拆解 AI TOPS:人工智慧晶片關鍵指標與 TOPS 晶片比較表

(圖說:美味的背後有多少辛苦的前置準備? 拍攝於 Le Bouchon Ogasawara 餐廳,渋谷,東京。圖片來源:Ernest。)



摘要 tl;dr

  • TOPS (Trillions of Operations Per Second) 是衡量 AI 晶片和 NPU 晶片運算能力的關鍵指標,反映處理器每秒可執行的萬億次運算數。
  • 用「煎蛋」類比舉例直觀理解 TOPS:普通 CPU 如同時只能煎一個蛋的廚師,而高 TOPS 效能的 AI 晶片則是能同時煎超級無敵多個蛋的超級廚師。
  • TOPS 是比較 AI 晶片性能的重要參考,但評估 AI 硬體時還需考慮能源效率、記憶體頻寬等因素,且 TOPS 值通常反映理論峰值,實際性能還需要綜合其他指標數據作為適合應用場景的判斷。

什麼是 TOPS (簡單生活版)

TOPS,全稱為 Trillions Operations Per Second (每秒兆次運算、每秒萬億次運算),是衡量人工智慧 (AI) 晶片或神經網路處理器 (NPU) 計算能力的重要指標。TOPS 用來表示一個處理器最高每秒能夠執行的運算數量,以萬億次來計算。未來如果運算能力繼續提升,則開頭的 T 將會換成其他更大的計量單位。

我們可以用日常生活中的例子來解釋,以更直觀地理解 TOPS

想像 AI 運算是煎蛋過程數據是那個被加熱的蛋

一個普通的廚師 (普通處理器、CPU) 可能一次只能煎一個蛋,而一個超級廚師 (AI 晶片) 可能同時煎 1 兆個蛋!TOPS 就像是衡量這個「超級廚師」能力的指標,告訴我們它每秒能「煎」多少個「數據蛋」。

TOPS 是理解和比較 AI 晶片性能的重要指標之一,但不是唯一。

在評估 AI 硬體或 AI 手機、AI 電腦時,請記得同時考慮其他因素,如能源使用效率、記憶體頻寬、軟體生態系統等。使用 TOPS 可以幫助我們比較不同 AI 晶片的計算能力,為選擇適合特定應用的 AI 硬體設備提供一個參考點。


什麼是 TOPS (硬要深入版)

在深入理解 TOPS 之前,我們需要先明白什麼是「operation」 (運算) :

在數位電路和計算機科學中,一個「operation」通常指的是一個基本的數學或邏輯運算。對於 AI 晶片或 NPU 來說,這些運算主要包括:

  1. 浮點運算:如加法、減法、乘法和除法。
  2. 矩陣運算:大規模的矩陣乘法是深度學習中最常見的運算之一。
  3. 向量運算:包括內積 (數量積)、外積 (向量積) 等。
  4. 激勵函數:如 ReLU、Sigmoid、Tanh 等。
  5. 卷積運算:在卷積神經網路 (CNN) 中廣泛使用。

這些運算通常以 FP32 (32 位元浮點數) 或 FP16 (16 位元浮點數) 格式進行。有些 AI 晶片還支持 INT8 (8 位元整數) 等低精度格式,以提高效能和降低能耗,通常用於推理。

TOPS 的計算方式可以簡化為:

TOPS = (每個時鐘周期的運算數) × (時鐘頻率) / 1 兆

例如,如果一個 AI 晶片在每個時鐘周期可以執行 1000 個運算,時鐘頻率為 1GHz,那麼它的理論峰值性能就是 1 TOPS。

1000 運算/周期×1GHz = 1000×10^9 運算/秒 = 10^12 運算/秒 = 1 TOPS

理解 TOPS 時請注意以下幾點:

  1. TOPS 通常表示理論峰值性能,實際性能可能會因為記憶體頻寬、晶片架構等因素而有所不同。
  2. 不同類型的運算 (如 FP32、FP16、INT8) 的 TOPS 值可能會不同。
  3. TOPS 值高不一定意味著在所有 AI 任務上都有更好的表現,因為實際性能還取決於軟體優化、特定任務的特性等。

