A basic introduction to GPU architecture. Based on Kayvon\'s "From Shader Code to a Teraflop: How GPU Shader Cores Work"
Updated to include the latest GPUs: AMD Tahiti (HD7970) and NVIDIA Kepler (GTX690)
How to Troubleshoot Apps for the Modern Connected Worker
Introduction To GPUs 2012
1. Introduction to GPU Compute
Architecture
Ofer Rosenberg
Based on
“From Shader Code to a Teraflop: How GPU Shader Cores Work”,
By Kayvon Fatahalian, Stanford University and Mike Houston, Fellow, AMD
2. Intro: some numbers…
Sources:
http://www.anandtech.com/show/6025/radeon-hd-7970-ghz-
edition-review-catching-up-to-gtx-680
http://ark.intel.com/products/65722
http://www.techpowerup.com/cpudb/1028/Intel_Xeon_E3-
1290V2.html
Intel IvyBridge 4C AMD Radeon HD 7970 Ratio
(E3-1290V2 )
Cores 4 32
Frequency 3700MHz 1000MHz
Process 22nm 28nm
Transistor Count 1400M 4310M ~3x
Power 87W 250W ~3x
Compute Power 237 GFLOPS 4096 GFLOPS ~17x
2
3. Content
1. Three major ideas that make GPU processing cores run
fast
2. Closer look at real GPU designs
– NVIDIA GTX 680
– AMD Radeon 7970
3. The GPU memory hierarchy: moving data to processors
4. Heterogeneous Cores
4. Part 1: throughput processing
• Three key concepts behind how modern GPU
processing cores run code
• Knowing these concepts will help you:
1. Understand space of GPU core (and throughput CPU core) designs
2. Optimize shaders/compute kernels
3. Establish intuition: what workloads might benefit from the design of
these architectures?
5. What’s in a GPU?
A GPU is a heterogeneous chip multi-processor (highly tuned for graphics)
Shader Shader Input Assembly
Tex
Core Core
Rasterizer
Shader Shader Tex
Core Core Output Blend
Shader Shader Video Decode
Tex
Core Core
Work HW
Shader Shader Tex Distributor or
Core Core SW?
6. A diffuse reflectance shader
sampler mySamp;
Texture2D<float3> myTex;
float3 lightDir; Shader programming model:
float4 diffuseShader(float3 norm, float2 uv) Fragments are processed
{
independently,
float3 kd;
but there is no explicit parallel
kd = myTex.Sample(mySamp, uv);
kd *= clamp( dot(lightDir, norm), 0.0, 1.0);
programming
return float4(kd, 1.0);
}
8. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
9. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
10. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
11. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
12. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
13. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
14. Execute shader
Fetch/
Decode
<diffuseShader>:
sample r0, v4, t0, s0
ALU mul r3, v0, cb0[0]
(Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0)
Execution mul o0, r0, r3
mul o1, r1, r3
Context
mul o2, r2, r3
mov o3, l(1.0)
15. “CPU-style” cores
Fetch/
Decode
Data cache
(a big one)
ALU
(Execute)
Execution Out-of-order control logic
Context
Fancy branch predictor
Memory pre-fetcher
16. Slimming down
Fetch/
Decode
Idea #1:
ALU
(Execute) Remove components that
help a single instruction
Execution
Context stream run fast
17. Two cores (two fragments in parallel)
fragment 1 fragment 2
Fetch/ Fetch/
Decode Decode
<diffuseShader>: <diffuseShader>:
sample r0, v4, t0, s0 ALU ALU sample r0, v4, t0, s0
mul r3, v0, cb0[0] mul r3, v0, cb0[0]
madd r3, v1, cb0[1], r3 (Execute) (Execute) madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3 madd r3, v2, cb0[2], r3
clmp r3, r3, l(0.0), l(1.0) clmp r3, r3, l(0.0), l(1.0)
mul o0, r0, r3 mul o0, r0, r3
mul o1, r1, r3 Execution Execution mul o1, r1, r3
mul o2, r2, r3 mul o2, r2, r3
mov o3, l(1.0) Context Context mov o3, l(1.0)
18. Four cores (four fragments in parallel)
Fetch/ Fetch/
Decode Decode
ALU ALU
(Execute) (Execute)
Execution Execution
Context Context
Fetch/ Fetch/
Decode Decode
ALU ALU
(Execute) (Execute)
Execution Execution
