Naming Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715

i-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

Architecture Bifrost Valhall 5th Gen

API Support Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

OpenGL ES 1.1 - 3.2 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Vulkan 1.0 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Vulkan 1.1 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Vulkan 1.2 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Vulkan 1.3 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
OpenCL 1.0, 1.1, 1.2 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
OpenCL 2.0 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
OpenCL 2.1 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
OpenCL 3.0 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓

Core Features Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

ASTC ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
AFBC ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
AFBC – RGBA16 ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
AFRC ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Shader framebuffer access ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Multiple Render Target[1] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
2xMSAA Automatically promoted to 4xMSAA ✓ ✓ ✓ ✓
4xMSAA ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
8xMSAA ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
16xMSAA ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
8-bit integer dot product ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
FP16 / R11G11B10
accelerated blending [5] ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Conservative rasterization ✓ ✓ ✓ ✓ ✓ ✓
Variable Rate Shading ✓ ✓ ✓ ✓ ✓ ✓
Ray tracing ○ ✓ ○ ✓ ○ ✓

Microarchitecture Features Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

Transaction elimination ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
Hidden surface removal ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
IDVS geometry pipeline ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓ ✓
DVS geometry pipeline ✓ ✓ ✓ ✓

○ — Implementation Dependent © Arm Ltd. 2024

 Core Config Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

Arithmetic units 3 3 1/2 2/3 2/3 3 2 2 2 4 3/4 1-4 4 4 4 4 4 4

Warp width 4 4 4 4 8 8 16 16 16 16 16 16 16 16 16 16 16 16
Thread count (max) 384 384 256/512 512/768 512/768 768 1024 1024 1024 2048 1536-2048 512-2048 2048 2048 2048 2048 2048 2048
FP16 operations/cycle 48 48 16/32 32/48 64/96 96 128 128 128 256 192-256 64-256 512 512 512 512 512 512
FP32 operations/cycle 24 24 8/16 16/24 32/48 48 64 64 64 128 96-128 32-128 256 256 256 256 256 256
Fragments/cycle 1 1 1/2 2 2 2 2 2 2 4 4 2-4 4 4 4 4 4 4
Pixels/cycle 1 1 1/2 1/2 2 2 2 2 2 4 4 2-4 4 4 4 4 4 4
Texels/cycle 1 1 1/2 1/2 2 2 4 4 4 8 4-8 2-8 8 8 8 8 8 8
Load/store cache size (bytes) 16K 16K 4K 16K 16K 16K 16K 16K 16K 32K 16K 8K-16K 32K 32K 32K 32K 32K 32K
Texture cache size (bytes) 8K 8K 16K 16K 16K 32K 32K 32K 32K 32K 32K 16K-32K 32K 32K 32K 32K 32K 32K
Tile bits/pixel [2] 128 256 256 256 256 256 256 256 256 256 256 256 256 256 256 256 256 256

Texturing Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

Bilinear samples/cycle 1 1 1/2 1/2 2 2 4 4 4 8 4-8 2-8 8 8 8 8 8 8

Trilinear filtering x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2
Nx anisotropic filtering [3] N/A xN(4) xN xN(4) xN xN xN xN xN xN xN xN xN xN xN xN xN xN
Depth format without reference x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1 x1
Depth format with reference x1 x1 x1 x1 x1 x1 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2
Data size over 32 bits/texel x1 x1 x1 x1 x1 x1 x2 x2 x2 x2 x2 x2 x2 x2 x1(6) x1(6) x1(6) x1(6)
ASTC without EXT_decode_mode x1 x1 x1 x1 x1 x1 x2 x2 x2 x2 x2 x2 x2 x2 x1(6) x1(6) x1(6) x1(6)
3D format with linear filtering x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2
N channel 32bit/channel format
with linear filtering
x4N x4N x4N x4N x4N x4N xN xN xN xN xN xN xN xN xN xN xN xN

N plane YUV format xN xN x1 x1 x1 x1 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2 x2