TOPS 比較表

(主要可看「INT8 Ops」欄位。可以左右滑動看到更多比較數據)

INT8 OpsFP32 FLOpsCompany NameTypeTarget MarketProduct FamilyProduct NameProduct GenerationCode NameRelease YearFirst Used OnFab ProcessCPUGPUNPUMemory TechMemory BandwidthTDP BaseRemark
73 TOPSn/aAMDSoCPCRyzen AI 300Ryzen AI 9 365n/aStrix Point2024n/aTSMC 4nm FinFETn/aAMD Radeon™ 880Mn/aDDR5-5600 or LPDDR5X-7500n/a28.0- Total 73 TOPS (50 TOPS from NPU).
80 TOPSn/aAMDSoCPCRyzen AI 300Ryzen AI 9 HX 370n/aStrix Point2024n/aTSMC 4nm FinFETn/aAMD Radeon™ 890Mn/aDDR5-5600 or LPDDR5X-7500n/a28.0- Total 80 TOPS (50 TOPS from NPU).
50 TOPSn/aAMDNPUn/aRyzenXDNA 2n/aAI2024Ryzen AI 9 HX 370n/an/an/an/an/an/an/an/a
1961.2 TOPS 3922.3 TOPS (with Sparsity)122.6 TFLOPSAMDGPUDatacenterAMD Data Center GPUs (AMD Instinct)MI300An/an/a2023n/an/an/an/an/aHBM35300 GB/s550.0n/a
2614.9 TOPS 5229.8 TOPS (with Sparsity)163.4 TFLOPSAMDGPUDatacenterAMD Data Center GPUs (AMD Instinct)MI300Xn/an/a2023n/aXCD: TSMC N5 IOD: TSMC N6n/an/an/aHBM35300 GB/s750.0n/a
2614.9 TOPS 5229.8 TOPS (with Sparsity)163.4 TFLOPSAMDGPUDatacenterAMD Data Center GPUs (AMD Instinct)MI325Xn/an/a2024n/aXCD: TSMC N5 IOD: TSMC N6n/an/an/aHBM3E6000 GB/s750.0n/a
n/an/aARMIPn/aNeoverseNeoverse E1n/an/an/an/an/an/an/an/an/an/an/an/a
n/an/aARMIPn/aNeoverseNeoverse N1n/aAres2019Ampere Altra, AWS Graviton2n/an/an/an/an/an/an/an/a
n/an/aARMIPDatacenter (Infrastructure Processor)NeoverseNeoverse N2n/aPerseus2020Microsoft Azure Cobalt 100n/an/an/an/an/an/an/an/a
n/an/aARMIPDatacenter (Infrastructure Processor)NeoverseNeoverse N3n/aHermesn/an/an/an/an/an/an/an/an/an/a
n/an/aARMIPDatacenter (Infrastructure Processor)NeoverseNeoverse V1n/aZeus2020AWS Graviton3n/an/an/an/an/an/an/a- first announcements coming out of Arm’s TechCon convention 2018 in San Jose.
n/an/aARMIPDatacenter (Infrastructure Processor)NeoverseNeoverse V2n/an/a2022NVIDIA Grace, AWS Graviton4, Google Axionn/an/an/an/an/an/an/an/a
n/an/aARMIPDatacenter (Infrastructure Processor)NeoverseNeoverse V3n/aPoseidonn/an/an/an/an/an/an/an/an/an/a
825 TOPS ???n/aAlibabaSoCDatacenter (AI inference)Hanguang 含光Hanguang 8001n/a2019n/aTSMC 12nmn/an/an/an/an/a280.0- 16x PCIe gen4 - SRAM, No DDR
n/an/aAlibabaSoCDatacenter (Infra)Yitian 倚天Yitian 7101n/a2021Alibaba ECS g8mN5128 Neoverse N2 coren/an/an/an/an/an/a
n/an/aAmazonSoCDatacenter (Infra) (Scale out)AWS GravitonGraviton1Alpine2018Amazon EC2 A1TSMC 16nmCortex A72n/an/aDDR4-160051.2 GB/s95.0- 32 lanes of PCIe gen3
n/an/aAmazonSoCDatacenter (Infra) (General Purpose)AWS GravitonGraviton 22Alpine+2019Amazon EC2 M6g, M6gd, C6g, C6gd, C6gn, R6g, R6gd, T4g, X2gd, G5g, Im4gn, Is4gen, I4gTSMC 7nm128 Neoverse N1 coren/an/aDDR4-3200204.8 GB/s110.0- 64 lanes of PCIe gen4
