Skip to content

Latest commit

 

History

History
109 lines (90 loc) · 3.93 KB

Adreno-600.md

File metadata and controls

109 lines (90 loc) · 3.93 KB

Examples

  • 640, 660

  • Oculus Quest 2, Meta Quest Pro, Pico 4 (Adreno 650)

  • Snapdragon XR1 (with Adreno 615)

  • Snapdragon XR2, Snapdragon XR2+ Gen 1 (with Adreno 650)

References

  1. Mesa driver details
  2. Qualcomm Details The Snapdragon 888
  3. Inside the Snapdragon 855’s iGPU
  4. Correction on Qualcomm iGPUs
  5. Vulkan features for Adreno 660, Turnip Adreno 650
  6. Qualcomm Announces Snapdragon 865 and 765(G)
  7. Adreno 660 Benchmarks

Notes

  • Has low resolution Z pass: "During the binning pass, a low resolution Z-buffer is constructed, and can reject LRZ-tile wide contributions to boost binning performance. This LRZ is then used during the rendering pass to reject pixels efficiently before testing against the full resolution Z-buffer."

  • Has forward pass (immediate mode rendering) for fullscreen quad/triangle.

  • Supports wave64 and wave128 execution. In Vulkan has maxSubgroupSize = 128.

  • Framebuffer compression formats:

    • tested: RGBA8, RGBA16_UNorm [7]
    • not supported: RGBA32F [7]

  • Adreno 660 core config:

    • 384 ALU
    • 900 MHz
    • 384 fp64 / clock, 345.6 gflops, 1/4x fp32
    • 1536 fp32 / clock, 1382.4 gflops
    • 3072 fp16 / clock, 2764.8 gflops, 2x fp32

  • Adreno 640 core config:

    • freq: 585 MHz
    • pipelines: 2 (queues?)
    • Shading units: 384
    • Total shaders: 768
    • FLOPS: 898.5 Gflops
    • 384 ALU
    • 9.4 GPixels/s
    • 898.5 FP32 GFLOPS
    • 1797.1 FP16 GFLOPS

  • Adreno 640 config: [3,4]

    • 2MB system level cache (between L2 and RAM).
    • 128KB L2 cache.
    • 1MB gmem
    • 2x Shader Processors.
    • per Shader Processor:
      • 32 KB local memory
      • 16 KB instruction cache
      • 3x Micro Shader Processor Texture Processor (uSPTP).
      • ROPs (2x 64 KB part of gmem?)
    • per uSPTP:
      • 1KB texture cache (L1)
      • 4x texture units
      • 16 KB instruction cache
      • 2x scheduler partitions
    • per uSPTP scheduler partition:
      • Scheduler with 8 entry (?)
      • 32 KB register file (?)
      • 128x FP16
      • 64x FP32
      • 32x INT32
      • 8x IMUL
      • 8x SFU
    • 6x uSPTPs.
    • 12x uSPTP scheduler partitions.
    • 768x FP32 units.

  • Adreno 640 performance: [3]

    • 408 GOPS FP32 Add (898.5 from specs)
    • 407 GOPS FP32 FMA
    • 55 GOPS FP32 Reciprocal
    • 55 GOPS FP32 InvSqrt
    • 432 GOPS FP16 FMA (incorrect [4]) (1797.1 from specs)
    • 461 GOPS FP16 Add (incorrect [4])
    • 84 GOPS INT64 Add
    • 10 GOPS INT64 Mul
    • 188 GOPS INT32 Add
    • 62 GOPS INT32 Mul
    • 218 GOPS INT16 Add
    • 208 GOPS INT16 Mul
    • 229 GOPS INT8 Add
    • 215 GOPS INT8 Mul
    • 88 GB/s local memory bandwidth
    • 37ns local memory latency
    • 192 GB/s texture cache bandwidth
    • 47 cycles L2 cache latency
    • Snapdragon 855’s system level cache offers about 40 GB/s, while main memory bandwidth is a touch below 30 GB/s.
    • In Slingshot Extreme, Adreno 640’s shaders are busy around 90% of the time, while ALU capacity sees 30-40% utilization.

  • new mixed-precision dot product as well as FP16/FP32 wave matrix-multiply instructions. [2]

  • Adreno 640 lets code allocate up to 32 KB of local memory per workgroup, but can only have 64 KB of local memory active at a time across the GPU. Therefore, it probably has one 32 KB local memory instance per 3 SPs. [3]

  • Fully utilizing the FP16 ALU requires wave128. [4]

  • each uSPTP should be capable of tracking 16 wave128 threads, or 8 per ALU partition. [4]