Jump to content

Mali (processor)

From Wikipedia, the free encyclopedia
(Redirected from Mali (GPU))
Mali
ARM Cortex A57 A53 big.LITTLE SoC with a Mali-T624 GPU
Release date2005
Architecture
  • Utgard
  • Midgard
  • Bifrost
  • Valhall
ModelsSee Variants
Cores1 to 32 cores
Fabrication process4 to 40 nm
API support
Direct3D9 to 12
OpenCL1.1 to 3.0
OpenGL2.0 to 3.0
Vulkan1.0 to 1.3

The Mali and Immortalis series of graphics processing units (GPUs) and multimedia processors are semiconductor intellectual property cores produced by Arm Holdings for licensing in various ASIC designs by Arm partners.

Mali GPUs were developed by Falanx Microsystems A/S, which was a spin-off of a research project from the Norwegian University of Science and Technology.[1] Arm Holdings acquired Falanx Microsystems A/S on June 23, 2006 and renamed the company to Arm Norway.[2]

It was originally named Malaik, but the team shortened the name to Mali, Serbo-Croatian for "small", which was thought to be fitting for a mobile GPU.[3]

On June 28, 2022, Arm announced their Immortalis series of GPUs with hardware-based ray tracing support.[4]

Graphics processors

[edit]

Utgard

[edit]

In 2005, Falanx announced their Utgard GPU Architecture, the Mali-200 GPU.[5] Arm followed up with the Mali-300, Mali-400, Mali-450, and Mali-470. Utgard was a non-unified GPU (discrete pixel and vertex shaders).[1]

Comparison of Mali Utgard graphics processing units
Model Launch date Type EUs/Shader core count Core clock rate

(MHz)

L2 cache size Fillrate GFLOPS
(per core)
OpenGL ES
M△/s GT/s (GP/s)
Mali-55/110[link removed] 2005 Fixed function pipeline[6] 1 2.8 0.1 ? 1.1
Mali-200 2007[7] Programmable pipeline[6] 1 5 ? 0.2 2.0
Mali-300 2010[8] 1 500 8 KiB 55 0.5 5
Mali-400 MP 2008 1–4 200–600 8–256 KiB 55 0.5 1.2–5.4
Mali-450 MP 2012 1–8 300–750 8–512 KiB 142 2.6 4.5–11.9
Mali-470 MP 2015 1–4 250–650 8–256 KiB 71 0.65 8–20.8

Midgard

[edit]

1st generation

[edit]

On November 10, 2010, Arm announced their Midgard 1st gen GPU Architecture, including the Mali-T604 and later the Mali-T658 GPU in 2011.[9][10][11][12] Midgard uses a Hierarchical Tiling system.[1]

2nd generation

[edit]

On August 6, 2012, Arm announced their Midgard 2nd gen GPU Architecture, including the Mali-T678 GPU.[13] Midgard 2nd gen introduced Forward Pixel Kill.[1][14]

3rd generation

[edit]

On October 29, 2013, Arm announced their Midgard 3rd gen GPU Architecture, including the Mali-T760 GPU.[15][1][16][17][18]

4th generation

[edit]

On October 27, 2014, Arm announced their Midgard 4th gen GPU Architecture, including the Mali-T860, Mali-T830, Mali-T820. Their flagship Mali-T880 GPU was announced on February 3, 2015. New microarchitectural features include:[19]

  • Up to 16 cores for the Mali-T880, with 256KB – 2MB L2 cache

Bifrost

[edit]

1st generation

[edit]

On May 27, 2016, Arm announced their Bifrost GPU Architecture, including the Mali-G71 GPU. New microarchitectural features include:[20][21]

  • Unified shaders with quad vectorization
  • Scalar ISA
  • Clauses execution
  • Full cache coherency
  • Up to 32 cores for the Mali-G71, with 128KB – 2MB L2 cache
  • Arm claims the Mali-G71 has 40% more performance density and 20% better energy efficiency than the Mali-T880

2nd generation

[edit]