n/an/aAmazonSoCDatacenter (Infra) (ML, HPC, SIMD)AWS GravitonGraviton 33n/a2021Amazon EC2 C7g, M7g, R7g; with local disk: C7gd, M7gd, R7gdTSMC 5nm64 Neoverse V1 coren/an/aDDR5-4800307.2 GB/s100.0- 32 lanes of PCIe gen5
n/an/aAmazonSoCDatacenter (Infra)AWS GravitonGraviton 3E3n/a2022Amazon EC2 C7gn, HPC7gn/a64 Neoverse V1 coren/an/an/an/an/an/a
n/an/aAmazonSoCDatacenter (Infra) (Scale up)AWS GravitonGraviton 44n/a2023Amazon EC2 R8gn/a96 Neoverse V2 coren/an/aDDR5-5600537.6 GB/sn/a- 96 lanes of PCIe gen5
63.3 TOPS0.97 TFLOPSAmazonSoCDatacenter (AI inference)AWS InferertiaInferertia 11n/a2018Amazon EC2 Inf1TSMC 16nm16 NeuroCore v1n/an/an/a50 GB/sn/an/a
380 TOPS2.9 TFLOPSAmazonSoCDatacenter (AI inference)AWS InferertiaInferertia 22n/a2022Amazon EC2 Inf2TSMC 5nm24 NeuroCore v2n/an/an/a820 GB/sn/an/a
380 TOPS2.9 TFLOPSAmazonSoCDatacenter (AI train)AWS TrainiumTrainium 11n/a2020Amazon EC2 Trn1TSMC 7nm32 NeuroCore v2n/an/an/a820 GB/sn/an/a
861 TOPS6.57 TFLOPSAmazonSoCDatacenter (AI train)AWS TrainiumTrainium 22n/a2023Amazon EC2 Trn2TSMC 4nm64 NeuroCore v2n/an/an/a4,096 GB/sn/an/a
11 TOPS748.8 GFLOPSAppleSoCMobileAA14 Bionicn/aAPL1W012020iPhone 12TSMC N5Firestorm + Icestormn/an/aLPDDR4X-426634.1 GB/sn/an/a
15.8 TOPS1.37 TFLOPSAppleSoCMobileAA15 Bionicn/aAPL1W072021iPhone 13TSMC N5PAvalanche + Blizzardn/an/aLPDDR4X-426634.1 GB/sn/an/a
17 TOPS1.789 TFLOPSAppleSoCMobileAA16 Bionicn/aAPL1W102022iPhone 14TSMC N4PEverest + Sawtoothn/an/aLPDDR5-640051.2 GB/sn/a- 6GB LPDDR5
35 TOPS2.147 TFLOPSAppleSoCMobileAA17 Pron/aAPL1V022023iPhone 15 Pro, iPhone 15 Pro MaxTSMC N3B6 cores (2 performance + 4 efficiency)Apple-designed 6-core16-core Neural EngineLPDDR5-640051.2 GB/sn/a- 8GB LPDDR5
35 TOPSn/aAppleSoCMobileAA18n/an/a2024iPhone 16TSMC N3P6 cores (2 performance + 4 efficiency)Apple-designed 5-core16-core Neural Enginen/an/an/an/a
35 TOPSn/aAppleSoCMobileAA18 Pron/an/a2024iPhone 16 ProTSMC N3P6 cores (2 performance + 4 efficiency)Apple-designed 6-core16-core Neural Enginen/an/an/an/a
11 TOPS2.6 TFLOPSAppleSoCMobile, PCMM1n/aAPL11022020n/aTSMC N5high-performance “Firestorm” + energy-efficient “Icestorm”n/an/aLPDDR4X-426668.3 GB/sn/an/a
11 TOPS10.4 TFLOPSAppleSoCMobile, PCMM1 Maxn/aAPL11052021n/aTSMC N5n/an/an/aLPDDR5-6400409.6 GB/sn/an/a
11 TOPSn/aAppleSoCMobile, PCMM1 Pron/aAPL11032021n/aTSMC N5n/an/an/aLPDDR5-6400204.8 GB/sn/an/a
