Design of Digital Circuits Lecture 21: SIMD Processors II and Graphics Processing Units

Transcription

1 Desig of Digital Circuits Lecture 21: SIMD Processors II ad Graphics Processig Uits Dr. Jua Gómez Lua Prof. Our Mutlu ETH Zurich Sprig May 2018

2 New Course: Bachelor s Semiar i Comp Arch Fall credit uits Rigorous semiar o fudametal ad cuttig-edge topics i computer architecture Critical presetatio, review, ad discussio of semial works i computer architecture We will cover may ideas & issues, aalyze their tradeoffs, perform critical thikig ad braistormig Participatio, presetatio, report ad review writig Stay tued for more iformatio 2

3 Ageda for Today & Next Few Lectures Sigle-cycle Microarchitectures Multi-cycle ad Microprogrammed Microarchitectures Pipeliig Issues i Pipeliig: Cotrol & Data Depedece Hadlig, State Maiteace ad Recovery, Out-of-Order Executio Other Executio Paradigms 3

4 Readigs for Today Peleg ad Weiser, MMX Techology Extesio to the Itel Architecture, IEEE Micro Lidholm et al., "NVIDIA Tesla: A Uified Graphics ad Computig Architecture," IEEE Micro

5 Other Approaches to Cocurrecy (or Istructio Level Parallelism)

6 Approaches to (Istructio-Level) Cocurrecy Pipeliig Out-of-order executio Dataflow (at the ISA level) Superscalar Executio VLIW Fie-Graied Multithreadig SIMD Processig (Vector ad array processors, GPUs) Decoupled Access Execute Systolic Arrays 6

7 SIMD Processig: Exploitig Regular (Data) Parallelism

8 Recall: Fly s Taxoomy of Computers Mike Fly, Very High-Speed Computig Systems, Proc. of IEEE, 1966 SISD: Sigle istructio operates o sigle data elemet SIMD: Sigle istructio operates o multiple data elemets Array processor Vector processor MISD: Multiple istructios operate o sigle data elemet Closest form: systolic array processor, streamig processor MIMD: Multiple istructios operate o multiple data elemets (multiple istructio streams) Multiprocessor Multithreaded processor 8

9 Recall: SIMD Processig Sigle istructio operates o multiple data elemets I time or i space Multiple processig elemets Time-space duality Array processor: Istructio operates o multiple data elemets at the same time usig differet spaces Vector processor: Istructio operates o multiple data elemets i cosecutive time steps usig the same space 9

10 Recall: Array vs. Vector Processors ARRAY PROCESSOR VECTOR PROCESSOR Istructio Stream LD VR ß A[3:0] ADD VR ß VR, 1 MUL VR ß VR, 2 ST A[3:0] ß VR Time Same same time LD0 LD1 LD2 LD3 LD0 Differet time AD0 AD1 AD2 AD3 LD1 AD0 MU0 MU1 MU2 MU3 LD2 AD1 MU0 ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0 Differet same space AD3 MU2 ST1 MU3 ST2 Same space ST3 Space Space 10

11 Recall: Memory Bakig Memory is divided ito baks that ca be accessed idepedetly; baks share address ad data buses (to miimize pi cost) Ca start ad complete oe bak access per cycle Ca sustai N parallel accesses if all N go to differet baks Bak 0 Bak 1 Bak 2 Bak 15 MDR MAR MDR MAR MDR MAR MDR MAR Data bus Address bus Picture credit: Derek Chiou CPU 11

12 Some Issues Stride ad bakig As log as they are relatively prime to each other ad there are eough baks to cover bak access latecy, we ca sustai 1 elemet/cycle throughput Storage of a matrix Row major: Cosecutive elemets i a row are laid out cosecutively i memory Colum major: Cosecutive elemets i a colum are laid out cosecutively i memory You eed to chage the stride whe accessig a row versus colum 12

13 Matrix Multiplicatio A ad B, both i row-major order A B A 4x6 B 6x10 C 4x10 Dot products of rows ad colums of A ad B A: Load A 0 ito vector register V 1 Each time, icremet address by oe to access the ext colum Accesses have a stride of 1 B: Load B 0 ito vector register V 2 Each time, icremet address by 10 Accesses have a stride of 10 Differet strides ca lead to bak coflicts How do we miimize them? 13