On May 29, 2017, Arm announced their Bifrost 2nd gen GPU Architecture, including the Mali-G72 GPU. New microarchitectural features include:[22][23]

  • Arithmetic optimizations and increased caches
  • Up to 32 cores for the Mali-G72, with 128KB – 2MB L2 cache
  • Arm claims the Mali-G72 has 20% more performance density and 25% better energy efficiency than the Mali-G71

3rd generation

[edit]

On May 31, 2018, Arm announced their Bifrost 3rd gen GPU Architecture, including the Mali-G76 GPU. New microarchitectural features include:[24][25]

  • 8 execution lanes per engine (up from 4). Doubled pixel and texel throughput
  • Up to 20 cores for the Mali-G76, with 512KB – 4MB L2 cache
  • Arm claims the Mali-G76 has 30% more performance density and 30% better energy efficiency than the Mali-G72

Valhall

[edit]

1st generation

[edit]

On May 27, 2019, Arm announced their Valhall GPU Architecture, including the Mali-G77 GPU, and in October Mali-G57 GPUs. New microarchitectural features include:[26][27][28]

  • New superscalar engine
  • Simplified scalar ISA
  • New dynamic scheduling
  • Up to 16 cores for the Mali-G77, with 512KB – 2MB L2 cache
  • Arm claims the Mali-G77 has 30% more performance density and 30% better energy efficiency than the Mali-G76

2nd generation

[edit]

On May 26, 2020, Arm announced their Valhall 2nd Gen GPU Architecture, including the Mali-G78. New microarchitectural features include:[29][30][31]

  • Asynchronous clock domains
  • New FMA units and increase Tiler throughput
  • Up to 24 cores for the Mali-G78, with 512KB – 2MB L2 cache
  • Arm Frame Buffer Compression (AFBC)
  • Arm claims the Mali-G78 has 15% more performance density and 10% better energy efficiency than the Mali-G77

3rd generation

[edit]

On May 25, 2021, Arm announced their Valhall 3rd Gen GPU Architecture (as part of TCS21), including the Mali-G710, Mali-G510, and Mali-G310 GPUs. New microarchitectural features include:[32][33][34]

  • Larger shader cores (2x compared to Valhall 2nd Gen)
  • New GPU frontend, Command Stream Frontend (CSF) replaces the Job Manager
  • Up to 16 cores for the Mali-G710, with 512KB – 2MB L2 cache
  • Arm claims the Mali-G710 has 20% more performance density and 20% better energy efficiency than the Mali-G78

4th generation

[edit]

On June 28, 2022, Arm announced their Valhall 4th Gen GPU Architecture (as part of TCS22), including the Immortalis-G715, Mali-G715, and Mali-G615 GPUs. New microarchitectural features include:[4][35]

  • Ray Tracing support (hardware-based)
  • Variable Rate Shading[36]
  • New Execution Engine, with doubled the FMA block, Matrix Multiply instruction support, and PPA improvements
  • Arm Fixed Rate Compression (AFRC)
  • Arm claims the Immortalis-G715 has 15% more performance & 15% better energy efficiency than the Mali-G710[37]

5th generation

[edit]

On May 29, 2023, Arm announced their 5th Gen Arm GPU Architecture (as part of TCS23), including the Immortalis-G720, Mali-G720 and Mali-G620 GPUs.[38][39][40] New microarchitectural features include:[41]

  • Deferred vertex shading (DVS) pipeline
  • Arm claims the Immortalis-G720 has 15% more performance and uses up to 40% less memory bandwidth than the Immortalis-G715

Technical details

[edit]

Like other embedded IP cores for 3D rendering acceleration, the Mali GPU does not include display controllers driving monitors, in contrast to common desktop video cards. Instead, the Mali ARM core is a pure 3D engine that renders graphics into memory and passes the rendered image over to another core to handle display.

ARM does, however, license display controller SIP cores independently of the Mali 3D accelerator SIP block, e.g. Mali DP500, DP550 and DP650.[42]

ARM also supplies tools to help in authoring OpenGL ES shaders named Mali GPU Shader Development Studio and Mali GPU User Interface Engine.