22 TOPS21 TFLOPSAppleSoCMobile, PCMM1 Ultran/aAPL1W062022n/aTSMC N5The M1 Ultra consists of two M1 Max units connected with UltraFusion Interconnect with a total of 20 CPU cores and 96 MB system level cache (SLC).n/an/aLPDDR5-6400819.2 GB/sn/an/a
15.8 TOPS2.863 TFLOPS, 3.578 TFLOPSAppleSoCMobile, PCMM2n/aAPL11092022n/aTSMC N5Phigh-performance @3.49 GHz “Avalanche” + energy-efficient @2.42 GHz “Blizzard”n/an/aLPDDR5-6400102.4 GB/sn/an/a
15.8 TOPS10.736 TFLOPS, 13.599 TFLOPSAppleSoCMobile, PCMM2 Maxn/aAPL11112023n/aTSMC N5Pn/an/an/aLPDDR5-6400409.6 GB/sn/an/a
15.8 TOPS5.726 TFLOPS, 6.799 TFLOPSAppleSoCMobile, PCMM2 Pron/aAPL11132023n/aTSMC N5Pn/an/an/aLPDDR5-6400204.8 GB/sn/an/a
31.6 TOPS21.473 TFLOPS, 27.199 TFLOPSAppleSoCMobile, PCMM2 Ultran/aAPL1W122023n/aTSMC N5Pn/an/an/aLPDDR5-6400819.2 GB/sn/an/a
18 TOPS2.826 TFLOPS, 3.533 TFLOPSAppleSoCMobile, PCMM3n/aAPL12012023MacBook ProTSMC N3Bn/an/an/aLPDDR5-6400102.4 GB/sn/an/a
18 TOPS10.598 TFLOPS, 14.131 TFLOPSAppleSoCMobile, PCMM3 Maxn/aAPL12042023n/aTSMC N3Bn/an/an/aLPDDR5-6400307.2 GB/s, 409.6 GB/sn/an/a
18 TOPS4.946 TFLOPS, 6.359 TFLOPSAppleSoCMobile, PCMM3 Pron/aAPL12032023n/aTSMC N3Bn/an/an/aLPDDR5-6400153.6 GB/sn/an/a
38 TOPS3.763 TFLOPSAppleSoCMobile, PCMM4n/aAPL12062024iPad Pro (7th generation)TSMC N3E10 cores (4 performance + 6 efficiency)Apple-designed 10-core16-core Neural EngineLPDDR5X-7500120 GB/sn/an/a
38 TOPSn/aAppleSoCMobile, PCMM4 Maxn/an/a2024MacBook Pro M4 MaxTSMC N3E14 cores (10 performance + 4 efficiency) 16 cores (12 performance + 4 efficiency)Apple-designed 16-core Apple-designed 20-core16-core Neural EngineLPDDR5X-8533409.6 GB/s (36GB), 546 GB/s (48GB, 64GB, 128GB)n/an/a
38 TOPSn/aAppleSoCMobile, PCMM4 Pron/an/a2024MacBook Pro M4 Pro, Mac mini M4 ProTSMC N3E12 cores (8 performance + 4 efficiency) 14 cores (10 performance + 4 efficiency)Apple-designed 32-core Apple-designed 40-core16-core Neural EngineLPDDR5X-8533273 GB/sn/an/a
n/an/aGoogleSoCDatacenter (Infra)GCP CPUAxionn/aAxion2024GCP Compute Engine ???n/a?? Neoverse V2 coren/an/an/an/an/an/a
1.6 TOPSn/aGoogleSoCMobileGoogle Tensor (Edge TPU)G11Whitechapel2021Pixel 6, Pixel 6 Pro, Pixel 6aSamsung 5 nm LPEOcta-core: 2.8 GHz Cortex-X1 (2×) 2.25 GHz Cortex-A76 (2×) 1.8 GHz Cortex-A55 (4×)Mali-G78 MP20 at 848 MHzGoogle Edge TPULPDDR551.2 GB/sn/an/a
n/an/aGoogleSoCMobileGoogle Tensor (Edge TPU)G22Cloudripper2022Pixel 7, Pixel 7 Pro, Pixel 7a, Pixel Fold, Pixel TabletSamsung 5 nm LPEOcta-core: 2.85 GHz Cortex-X1 (2×) 2.35 GHz Cortex-A78 (2×) 1.8 GHz Cortex-A55 (4×)Mali-G710 MP7 at 850 MHzGoogle Edge TPULPDDR551.2 GB/sn/an/a