Bifrost ISA Config Mali-G71 Mali-G72 Mali-G31 Mali-G51 Mali-G52 Mali-G76 1. OpenGL ES has 4 render targets and Vulkan 8
Thread count (max) 384 384 256/512 512/768 768 768 2. Tile storage per pixel may be able to exceed this,
Max work registers (32b) 64 64 64 64 64 64 but with reduced tile size. Theoretical limit is higher
Thread count with 0-32 work registers 384 384 256/512 512/768 768 768 from Mali-G710 onward, but 256 is recommendation
Thread count with 33-64 work registers 384 384 128/256 256/384 384 384 3. Worst-case anisotropic filtering performance with
a MAX_ANISOTROPY = N
Valhall ISA Config Mali-G57 Mali-G77 Mali-G78 Mali-G710 Mali-G510 Mali-G310 Mali-G715 Immortalis-G715 4. Mali-G72 r0p3 / Mali-G51 r1p1 or higher required
Thread count (max) 1024 1024 1024 2048 1536-2048 512-2048 2048 2048 5. All have float blending. Valhall adds hardware
Max work registers (32b) 64 64 64 64 64 64 64 64 acceleration for standard blend operations
Thread count with 0-32 work registers 1024 1024 1024 2048 1536-2048 512-2048 2048 2048 6. Only fp16 and UNORM10 formats fully achieve x1
Thread count with 33-64 work registers 512 512 512 1024 768-1024 256-1024 1024 1024

5th Gen ISA Config Mali-G720 Immortalis-G720 Mali-G725 Immortalis-G925

Thread count (max) 2048 2048 2048 2048

Max work registers (32b) 64 64 64 64
Thread count with 0-32 work registers 2048 2048 2048 2048
Thread count with 33-64 work registers 1024 1024 1024 1024

© Arm Ltd. 2024

This reference sheet covers from the Bifrost Finally, the architecture-specific tables give thread counts NOTE: Mali-G78AE has the same base configurations
Mali-G71, to 5th Gen GPUs up to Immortalis-G925. and registers for the chips. For more on the generations and support as Mali-G78, but includes extra safety features.
of Arm architectures see links below. Mali-G6 series has the same specifications as Mali-G7
The API Support, Core Features and Microarchitecture series for the values in this sheet.
Features tables cover which GPUs support which For a general picture of Arm GPU architectures see:
technologies. For more on given technologies see
links below. Arm GPU Architectures

The Core Config table details the specs of the chips, rather Specific Architecture pages:
than just whether features are available. As such for each
GPU it has threads in a warp, total threads, and operations/ Bifrost (Mali-G71 — Mali-G76)
texels etc per clock cycle, as well as cache sizes. Note Valhall (Mali-G57 — Immortalis-G715)
that for tile write rate on Arm chips this is both fragments
5th Gen (Mali-G720 — Immortalis-G720)
written into the tile and the pixels written back out of the
tile. Thread count is the total shader core hardware capacity; Performance Counters
note that for OpenGL ES only 128 threads are exposed.
For Mali-G310 and Mali-G510 Core Config has ranges For further reference on the technologies mentioned
depending on implementation — please check with device in the sheet, please refer to these webpages:
manufacturer for exact specification.
ASTC (Adaptive Scalable Texture Compression)
For Texturing, to work out cycles/sample for more AFBC (Arm FrameBuffer Compression)
complicated filters than bilinear, apply the multiplications
MSAA (Multi-Sample Anti-Aliasing)
in the tables on top of the bilinear performance to combine
Transaction Elimination
to the required filter. Remember to invert the bilinear
samples/cycle to get cycles/sample. For example, a simple Hidden Surface Removal
trilinear will be 2 x 1 cycles/sample on a Mali-G72, and 2 x IDVS (Index-Driven Vertex Shading)
0.25 cycles/sample on a Mali-G77. To add in 4x anisotropic DVS (Deferred Vertex Shading)
filtering, multiply by a further 4x. Note that anisotropic filter
Shader Framebuffer Access (GLES)
scaling is the worst-case number caused by the maximum
number of sample taps, it will usually be less than this. Shader Framebuffer Access (Vulkan)
Texture performance will differ from Image performance.
Depth performance with/without reference refers to e.g., For free GPU profiling tools, see:
a shadow sampler with reference comparison returning
a weighted bool vs a normal sample returning the actual Arm Mobile Studio
depth value.

© Arm Ltd. 2024