14 Miimizig Bak Coflicts More baks Better data layout to match the access patter Is this always possible? Better mappig of address to bak E.g., radomized mappig Rau, Pseudo-radomly iterleaved memory, ISCA

15 Recall: Questios (II) What if vector data is ot stored i a strided fashio i memory? (irregular memory access to a vector) Idea: Use idirectio to combie/pack elemets ito vector registers Called scatter/gather operatios 15

16 Gather/Scatter Operatios Wat to vectorize loops with idirect accesses: for (i=0; i<n; i++) A[i] = B[i] + C[D[i]] Idexed istructio (Gather) LV vd, rd # Load idices i D vector LVI vc, rc, vd # Load idirect from rc base LV vb, rb # Load B vector ADDV.D va,vb,vc # Do add SV va, ra # Store result 16

17 Gather/Scatter Operatios Gather/scatter operatios ofte implemeted i hardware to hadle sparse vectors (matrices) Vector s ad stores use a idex vector which is added to the base register to geerate the addresses Scatter example Idex Vector Data Vector (to Store) Stored Vector (i Memory) Base Base+1 X Base Base+3 X Base+4 X Base+5 X Base Base

18 Array vs. Vector Processors, Revisited Array vs. vector processor distictio is a purist s distictio Most moder SIMD processors are a combiatio of both They exploit data parallelism i both time ad space GPUs are a prime example we will cover i a bit more detail 18

19 Recall: Array vs. Vector Processors ARRAY PROCESSOR VECTOR PROCESSOR Istructio Stream LD VR ß A[3:0] ADD VR ß VR, 1 MUL VR ß VR, 2 ST A[3:0] ß VR Time Same same time LD0 LD1 LD2 LD3 LD0 Differet time AD0 AD1 AD2 AD3 LD1 AD0 MU0 MU1 MU2 MU3 LD2 AD1 MU0 ST0 ST1 ST2 ST3 LD3 AD2 MU1 ST0 Differet same space AD3 MU2 ST1 MU3 ST2 Same space ST3 Space Space 19

20 Vector Istructio Executio VADD A,B à C Executio usig oe pipelied fuctioal uit Executio usig four pipelied fuctioal uits A[6] B[6] A[24] B[24] A[25] B[25] A[26] B[26] A[27] B[27] A[5] B[5] A[20] B[20] A[21] B[21] A[22] B[22] A[23] B[23] A[4] B[4] A[16] B[16] A[17] B[17] A[18] B[18] A[19] B[19] A[3] B[3] A[12] B[12] A[13] B[13] A[14] B[14] A[15] B[15] C[2] C[8] C[9] C[10] C[11] C[1] C[4] C[5] C[6] C[7] Time Time C[0] Slide credit: Krste Asaovic C[0] C[1] Space C[2] C[3] 20

21 Vector Uit Structure Fuctioal Uit Partitioed Vector Registers Elemets 0, 4, 8, Elemets 1, 5, 9, Elemets 2, 6, 10, Elemets 3, 7, 11, Lae Memory Subsystem Slide credit: Krste Asaovic 21

22 Vector Istructio Level Parallelism Ca overlap executio of multiple vector istructios Example machie has 32 elemets per vector register ad 8 laes Completes 24 operatios/cycle while issuig 1 vector istructio/cycle Load Uit Multiply Uit Add Uit mul time add mul add Istructio issue Slide credit: Krste Asaovic 22

23 Automatic Code Vectorizatio Scalar Seuetial Code for (i=0; i < N; i++) C[i] = A[i] + B[i]; Vectorized Code Iter. 1 add Time add add store store store Iter. 2 Iter. 1 Iter. 2 Vector Istructio add store Vectorizatio is a compile-time reorderig of operatio seuecig Þ reuires extesive loop depedece aalysis Slide credit: Krste Asaovic 23

24 Vector/SIMD Processig Summary Vector/SIMD machies are good at exploitig regular datalevel parallelism Same operatio performed o may data elemets Improve performace, simplify desig (o itra-vector depedecies) Performace improvemet limited by vectorizability of code Scalar operatios limit vector machie performace Remember Amdahl s Law CRAY-1 was the fastest SCALAR machie at its time! May existig ISAs iclude (vector-like) SIMD operatios Itel MMX/SSE/AVX, PowerPC AltiVec, ARM Advaced SIMD 24