Context Context
22. Add ALUs
Fetch/
Idea #2:
Decode Amortize cost/complexity of
ALU 1 ALU 2 ALU 3 ALU 4 managing an instruction
ALU 5 ALU 6 ALU 7 ALU 8 stream across many ALUs
Ctx Ctx Ctx Ctx
Ctx Ctx Ctx Ctx
SIMD processing
Shared Ctx Data
23. Modifying the shader
Fetch/
Decode <diffuseShader>:
sample r0, v4, t0, s0
mul r3, v0, cb0[0]
ALU 1 ALU 2 ALU 3 ALU 4
madd r3, v1, cb0[1], r3
madd r3, v2, cb0[2], r3
ALU 5 ALU 6 ALU 7 ALU 8
clmp r3, r3, l(0.0), l(1.0)
mul o0, r0, r3
mul o1, r1, r3
Ctx Ctx Ctx Ctx mul o2, r2, r3
mov o3, l(1.0)
Ctx Ctx Ctx Ctx
Original compiled shader:
Shared Ctx Data
Processes one fragment using
scalar ops on scalar registers
24. Modifying the shader
Fetch/
Decode <VEC8_diffuseShader>:
VEC8_sample vec_r0, vec_v4, t0, vec_s0
VEC8_mul vec_r3, vec_v0, cb0[0]
ALU 1 ALU 2 ALU 3 ALU 4 VEC8_madd vec_r3, vec_v1, cb0[1], vec_r3
VEC8_madd vec_r3, vec_v2, cb0[2], vec_r3
ALU 5 ALU 6 ALU 7 ALU 8 VEC8_clmp vec_r3, vec_r3, l(0.0), l(1.0)
VEC8_mul vec_o0, vec_r0, vec_r3
VEC8_mul vec_o1, vec_r1, vec_r3
VEC8_mul vec_o2, vec_r2, vec_r3
Ctx Ctx Ctx Ctx VEC8_mov o3, l(1.0)
Ctx Ctx Ctx Ctx
New compiled shader:
Shared Ctx Data
Processes eight fragments using
vector ops on vector registers
25. Modifying the shader
1 2 3 4
5 6 7 8
Fetch/
Decode <VEC8_diffuseShader>:
VEC8_sample vec_r0, vec_v4, t0, vec_s0
VEC8_mul vec_r3, vec_v0, cb0[0]
ALU 1 ALU 2 ALU 3 ALU 4 VEC8_madd vec_r3, vec_v1, cb0[1], vec_r3
VEC8_madd vec_r3, vec_v2, cb0[2], vec_r3
ALU 5 ALU 6 ALU 7 ALU 8 VEC8_clmp vec_r3, vec_r3, l(0.0), l(1.0)
VEC8_mul vec_o0, vec_r0, vec_r3
VEC8_mul vec_o1, vec_r1, vec_r3
VEC8_mul vec_o2, vec_r2, vec_r3
Ctx Ctx Ctx Ctx VEC8_mov o3, l(1.0)
Ctx Ctx Ctx Ctx
Shared Ctx Data
26. 128 fragments in parallel
16 cores = 128 ALUs , 16 simultaneous instruction streams
27. vertices/fragments
128 [ primitives
OpenCL work items ] in parallel
vertices
primitives
fragments
28. But what about branches?
1 2 ... ... 8
Time (clocks)
ALU 1 ALU 2 . . . . . . ALU 8
<unconditional
shader code>
if (x > 0) {
y = pow(x, exp);
y *= Ks;
refl = y + Ka;
} else {
x = 0;
refl = Ka;
}
<resume unconditional
shader code>
29. But what about branches?
1 2 ... ... 8
Time (clocks)
ALU 1 ALU 2 . . . . . . ALU 8
<unconditional
shader code>
T T F T F F F F if (x > 0) {
y = pow(x, exp);
y *= Ks;
refl = y + Ka;
} else {
x = 0;
refl = Ka;
}
<resume unconditional
shader code>
30. But what about branches?
1 2 ... ... 8
Time (clocks)
ALU 1 ALU 2 . . . . . . ALU 8
<unconditional
shader code>
T T F T F F F F if (x > 0) {
y = pow(x, exp);
y *= Ks;
refl = y + Ka;
} else {
x = 0;
refl = Ka;
}
<resume unconditional
Not all ALUs do useful work! shader code>
Worst case: 1/8 peak
performance
31. But what about branches?
1 2 ... ... 8
Time (clocks)
ALU 1 ALU 2 . . . . . . ALU 8
<unconditional
shader code>
T T F T F F F F if (x > 0) {
y = pow(x, exp);
y *= Ks;
refl = y + Ka;
} else {
x = 0;
refl = Ka;
}
<resume unconditional
shader code>
32. Clarification
SIMD processing does not imply SIMD instructions
• Option 1: explicit vector instructions
– x86 SSE, AVX, Intel Larrabee
• Option 2: scalar instructions, implicit HW vectorization
– HW determines instruction stream sharing across ALUs (amount of sharing
hidden from software)
– NVIDIA GeForce (“SIMT” warps), ATI Radeon architectures (“wavefronts”)
In practice: 16 to 64 fragments share an instruction stream.
33. Stalls!
Stalls occur when a core cannot run the next instruction
because of a dependency on a previous operation.
Texture access latency = 100’s to 1000’s of cycles
We’ve removed the fancy caches and logic that helps avoid stalls.
34. But we have LOTS of independent fragments.
Idea #3:
Interleave processing of many fragments on a single
core to avoid stalls caused by high latency operations.
