Rumor: the Exynos 2500 might use Google's TPU for AI acceleration

03 April 2024
The TPU is used in Pixel phones for on-device AI tasks. And since Samsung will be fabbing the Tensor G4, there might be similarities between it and the Exynos.

Sort by:

  • S
  • Sir-Shadow
  • 7jQ
  • 11 Apr 2024

Anonymous, 03 Apr 2024Google is not a hardware company.I remember the days people would say Google is nothing more than a search engine.

    Kangal, 06 Apr 2024Feels like Google's Phones release 6-9 months behind t... moreFeels like Google's Phones release 6-9 months behind the other flagships, and offer chipset that is only competitive to 1-2 generations behind. Yet their phones sell at a higher price. For instance;

    - Samsung S23+, with QC 8g2, released Q1 2023, discounted to USD $800
    - Google Pixel 8 Pro, with Tensor G3, released Q4 2023, released for USD $999

    Now Compare Chipsets:
    - 1710 sCPU, 6500 mCPU, 84 iGPU = 2024 Q1 - QC 8g3
    - 1680 sCPU, 6800 mCPU, 82 iGPU = 2023 Q4 - MTK D9300
    - 1670 sCPU, 6300 mCPU, 70 iGPU = 2024 Q1 - Exynos 2400
    - 1550 sCPU, 4950 mCPU, 65 iGPU = 2023 Q1 - QC 8g2
    - 1390 sCPU, 4270 mCPU, 54 iGPU = 2022 Q4 - MTK D9200
    - 1340 sCPU, 3980 mCPU, 45 iGPU = 2022 Q3 - QC 8g1+
    - 1200 sCPU, 3600 mCPU, 43 iGPU = 2022 Q1 - MTK D9000
    - 1190 sCPU, 3650 mCPU, 41 iGPU = 2022 Q1 - QC 8g1
    - 1260 sCPU, 3530 mCPU, 37 iGPU = 2023 Q4 - Tensor G3
    - 1150 sCPU, 3520 mCPU, 31 iGPU = 2022 Q1 - Exynos 2200
    - 1120 sCPU, 3500 mCPU, 28 iGPU = 2021 Q1 - QSD 888
    - 1050 sCPU, 3290 mCPU, 31 iGPU = 2022 Q4 - Tensor G2
    - 1080 sCPU, 3460 mCPU, 27 iGPU = 2021 Q1 - Exynos 2100
    - 1040 sCPU, 2840 mCPU, 30 iGPU = 2021 Q4 - Tensor G1
    -- 820 sCPU, 2830 mCPU, 19 iGPU = 2021 Q1 - MTK D1200

      Feels like Google's Phones release 6-9 months behind the other flagships, and offer chipset that is only competitive to 1-2 generations behind. Yet their phones sell at a higher price. For instance;

      - Samsung S23+, with QC 8g2, released Q1 2023, discounted to USD $800
      - Google Pixel 8 Pro, with Tensor G3, released Q4 2023, released for USD $999

      Now Compare Chipsets:
      (QC 8g3, D9300, Exynos 2400)
      - 1550 sCPU, 4950 mCPU, 65 iGPU = 2023 Q1 - QC 8g2
      - 1390 sCPU, 4270 mCPU, 54 iGPU = 2022 Q4 - MTK D9200
      - 1340 sCPU, 3980 mCPU, 45 iGPU = 2022 Q3 - QC 8g1+
      - 1200 sCPU, 3600 mCPU, 43 iGPU = 2022 Q1 - MTK D9000
      - 1190 sCPU, 3650 mCPU, 41 iGPU = 2022 Q1 - QC 8g1
      - 1260 sCPU, 3530 mCPU, 37 iGPU = 2023 Q4 - Tensor G3
      - 1150 sCPU, 3520 mCPU, 31 iGPU = 2022 Q1 - Exynos 2200
      - 1050 sCPU, 3290 mCPU, 31 iGPU = 2022 Q4 - Tensor G2