25 SIMD Operatios i Moder ISAs

26 SIMD ISA Extesios Sigle Istructio Multiple Data (SIMD) extesio istructios Sigle istructio acts o multiple pieces of data at oce Commo applicatio: graphics Perform short arithmetic operatios (also called packed arithmetic) For example: add four 8-bit umbers Must modify ALU to elimiate carries betwee 8-bit values padd8 $s2, $s0, $s Bit positio a 3 a 2 a 1 a 0 $s0 + b 3 b 2 b 1 b 0 $s1 a 3 + b 3 a 2 + b 2 a 1 + b 1 a 0 + b 0 $s2 26

27 Itel Petium MMX Operatios Idea: Oe istructio operates o multiple data elemets simultaeously À la array processig (yet much more limited) Desiged with multimedia (graphics) operatios i mid No VLEN register Opcode determies data type: 8 8-bit bytes 4 16-bit words 2 32-bit doublewords 1 64-bit uadword Stride is always eual to 1. Peleg ad Weiser, MMX Techology Extesio to the Itel Architecture, IEEE Micro,

28 MMX Example: Image Overlayig (I) Goal: Overlay the huma i image 1 o top of the backgroud i image 2 for (i=o: i<image-size; i++) i if (x[il == Blue) ew-image[i] =y[il; else ew-image[il = x[il; 1 Peleg ad Weiser, MMX Techology Extesio to the Itel Architecture, IEEE Micro,

29 MMX Example: Image Overlayig (II) Y = Blossom image X = Woma s image Peleg ad Weiser, MMX Techology Extesio to the Itel Architecture, IEEE Micro,

30 GPUs (Graphics Processig Uits)

31 GPUs are SIMD Egies Udereath The istructio pipelie operates like a SIMD pipelie (e.g., a array processor) However, the programmig is doe usig threads, NOT SIMD istructios To uderstad this, let s go back to our parallelizable code example But, before that, let s distiguish betwee Programmig Model (Software) vs. Executio Model (Hardware) 31

32 Programmig Model vs. Hardware Executio Model Programmig Model refers to how the programmer expresses the code E.g., Seuetial (vo Neuma), Data Parallel (SIMD), Dataflow, Multi-threaded (MIMD, SPMD), Executio Model refers to how the hardware executes the code udereath E.g., Out-of-order executio, Vector processor, Array processor, Dataflow processor, Multiprocessor, Multithreaded processor, Executio Model ca be very differet from the Programmig Model E.g., vo Neuma model implemeted by a OoO processor E.g., SPMD model implemeted by a SIMD processor (a GPU) 32

33 How Ca You Exploit Parallelism Here? for (i=0; i < N; i++) Scalar Seuetial Code C[i] = A[i] + B[i]; Iter. 1 add store Let s examie three programmig optios to exploit istructio-level parallelism preset i this seuetial code: Iter Seuetial (SISD) add 2. Data-Parallel (SIMD) store 3. Multithreaded (MIMD/SPMD) 33

34 Prog. Model 1: Seuetial (SISD) for (i=0; i < N; i++) C[i] = A[i] + B[i]; Scalar Seuetial Code Ca be executed o a: Iter. 1 Pipelied processor Out-of-order executio processor add Idepedet istructios executed whe ready store Differet iteratios are preset i the istructio widow ad ca execute i parallel i multiple fuctioal uits Iter. 2 add I other words, the loop is dyamically urolled by the hardware Superscalar or VLIW processor store Ca fetch ad execute multiple istructios per cycle 34

35 Prog. Model 2: Data Parallel (SIMD) for (i=0; i < N; i++) C[i] = A[i] + B[i]; Scalar Seuetial Code Vector Istructio Vectorized Code VLD A à V1 Iter. 1 VLD B à V2 add add VADD V1 + V2 à V3 store store VST V3 à C Iter. 2 Iter. 1 Iter. 2 Realizatio: Each iteratio is idepedet add store Idea: Programmer or compiler geerates a SIMD istructio to execute the same istructio from all iteratios across differet data Best executed by a SIMD processor (vector, array) 35

36 Prog. Model 3: Multithreaded for (i=0; i < N; i++) C[i] = A[i] + B[i]; Scalar Seuetial Code Iter. 1 Iter. 2 Iter. 1 add store add store add store Iter. 2 Realizatio: Each iteratio is idepedet Idea: Programmer or compiler geerates a thread to execute each iteratio. Each thread does the same thig (but o differet data) Ca be executed o a MIMD machie 36