Display controllers such as the ARM HDLCD display controller are available separately.[43]

Variants

[edit]

The Mali core grew out of the cores previously produced by Falanx and currently constitute:[44]

Model Microarchi-
tecture
Type Launch date EUs/Shader core count Shading Units (per core) Total Shaders Fab (nm) Die size
(mm2)
Core clock rate (MHz) L2 cache size (KiB) Fillrate GFLOPS
(per core)
GFLOPS
(total)
API (version)
M△/s GT/s (GP/s) Vulkan OpenGL ES OpenCL
Mali-T604[45] Midgard 1st gen Unified shader model +

SIMD ISA

Nov 2010[46] 1–4 8 8–32 32
28
? 533 32–256 133 0.6 @ 600 MHz 9.6 @ 600 MHz 9.6–38.4 @ 600 MHz N/a 3.1 1.1 Full Profile
Mali-T658[45] Nov 2011[47] 1–8 16 16–128 ? ? ? 19.2 @ 600 MHz 19.2–153.6 @ 600 MHz
Mali-T622 Midgard 2nd gen Jun 2013[48] 1–2 4 4–8 32
28
? 533 ? ? 4.8 @ 600 MHz 4.8–9.6 @ 600 MHz
Mali-T624 Aug 2012 1–4 8 8–32 ? 533–600 ? ? 9.6 @ 600 MHz 9.6–38.4 @ 600 MHz
Mali-T628 1–8 16 16–128 ? 533–695 ? ? 19.2 @ 600 MHz 19.2–153.6 @ 600 MHz
Mali-T678[49] 1–8 28 ? ? ? ?
Mali-T720 Midgard 3rd gen Oct 2013 1–8 10 10–80 28
14
10
? 400–700 600 (MP8@
600 MHz)
0.6 @ 600 MHz 12 @ 600 MHz 12–96 @ 600 MHz
Mali-T760 1–16 14 14–224 28
20
14
1.75 mm2 per shader core at 14 nm[50] 600–772 256–2048[51] 1300 16.8 @ 600 MHz 16.8–268.8 @ 600 MHz 1.0[52] 3.2[53] 1.2 Full Profile
Mali-T820 Midgard 4th gen Q4 2015 1–4 8 8–32 28 ? 600 32–256[51] 400 9.6 @ 600 MHz 9.6–38.4 @ 600 MHz
Mali-T830 16 16–64 28
16
14
? 600–950 400 19.2 @ 600 MHz 19.2–76.8 @ 600 MHz
Mali-T860 1–16 14 14–224 ? 350–700 256–2048[51] 1300 16.8 @ 600 MHz 16.8–268.8 @ 600 MHz
Mali-T880 Q2 2016 1–16 21 21–351 20
16
14
? 650–1000 1700 25.2 @ 600 MHz 25.2–403.2 @ 600 MHz
Mali-G31 Bifrost 1st gen Unified shader model + Unified memory +

scalar, clause-based ISA

Q1 2018 1–6[54] 4 or 8 4–48 28
12
? 650 32–512 0.5 @ 1000 MHz 8–16 @ 1000 MHz 48–576 @ 1000 MHz 1.2[55] 2.0 Full Profile
Mali-G51[56] Q4 2016 1–6[57] 8 or 12 8–72 28
16
14
12
10
? 1000 16–24 @ 1000 MHz 16–144 @ 1000 MHz 1.0[58]
Mali-G71 Q2 2016 1–32 12 12–384 16
14
10
? 546–1037 128–2048 1850 1 @ 1000 MHz 24 @ 1000 MHz 24–768 @ 1000 MHz
Mali-G52 Bifrost 2nd gen Q1 2018 1–6 16 or 24 16–144 16
12
8
7
? 850 32-512 2 @ 1000 MHz 32–48 @ 1000 MHz 32–288 @ 1000 MHz 2.1 Full Profile
Mali-G72 Q2 2017 1–32 12 12–384 16
12
10
1.36 mm2 per shader core at 10 nm[59] 572–1050 128–2048 1 @ 1000 MHz 24 @ 1000 MHz 24–768 @ 1000 MHz 2.0 Full Profile
Mali-G76 Bifrost 3rd gen Q2 2018 4–20 24 96–480 12
8
7
? 600–800 512–4096 ? 2 @ 1000 MHz 2 @ 1000 MHz 48 @ 1000 MHz 192–960 @ 1000 MHz 1.1 2.1 Full Profile
Mali-G57 Valhall 1st gen Superscalar engine + Unified memory +