        • o
        • ok
        • TjU
        • 04 Apr 2024

        om, 04 Apr 2024Intel and many other need to start releasing SOC with all k... moreSame modell as open source contributor everywhere, the future use to big for single entities

          • o
          • om
          • TjU
          • 04 Apr 2024

          Intel and many other need to start releasing SOC with all kinds of ISA moving away from single companies IP

            Hemedans, 03 Apr 2024Yeah I saw it, maybe because GB6 don't utilise all cor... moreYup, but gb6 is said to be testing more realistic usage. And with similar core and node sd always beat mtk in efficiency like 8+g1 vs d9000 or 865 vs 1200. And i am only talking about just cpu anyway, when gpu is added the driver quality or qualcomm specific optimizations if any will change results somewhat.

              1989 nothinghappened, 03 Apr 2024Click on geekbench 6, look at 4w point.Yeah I saw it, maybe because GB6 don't utilise all cores, but still in real world scenario Cpu+Gpu would rarely surpass 8W, you still need those LW data to get full picture.

                Hemedans, 03 Apr 2024Same socpk are we using same site? With D9300 we only have ... moreClick on geekbench 6, look at 4w point.

                  1989 nothinghappened, 03 Apr 2024Gb6 from socpk d9300 is faster even at 4w. Which light game... moreSame socpk are we using same site? With D9300 we only have two dots because device is not rooted yet

                    Hemedans, 03 Apr 2024The answer we don't know, there is no data at low watt... moreGb6 from socpk d9300 is faster even at 4w. Which light game? Test where?

                      1989 nothinghappened, 03 Apr 2024Its not. Socpk.com has the proof.The answer we don't know, there is no data at low wattage, but judging by fact D9300 is better over 5W and match 8 gen 3 around 5W it will probably lose at low wattage. That's why I heavy games like Genshini D9300 use less power and light games it use more power than 8 gen 3.

                        • T
                        • Tinyman circle 1989
                        • U@6
                        • 03 Apr 2024

                        Dave. H, 03 Apr 2024TPU is very efficient for specific tasks, I'm not sure... moreVery useful for LLM's

                          • D
                          • Dave. H
                          • raQ
                          • 03 Apr 2024

                          TPU is very efficient for specific tasks, I'm not sure if your smartphone needs it

                            Tinyman circle1989, 03 Apr 2024Socpk tests with cold conditions and a fan lolSocpk also has a efficiency chart with power usage and scores. Fan or cold conditions doesnt help there.

                              • T
                              • Tinyman circle1989
                              • U@6
                              • 03 Apr 2024

                              1989 nothinghappened, 03 Apr 2024Its not. Socpk.com has the proof.Socpk tests with cold conditions and a fan lol

                                Tinyman 1989, 03 Apr 20241+5+2 from SD Gen 3 is faster and more efficient than Media... moreIts not. Socpk.com has the proof.

                                  • b
                                  • bl
                                  • S37
                                  • 03 Apr 2024

                                  Tinyman 1989, 03 Apr 20241+5+2 from SD Gen 3 is faster and more efficient than Media... moreMediatek don't know that software is poor thing, AMD have some hint Intel have great envelopes but nothing to show at this point, Nvidia is all about software the famous CUDA stack

                                    • T
                                    • Tinyman 1989
                                    • U@6
                                    • 03 Apr 2024

                                    1989 nothinghappened, 03 Apr 2024Its also slower and less efficient. Even helio g99 gets the... more1+5+2 from SD Gen 3 is faster and more efficient than Mediatek 4+4

                                      • o
                                      • on
                                      • S37
                                      • 03 Apr 2024

                                      The natural part is software hardware engineers have to from the beginning cooperate, but biologists and all other fields of science is the ones who will drive development forward

                                        Tensor G5, maybe?

                                          1 of 2