37 Prog. Model 3: Multithreaded for (i=0; i < N; i++) C[i] = A[i] + B[i]; add store add store Iter. 1 Iter. 2 Realizatio: Each iteratio is idepedet Idea: This Programmer particular or model compiler is also geerates called: a thread to execute each iteratio. Each thread does the same SPMD: thig (but Sigle o Program differet data) Multiple Data Ca Ca be executed be executed o a MIMD o a SIMT SIMD machie machie Sigle Istructio Multiple Thread 37

38 A GPU is a SIMD (SIMT) Machie Except it is ot programmed usig SIMD istructios It is programmed usig threads (SPMD programmig model) Each thread executes the same code but operates a differet piece of data Each thread has its ow cotext (i.e., ca be treated/restarted/executed idepedetly) A set of threads executig the same istructio are dyamically grouped ito a warp (wavefrot) by the hardware A warp is essetially a SIMD operatio formed by hardware! 38

39 SPMD o SIMT Machie for (i=0; i < N; i++) C[i] = A[i] + B[i]; Warp 0 at PC X Warp 0 at PC X+1 add add Warp 0 at PC X+2 store store Warp 0 at PC X+3 Iter. 1 Iter. 2 Warp: A set of threads that execute Realizatio: Each iteratio is idepedet the same istructio (i.e., at the same PC) Idea: This Programmer particular or model compiler is also geerates called: a thread to execute each iteratio. Each thread does the same SPMD: thig Sigle (but o Program differet data) Multiple Data Ca A GPU Ca be executed be executes executed o it a usig MIMD o a SIMD the machie SIMT machie model: Sigle Istructio Multiple Thread 39

40 Graphics Processig Uits SIMD ot Exposed to Programmer (SIMT)

41 SIMD vs. SIMT Executio Model SIMD: A sigle seuetial istructio stream of SIMD istructios à each istructio specifies multiple data iputs [VLD, VLD, VADD, VST], VLEN SIMT: Multiple istructio streams of scalar istructios à threads grouped dyamically ito warps [LD, LD, ADD, ST], NumThreads Two Major SIMT Advatages: Ca treat each thread separately à i.e., ca execute each thread idepedetly (o ay type of scalar pipelie) à MIMD processig Ca group threads ito warps flexibly à i.e., ca group threads that are supposed to truly execute the same istructio à dyamically obtai ad maximize beefits of SIMD processig 41

42 Multithreadig of Warps for (i=0; i < N; i++) C[i] = A[i] + B[i]; Assume a warp cosists of 32 threads If you have 32K iteratios, ad 1 iteratio/thread à 1K warps Warps ca be iterleaved o the same pipelie à Fie graied multithreadig of warps Warp 10 at PC X add store add store Warp 20 at PC X+2 Iter * Iter *

43 Warps ad Warp-Level FGMT Warp: A set of threads that execute the same istructio (o differet data elemets) à SIMT (Nvidia-speak) All threads ru the same code Warp: The threads that ru legthwise i a wove fabric Thread Warp Scalar Scalar Scalar Thread Thread Thread W X Y Commo PC Scalar Thread Z Thread Warp 3 Thread Warp 8 Thread Warp 7 SIMD Pipelie 43

44 High-Level View of a GPU 44

45 Latecy Hidig via Warp-Level FGMT Warp: A set of threads that execute the same istructio (o differet data elemets) Thread Warp 3 Thread Warp 8 Warps available for schedulig Fie-graied multithreadig Oe istructio per thread i pipelie at a time (No iterlockig) Iterleave warp executio to hide latecies Register values of all threads stay i register file FGMT eables log latecy tolerace Millios of pixels Thread Warp 7 RF ALU All Hit? I-Fetch Decode RF ALU D-Cache Writeback Data RF ALU SIMD Pipelie Warps accessig memory hierarchy Miss? Thread Warp 1 Thread Warp 2 Thread Warp 6 Slide credit: Tor Aamodt 45