27 TOPSn/aGoogleSoCMobileGoogle Tensor (Edge TPU)G33Zuma (Dev Board: Ripcurrent)2023Pixel 8, Pixel 8 Pro, Pixel 8aSamsung 4nm LPPNona-core: 2.91 GHz Cortex-X3 (1×) 2.37 GHz Cortex-A715 (4×) 1.7 GHz Cortex-A510 (4×)Mali-G715 MP10 at 890 MHzGoogle Edge TPU (Rio)LPDDR5X68.2 GB/sn/an/a
45 TOPSn/aGoogleSoCMobileGoogle Tensor (Edge TPU)G44Zuma Pro2024Pixel 9, Pixel 9 ProSamsung 4nm LPPOcta-core: 3.1 GHz Cortex-X4 (1×) 2.6 GHz Cortex-A720 (3×) 1.92 GHz Cortex-A520 (4×)Mali-G715 MP10 at 940 MHzn/aLPDDR5Xn/an/a- 8Gen3 = 45 TOPS, D9300 = 48 TOPS
n/an/aGoogleSoCMobileGoogle Tensor (Edge TPU)G55Laguna Beach (Dev Board: Deepspace)2025Pixel 10, Pixel 10 ProTSMC N3 + InFO-POP packagingn/an/an/an/an/an/an/a
23 TOPSn/aGoogleSoCDatacenter (AI inference)TPUTPUv11n/a2015n/a28nmn/an/an/aDDR3-213334 GB/s75.0- The core of TPU: Systolic Array - Matrix Multiply Unit (MXU): a big systolic array - PCIe Gen3 x16
45 TOPS3 TFLOPSGoogleSoCDatacenter (AI inference)TPUTPUv22n/a2017n/a16nmn/an/an/an/a600 GB/s280.0- 16GB HBM - BF16
123 TOPS4 TFLOPSGoogleSoCDatacenter (AI inference)TPUTPUv33n/a2018n/a16nmn/an/an/an/a900 GB/s220.0n/a
275 TOPSn/aGoogleSoCDatacenter (AI inference)TPUTPUv44n/a2021n/a7nmn/an/an/an/a1,200 GB/s170.0- 32GB HBM2
393 TOPSn/aGoogleSoCDatacenter (AI inference)TPUTPUv5e5n/a2023n/an/an/an/an/an/a819 GB/sn/an/a
918 TOPSn/aGoogleSoCDatacenter (AI inference)TPUTPUv5p5n/a2023n/an/an/an/an/an/a2,765 GB/sn/an/a
n/an/aGoogleSoCDatacenter (AI inference)TPUTPUv6? Trillium?6n/a2024n/an/an/an/an/an/an/an/an/a
n/a31 TFLOPSGraphcoreSoCDatacenterColossusColossus MK1 GC2 IPU1n/a2017n/aTSMC 16nm1216 processor coresn/an/an/a45,000 GB/sn/an/a
n/a62 TFLOPSGraphcoreSoCDatacenterColossusColossus MK2 GC200 IPU2n/a2020n/aTSMC 7nm1472 processor coresn/an/an/a47,500 GB/sn/an/a
n/an/aGraphcoreSoCDatacenterColossusColossus MK3 (TBD)3n/an/an/an/an/an/an/an/an/an/an/a
n/an/aIntelSoCHP Mobile, PCn/an/an/aArrow Laken/an/an/an/an/an/an/an/an/an/a
120 TOPSn/aIntelSoCLP MobileCore UltraCore UltraSeries 2Lunar Lake2024n/aTSMC N3B (Compute tile), TSMC N6 (Platform controoler tile)P-core: Lion Cove E-core: SkymontXe2NPU 4n/an/an/a- Total 120 TOPS (48 TOPS from NPU 4 + 67 TOPS from GPU + 5 TOPS from CPU).
34 TOPSn/aIntelSoCMobileCore UltraCore UltraSeries 1Meteor Lake2023n/aIntel 4 (7nm EUV, Compute tile), TSMC N5 (Graphics tile), TSMC N6 (Soc tile, I/O extender tile)P-core: Redwood Cove E-core: CrestmontXe-LPGNPU 3720n/an/an/a- Total 34 TOPS (11 TOPS from NPU + 18 TOPS from GPU + 5 TOPS from CPU).
0.5 TOPSn/aIntelNPUn/an/aNPU 11n/a2018n/an/an/an/an/an/an/an/an/a
7 TOPSn/aIntelNPUn/an/aNPU 22n/a2021n/an/an/an/an/an/an/an/an/a