simplified scalar ISA

Q2 2019 1–6 32 32–192 12
7
6
? 950[60] 64–512 ? 4 @ 1000 MHz 64 @ 1000 MHz 64–384 @ 1000 MHz
Mali-G77 7–16 224–512 7
6
? 695–850 512–2048 ? 448–1024 @ 1000 MHz
Mali-G68 Valhall 2nd gen Q2 2020 1–6 32–192 6
5
3
64–384 @ 1000 MHz 1.2 3.0 Full Profile
Mali-G78 7–24 224–768 5 759-848 448–1536 @ 1000 MHz
Mali-G310 Valhall 3rd gen Q2 2021 1 16 or 32 or 64 16–64 6
5
4
256–1024 2, 4 or 8 @ 1000 MHz 2 or 4 @ 1000 MHz 32–128 @ 1000 MHz
Mali-G510 2–6 48 or 64 96–384 4 or 8 @ 1000 MHz 4 @ 1000 MHz 96–128 @ 1000 MHz 192–768 @ 1000 MHz
Mali-G610 1–6 64 64–384 512–2048 8 @ 1000 MHz 128 @ 1000 MHz 128–768 @ 1000 MHz
Mali-G710 7–16 448–1024 650,850
900
2648 896–2048 @ 1000 MHz
Mali-G615 Valhall 4th gen Q2 2022 1–6 128 128–768 4 256 @ 1000 MHz 256–1536 @ 1000 MHz 1.3[61]
Mali-G715 7–9 896–1152 1792–2304 @ 1000 MHz
Immortalis-G715 10–16 1280–2048 2560–4096 @ 1000 MHz
Mali-G620 5th Gen[62] Deferred Vertex Shading (DVS) Q2 2023 1–5 128–640 256–1024 256–1280 @ 1000 MHz
Mali-G720 6–9 768–1152 512–2048 1536–2304 @ 1000 MHz
Immortalis-G720 Q4 2023 10–16 1280–2048 2560–4096 @ 1000 
MHz
Mali-G625 Q2 2024 1–5 128–640 4
3
256–1024 256–1280 @ 1000 MHz
Mali-G725 6–9 768–1152 512–4096 1536–2304 @ 1000 MHz
Immortalis-G925 10–24 1280–3072 2560–6144 @ 1000 
MHz
Mali G1-Pro Q3 2025 1–5 128–640 3 512–2048 256–1280 @ 1000 MHz 1.4
Mali G1-Premium 6–9 768–1152 512–4096 1536–2304 @ 1000 MHz
Mali G1-Ultra 10–24 1280–3072 2560–6144 @ 1000 
MHz
Model Microarchi-
tecture
Type Launch date EUs/Shader core count Shading Units (per core) Total Shaders Fab

(nm)

Die size
(mm2)
Core clock rate (MHz) Max L2 cache size (KiB) Fillrate (per core) FP32 GFLOPS
(per core)
GFLOPS
(total)
Vulkan Open
GL/ES
Open
CL

Some microarchitectures (or just some chips?) support cache coherency for the L2 cache with the CPU.[63][64]

Adaptive Scalable Texture Compression (ASTC) is supported by Mali-T620, T720/T760, T820/T830/T860/T880[65] and Mali-G series.

Implementations

[edit]

The Mali GPU variants can be found in the following systems on chips (SoCs):