46 Warp Executio (Recall the Slide) 32-thread warp executig ADD A[tid],B[tid] à C[tid] Executio usig oe pipelied fuctioal uit Executio usig four pipelied fuctioal uits A[6] B[6] A[24] B[24] A[25] B[25] A[26] B[26] A[27] B[27] A[5] B[5] A[20] B[20] A[21] B[21] A[22] B[22] A[23] B[23] A[4] B[4] A[16] B[16] A[17] B[17] A[18] B[18] A[19] B[19] A[3] B[3] A[12] B[12] A[13] B[13] A[14] B[14] A[15] B[15] C[2] C[8] C[9] C[10] C[11] C[1] C[4] C[5] C[6] C[7] Time Time C[0] Slide credit: Krste Asaovic C[0] C[1] Space C[2] C[3] 46

47 SIMD Executio Uit Structure Fuctioal Uit Registers for each Thread Registers for thread IDs 0, 4, 8, Registers for thread IDs 1, 5, 9, Registers for thread IDs 2, 6, 10, Registers for thread IDs 3, 7, 11, Lae Memory Subsystem Slide credit: Krste Asaovic 47

48 Warp Istructio Level Parallelism Ca overlap executio of multiple istructios Example machie has 32 threads per warp ad 8 laes Completes 24 operatios/cycle while issuig 1 warp/cycle Load Uit Multiply Uit Add Uit W0 W1 time W2 W3 W4 W5 Warp issue Slide credit: Krste Asaovic 48

49 SIMT Memory Access Same istructio i differet threads uses thread id to idex ad access differet data elemets + Let s assume N=16, 4 threads per warp à 4 warps Threads Data elemets Warp 0 Warp 1 Warp 2 Warp 3 Slide credit: Hyesoo Kim 48

50 Sample GPU SIMT Code (Simplified) CPU code for (ii = 0; ii < ; ++ii) { C[ii] = A[ii] + B[ii]; } CUDA code // there are threads global void KerelFuctio( ) { it tid = blockdim.x * blockidx.x + threadidx.x; it vara = aa[tid]; it varb = bb[tid]; C[tid] = vara + varb; } Slide credit: Hyesoo Kim 48

51 Sample GPU Program (Less Simplified) Slide credit: Hyesoo Kim 51

52 Warp-based SIMD vs. Traditioal SIMD Traditioal SIMD cotais a sigle thread Seuetial istructio executio; lock-step operatios i a SIMD istructio Programmig model is SIMD (o extra threads) à SW eeds to kow vector legth ISA cotais vector/simd istructios Warp-based SIMD cosists of multiple scalar threads executig i a SIMD maer (i.e., same istructio executed by all threads) Does ot have to be lock step Each thread ca be treated idividually (i.e., placed i a differet warp) à programmig model ot SIMD SW does ot eed to kow vector legth Eables multithreadig ad flexible dyamic groupig of threads ISA is scalar à SIMD operatios ca be formed dyamically Essetially, it is SPMD programmig model implemeted o SIMD hardware 52

53 SPMD Sigle procedure/program, multiple data This is a programmig model rather tha computer orgaizatio Each processig elemet executes the same procedure, except o differet data elemets Procedures ca sychroize at certai poits i program, e.g. barriers Essetially, multiple istructio streams execute the same program Each program/procedure 1) works o differet data, 2) ca execute a differet cotrol-flow path, at ru-time May scietific applicatios are programmed this way ad ru o MIMD hardware (multiprocessors) Moder GPUs programmed i a similar way o a SIMD hardware 53

54 SIMD vs. SIMT Executio Model SIMD: A sigle seuetial istructio stream of SIMD istructios à each istructio specifies multiple data iputs [VLD, VLD, VADD, VST], VLEN SIMT: Multiple istructio streams of scalar istructios à threads grouped dyamically ito warps [LD, LD, ADD, ST], NumThreads Two Major SIMT Advatages: Ca treat each thread separately à i.e., ca execute each thread idepedetly o ay type of scalar pipelie à MIMD processig Ca group threads ito warps flexibly à i.e., ca group threads that are supposed to truly execute the same istructio à dyamically obtai ad maximize beefits of SIMD processig 54

55 Threads Ca Take Differet Paths i Warp-based SIMD Each thread ca have coditioal cotrol flow istructios Threads ca execute differet cotrol flow paths A B Thread Warp Commo PC C D F Thread 1 Thread 2 Thread 3 Thread 4 E G Slide credit: Tor Aamodt 55