11.5 TOPSn/aIntelNPUn/an/aNPU 33n/a2023n/an/an/an/an/an/an/an/an/a
48 TOPSn/aIntelNPUn/an/aNPU 44n/a2024Lunar Laken/an/an/an/an/an/an/an/a
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9000 天璣 90009000n/a2021Redmi K50 Pro OPPO Find X5 Pro 天璣版 vivo X80 / X80 Pro 天璣版TSMC N41× Cortex-X2 @ 3.05 GHz 3× Cortex-A710 @ 2.85 GHz 4× Cortex-A510 @ 1.8 GHzMali-G710 MP10 @ 850 MHzMediaTek APU 590n/an/an/a- 5G NR Sub-6GHz, LTE
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9000+ 天璣 9000+9000n/a2022小米12 Pro 天璣版 華碩 ROG Phone 6D Ultimate iQOO Neo 7 OPPO Find N2 FlipTSMC N41× Cortex-X2 @ 3.2 GHz 3× Cortex-A710 @ 2.85 GHz 4× Cortex-A510 @ 1.8 GHzMali-G710 MC10MediaTek APU 590n/an/an/a- 5G NR Sub-6GHz, LTE
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9200 天璣 92009000n/a2022vivo X90, vivo X90 Pro OPPO Find X6 OPPO Find N3 FlipTSMC N41× Cortex-X3 @ 3.05GHz 3× Cortex-A715 @ 2.85GHz 4× Cortex-A510 @ 1.8GHzMali-Immortalis-G715 MP11 @ 981 MHzMediaTek APU 690n/an/an/a- 5G NR Sub-6 GHz, 5G mmWave, LTE
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9200+ 天璣 9200+9000n/a2023iQOO Neo8 Pro vivo X90s Redmi K60至尊版TSMC N41× Cortex-X3 @ 3.35 GHz 3× Cortex-A715 @ 3.0 GHz 4× Cortex-A510 @ 2.0 GHzMali-Immortalis-G715 MC11MediaTek APU 690n/an/an/a- 5G NR Sub-6 GHz, 5G mmWave, LTE
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9300 天璣 93009000n/a2023vivo X100, vivo X100 Pro OPPO Find X7TSMC N4P1× Cortex-X4 @ 3.25 GHz 3× Cortex-X4 @ 2.85 GHz 4× Cortex-A720 @ 2.0 GHzMali-Immortalis-G720 MC12 @ 1300 MHzMediaTek APU 790n/an/an/a- 5G NR (Sub-6 GHz & mmWave), 4G LTE, quad-band GNSS (BeiDou, Galileo, GLONASS, GPS, NavIC, QZSS), Bluetooth 5.4, Wi-Fi 7 (2x2)
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9300+ 天璣 9300+9000n/a2024vivo X100S, vivo X100X ProTSMC N4P1× Cortex-X4 @ 3.4 GHz 3× Cortex-X4 @ 2.85 GHz 4× Cortex-A720 @ 2.0 GHzMali-Immortalis-G720 MC12 @ 1300 MHzMediaTek APU 790n/an/an/a- 5G NR (Sub-6 GHz & mmWave), 4G LTE, quad-band GNSS (BeiDou, Galileo, GLONASS, GPS, NavIC, QZSS), Bluetooth 5.4, Wi-Fi 7 (2x2)
n/an/aMediaTekSoCMobileDimensity 天璣Dimensity 9400 天璣 94009000n/a2024vivo X200, OPPO Find X8 / ProTSMC N31× Cortex-X925 @ 3.63 GHz 3× Cortex-X4 @ 2.8 GHz 4× Cortex-A725 @ 2.1 GHzMali-Immortalis-G925 MC12 @ ??? MHzn/an/an/an/an/a
n/an/aMicrosoftSoCDatacenter (Infra)Azure CobaltCobalt 1001n/a2024Azure VM Dpsv6, Dplsv6, Epsv6n/a128 Neoverse V2 coren/an/aLPDDR5 ???n/an/a- PCIe gen5 - CXL 1.1 - from project start to silicon in 13 months.
1,600 TOPSn/aMicrosoftSoCDatacenter (AI inference)Azure MaiaMaia 1001n/a2024Microsoft CopilotTSMC N5 + CoWoS-Sn/an/an/an/a18,000 GB/s ???500.0- 32Gb/s PCIe gen5x8 - Design to TDP = 700W - Provision TDP = 500W
