 2 × 128-bit Addition AVX2 & FMA (256-bit)
DP 12 FLOPs/cycle 2 x 4 FLOPs + 2 x 2 FLOPs
SP 24 FLOPs/cycle 2 x 8 FLOPs + 2 x 4 FLOPs
Zen 2 EUs 2 × 256-bit FMA + 2 x 256-bit Addition AVX2 & FMA (256-bit)
DP 24 FLOPs/cycle 2 x 8 FLOPs + 2 x 4 FLOPs
SP 48 FLOPs/cycle 2 x 16 FLOPs + 2 x 8 FLOPs
Centaur Microarchitectures
CHA EUs 2 × 256-bit FMA AVX-512 & FMA (512-bit)
DP 16 FLOPs/cycle 2 x 8 FLOPs
SP 32 FLOPs/cycle 2 x 16 FLOPs

ARM

Microarchitecture FLOPs ISA
ARM Microarchitectures
Cortex-A57 EUs 1 × 128-bit FMA ARMv8 (128-bit)
DP 4 FLOPs/cycle 4 FLOPs
SP 8 FLOPs/cycle 8 FLOPs
Cortex-A76 EUs 2 × 128-bit FMA ARMv8 (128-bit)
DP 8 FLOPs/cycle 2 x 4 FLOPs
SP 16 FLOPs/cycle 2 x 8 FLOPs
AppliedMicro/Ampere Computing Microarchitectures
Storm
Shadowcat
Skylark 		EUs 1 × 64-bit FMA ARMv8 (128-bit)
DP 2 FLOPs/cycle 2 FLOPs
SP 4 FLOPs/cycle 4 FLOPs
Cavium Microarchitectures
Vulcan EUs 2 × 128-bit FMA ARMv8 (128-bit)
DP 8 FLOPs/cycle 2 x 4 FLOPs
SP 16 FLOPs/cycle 2 x 8 FLOPs
Samsung Microarchitectures
M1
M2 		EUs 1 × 128-bit FMA + 1 × 128-bit Addition ARMv8 (128-bit)
DP 6 FLOPs/cycle 1 x 4 FLOPs + 1 x 2 FLOPs
SP 12 FLOPs/cycle 1 x 8 FLOPs + 1 x 4 FLOPs
M3 EUs 3 × 128-bit FMA ARMv8 (128-bit)
DP 12 FLOPs/cycle 3 x 4 FLOPs
SP 24 FLOPs/cycle 3 x 8 FLOPs
Phytium Microarchitectures
Xiaomi EUs 1 × 128-bit FMA ARMv8 (128-bit)
DP 4 FLOPs/cycle 1 x 4 FLOPs
SP 8 FLOPs/cycle 1 x 8 FLOPs
HiSilicon Microarchitectures
TaiShan v110 EUs 1 × 128-bit FMA ARMv8 (128-bit)
DP 4 FLOPs/cycle 1 x 4 FLOPs
SP 8 FLOPs/cycle 1 x 8 FLOPs

See also

Retrieved from "https://en.wikichip.org/w/index.php?title=flops&oldid=96501"