56 Cotrol Flow Problem i GPUs/SIMT A GPU uses a SIMD pipelie to save area o cotrol logic Groups scalar threads ito warps Brach Brach divergece occurs whe threads iside warps brach to differet executio paths Path A Path B This is the same as coditioal/predicated/masked executio. Recall the Vector Mask ad Masked Vector Operatios? Slide credit: Tor Aamodt 56

57 Remember: Each Thread Is Idepedet Two Major SIMT Advatages: Ca treat each thread separately à i.e., ca execute each thread idepedetly o ay type of scalar pipelie à MIMD processig Ca group threads ito warps flexibly à i.e., ca group threads that are supposed to truly execute the same istructio à dyamically obtai ad maximize beefits of SIMD processig If we have may threads We ca fid idividual threads that are at the same PC Ad, group them together ito a sigle warp dyamically This reduces divergece à improves SIMD utilizatio SIMD utilizatio: fractio of SIMD laes executig a useful operatio (i.e., executig a active thread) 57

58 Dyamic Warp Formatio/Mergig Idea: Dyamically merge threads executig the same istructio (after brach divergece) Form ew warps from warps that are waitig Eough threads brachig to each path eables the creatio of full ew warps Warp X Warp Z Warp Y 58

59 Dyamic Warp Formatio/Mergig Idea: Dyamically merge threads executig the same istructio (after brach divergece) Brach Path A Path B Fug et al., Dyamic Warp Formatio ad Schedulig for Efficiet GPU Cotrol Flow, MICRO

60 Dyamic Warp Formatio Example B x/1110 y/0011 A x/1111 y/1111 C x/1000 y/0010 D x/0110 y/0001 F x/0001 y/1100 E x/1110 y/0011 A D Leged A Executio of Warp x at Basic Block A A ew warp created from scalar threads of both Warp x ad y executig at Basic Block D Executio of Warp y at Basic Block A Baselie Dyamic Warp Formatio G x/1111 y/1111 A A B B C C D D E E F F G G A A A A B B C D E E F G G A A Time Time Slide credit: Tor Aamodt 60

61 Hardware Costraits Limit Flexibility of Warp Groupig Fuctioal Uit Registers for each Thread Registers for thread IDs 0, 4, 8, Registers for thread IDs 1, 5, 9, Registers for thread IDs 2, 6, 10, Registers for thread IDs 3, 7, 11, Lae Ca you move ay thread flexibly to ay lae? Memory Subsystem Slide credit: Krste Asaovic 61

62 Desig of Digital Circuits Lecture 21: SIMD Processors II ad Graphics Processig Uits Dr. Jua Gómez Lua Prof. Our Mutlu ETH Zurich Sprig May 2018

63 We did ot cover the followig slides i lecture. These are for your preparatio for the ext lecture.

64 A Example GPU

65 NVIDIA GeForce GTX 285 NVIDIA-speak: 240 stream processors SIMT executio Geeric speak: 30 cores 8 SIMD fuctioal uits per core Slide credit: Kayvo Fatahalia 65

66 NVIDIA GeForce GTX 285 core 64 KB of storage for thread cotexts (registers) = SIMD fuctioal uit, cotrol shared across 8 uits = multiply-add = multiply = istructio stream decode = executio cotext storage Slide credit: Kayvo Fatahalia 66

67 NVIDIA GeForce GTX 285 core 64 KB of storage for thread cotexts (registers) Groups of 32 threads share istructio stream (each group is a Warp) Up to 32 warps are simultaeously iterleaved Up to 1024 thread cotexts ca be stored Slide credit: Kayvo Fatahalia 67

68 NVIDIA GeForce GTX 285 Tex Tex Tex Tex Tex Tex Tex Tex Tex Tex 30 cores o the GTX 285: 30,720 threads Slide credit: Kayvo Fatahalia 68

69 Evolutio of NVIDIA GPUs #Stream Processors GFLOPS Stream Processors GFLOPS 0 GTX 285 (2009) GTX 480 (2010) GTX 780 (2013) GTX 980 (2014) P100 (2016) V100 (2017) 0 69

70 NVIDIA V100 NVIDIA-speak: 5120 stream processors SIMT executio Geeric speak: 80 cores 64 SIMD fuctioal uits per core Tesor cores for Machie Learig 70

71 NVIDIA V100 Block Diagram 80 cores o the V

72 NVIDIA V100 Core 15.7 TFLOPS Sigle Precisio 7.8 TFLOPS Double Precisio 125 TFLOPS for Deep Learig (Tesor cores) 72