n/a15.1 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4060n/aAD107-4002023n/aTSMC N4n/an/an/aGDDR6272 GB/s115.0- PCIe 4.0 x8
n/a22.1 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4060 Tin/aAD106-3512023n/aTSMC N4n/an/an/aGDDR6288 GB/s160.0- PCIe 4.0 x8
n/a29.1 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4070n/aAD104-2502023n/aTSMC N4n/an/an/aGDDR6X504 GB/s200.0- PCIe 4.0 x16
n/a35.48 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4070 Supern/aAD104-3502024n/aTSMC N4n/an/an/aGDDR6X504 GB/s220.0- PCIe 4.0 x16
n/a40.1 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4070 Tin/aAD104-4002023n/aTSMC N4n/an/an/aGDDR6X504 GB/s285.0- PCIe 4.0 x16
n/a44.10 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4070 Ti Supern/aAD103-2752024n/aTSMC N4n/an/an/aGDDR6X672 GB/s285.0- PCIe 4.0 x16
n/a48.7 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4080n/aAD103-3002022n/aTSMC N4n/an/an/aGDDR6X717 GB/s320.0- PCIe 4.0 x16
n/a52.22 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4080 Supern/aAD103-4002024n/aTSMC N4n/an/an/aGDDR6X736 GB/s320.0- PCIe 4.0 x16
n/a82.6 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4090n/aAD102-3002022n/aTSMC N4n/an/an/aGDDR6X1008 GB/s450.0- PCIe 4.0 x16
n/a73.5 TFLOPSNVIDIAGPUDesktopGeForce RTX 40GeForce RTX 4090 Dn/aAD102-2502023n/aTSMC N4n/an/an/aGDDR6X1008 GB/s425.0- PCIe 4.0 x16
n/a124.96 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)A10Amperen/a2021n/an/an/a1× GA102-890-A1n/aGDDR6600 GB/sn/an/a
624 TOPS312.0 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)A100Amperen/a2020n/aTSMC N7n/a1× GA100-883AA-A1n/aHBM21555 GB/s400.0n/a
n/a73.728 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)A16Amperen/a2021n/an/an/a4× GA107n/aGDDR64x 200 GB/sn/an/a
n/a18.124 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)A2Amperen/a2021n/an/an/a1× GA107n/aGDDR6200 GB/s60.0n/a
n/a165.12 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)A30Amperen/a2021n/an/an/a1× GA100n/aHBM2933.1 GB/sn/an/a
n/a149.68 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)A40Amperen/a2020n/an/an/a1× GA102n/aGDDR6695.8 GB/sn/an/a
3500 TOPS (3.5 POPS)n/aNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)B100 (SXM6 card)Blackwelln/a2024n/aTSMC 4NP (custom N4P)n/an/an/aHBM3E8000 GB/s700.0n/a
4500 TOPS (4.5 POPS)n/aNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)B200 (SXM6 card)Blackwelln/a2024n/aTSMC 4NP (custom N4P)n/an/an/aHBM3E8000 GB/s1000.0n/a
n/a756.449 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)H100 (PCIe card)Hoppern/a2022n/aTSMC 4N (custom N4)n/a1× GH100n/aHBM2E2039 GB/sn/an/a
1980 TOPS (1.98 POPS)989.43 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)H100 (SXM5 card)Hoppern/a2022n/aTSMC 4N (custom N4)n/a1× GH100n/aHBM33352 GB/s700.0n/a
1980 TOPS (1.98 POPS)67 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)H200 (SXM5 card)Hoppern/a2023n/aTSMC 4N (custom N4)n/an/an/aHBM3E4800 GB/s1000.0n/a
n/a121.0 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)L4Ada Lovelacen/a2023n/an/an/a1x AD104n/aGDDR61563 GB/sn/an/a
n/a362.066 TFLOPSNVIDIAGPUDatacenterNvidia Data Center GPUs (Nvidia Tesla)L40Ada Lovelacen/a2022n/an/an/a1× AD102n/aGDDR62250 GB/sn/an/a
n/a2.774 TFLOPSQualcommSoCMobileSnapdragon 8Snapdragon 8 Gen 38n/a2023n/aTSMC N4P1× 3.30 GHz Kryo Prime (Cortex-X4) + 3× 3.15 GHz Kryo Gold (Cortex-A720) + 2× 2.96 GHz Kryo Gold (Cortex-A720) + 2× 2.27 GHz Kryo Silver (Cortex-A520)Adreno 750 @ 903 MHzn/aLPDDR5X76.8 GB/sn/an/a
n/a1.689 TFLOPSQualcommSoCMobileSnapdragon 8Snapdragon 8s Gen 38n/a2024n/aTSMC N4P1× 3.0 GHz Kryo Prime (Cortex-X4) + 4× 2.8 GHz Kryo Gold (Cortex-A720) + 3× 2.0 GHz Kryo Silver (Cortex-A520)Adreno 735 @ 1100 MHzn/aLPDDR5X76.8 GB/sn/an/a
45 TOPS4.6 TFLOPSQualcommSoCPCSnapdragon XSnapdragon X EliteXn/a2023n/aTSMC N4OryonAdreno X1HexagonLPDDR5X-8448 @ 4224 MHz135 GB/sn/a- Total 75 TOPS (45 TOPS from NPU).
45 TOPS3.8 TFLOPSQualcommSoCPCSnapdragon XSnapdragon X PlusXn/a2024n/aTSMC N4OryonAdreno X1-45 1107 MHz (1.7 TFLOPS) Adreno X1-45 (2.1 TFLOPS) Adreno X1-85 1250 MHz (3.8 TFLOPS)HexagonLPDDR5X-8448 @ 4224 MHz135 GB/sn/an/a
45 TOPSn/aQualcommNPUn/aHexagonHexagonn/an/an/aSnapdragon X Plusn/an/an/an/an/an/an/a- Hexagon is the brand name for a family of digital signal processor (DSP) and later neural processing unit (NPU) products by Qualcomm. Hexagon is also known as QDSP6, standing for “sixth generation digital signal processor.”
n/a2.1 TFLOPSQualcommGPUn/aAdrenoAdreno X1-45XAdreno 726n/an/aTSMC N4n/an/an/aLPDDR5X-8448 @ 4224 MHz or LPDDR5X-8533 @ 4266.5 MHz125.1 GB/s or 136.5 GB/sn/a- The Adreno X1-45 is internally called the Adreno 726, suggesting it’s a scaled-up of the Adreno 725 from the Snapdragon 7+ Gen 2.
n/a4.6 TFLOPSQualcommGPUn/aAdrenoAdreno X1-85XAdreno 741n/aSnapdragon X PlusTSMC N4n/an/an/aLPDDR5X-8448 @ 4224 MHz or LPDDR5X-8533 @ 4266.5 MHz125.1 GB/s or 136.5 GB/sn/a- The Adreno X1-85 is internally called the Adreno 741, suggesting it’s a scaled-up of the Adreno 730 from the Snapdragon 8 Gen 1/8+ Gen 1.

參考資料 Reference

Loading comments…