35. Hiding shader stalls
Time (clocks) Frag 1 … 8
Fetch/
Decode
ALU 1 ALU 2 ALU 3 ALU 4
ALU 5 ALU 6 ALU 7 ALU 8
Ctx Ctx Ctx Ctx
Ctx Ctx Ctx Ctx
Shared Ctx Data
36. Hiding shader stalls
Time (clocks) Frag 1 … 8 Frag 9 … 16 Frag 17 … 24 Frag 25 … 32
1 2 3 4
Fetch/
Decode
ALU 1 ALU 2 ALU 3 ALU 4
ALU 5 ALU 6 ALU 7 ALU 8
1 2
3 4
43. Four large contexts (low latency hiding ability)
Fetch/
Decode
ALU 1 ALU 2 ALU 3 ALU 4
ALU 5 ALU 6 ALU 7 ALU 8
1 2
3 4
44. Clarification
Interleaving between contexts can be managed by
hardware or software (or both!)
• NVIDIA / AMD Radeon GPUs
– HW schedules / manages all contexts (lots of them)
– Special on-chip storage holds fragment state
• Intel Larrabee
– HW manages four x86 (big) contexts at fine granularity
– SW scheduling interleaves many groups of fragments on each HW context
– L1-L2 cache holds fragment state (as determined by SW)
45. Example chip
16 cores
8 mul-add ALUs per core
(128 total)
16 simultaneous
instruction streams
64 concurrent (but interleaved)
instruction streams
512 concurrent fragments
= 256 GFLOPs (@ 1GHz)
46. Summary: three key ideas
1. Use many “slimmed down cores” to run in parallel
2. Pack cores full of ALUs (by sharing instruction stream across
groups of fragments)
– Option 1: Explicit SIMD vector instructions
– Option 2: Implicit sharing managed by hardware
3. Avoid latency stalls by interleaving execution of many groups of
fragments
– When one group stalls, work on another group
47. Part 2:
Putting the three ideas into practice:
A closer look at real GPUs
NVIDIA GeForce GTX 680
AMD Radeon HD 7970
48. Disclaimer
• The following slides describe “a reasonable way to think”
about the architecture of commercial GPUs
• Many factors play a role in actual chip performance
49. NVIDIA GeForce GTX 680 (Kepler)
• NVIDIA-speak:
– 1536 stream processors (“CUDA cores”)
– “SIMT execution”
• Generic speak:
– 8 cores
– 6 groups of 32-wide SIMD functional units per core
50. NVIDIA GeForce GTX 680 “core”
Fetch/
Decode = SIMD function unit,
control shared across 32 units
(1 MUL-ADD per clock)
• Groups of 32 [fragments/vertices/CUDA
threads] share an instruction stream
• Up to 64 groups are simultaneously
interleaved
Execution contexts • Up to 2048 individual contexts can be
(256 KB) stored
“Shared” memory
(16+48 KB)
Source: http://www.geforce.com/Active/en_US/en_US/pdf/GeForce-GTX-680-Whitepaper-FINAL.pdf
http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
51. NVIDIA GeForce GTX 680 “core”
Fetch/
Decode = SIMD function unit,
control shared across 32 units
Fetch/
Decode (1 MUL-ADD per clock)
Fetch/
Decode
• The core contains 192 functional units
Fetch/
Decode
• Six groups are selected each clock
Fetch/
Decode (decode, fetch, and execute six
Fetch/ instruction streams in parallel)
Decode
Execution contexts
(128 KB)
“Shared” memory
(16+48 KB)
Source: http://www.geforce.com/Active/en_US/en_US/pdf/GeForce-GTX-680-Whitepaper-FINAL.pdf
http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
52. NVIDIA GeForce GTX 680 “SMX”
Fetch/
Decode = CUDA core
(1 MUL-ADD per clock)
Fetch/
Decode
Fetch/
Decode
• The SMX contains 192 CUDA cores
Fetch/
Decode
• Six warps are selected each clock
Fetch/
Decode (decode, fetch, and execute six
Fetch/ warps in parallel)
Decode
Execution contexts • Up to 64 warps are interleaved,
(128 KB) totaling 2048 CUDA threads
“Shared” memory
(16+48 KB)
Source: http://www.geforce.com/Active/en_US/en_US/pdf/GeForce-GTX-680-Whitepaper-FINAL.pdf
http://www.nvidia.com/content/PDF/kepler/NVIDIA-Kepler-GK110-Architecture-Whitepaper.pdf
53. NVIDIA GeForce GTX 680
There are 8 of
these things on the
GTX 680:
That’s 16,384
fragments!
Or 16,384 CUDA
threads!
54. AMD Radeon HD 7970 (Tahiti)
• AMD-speak:
– 2048 stream processors
– “GCN”: Graphics Core Next
• Generic speak:
– 32 cores
– 4 groups of 64-wide SIMD functional units per core
55. AMD Radeon HD 7970 “core”
Fetch/ = SIMD function unit,
Decode
control shared across 64 units
(1 MUL-ADD per clock)
• Groups of 64
[fragments/vertices/OpenCL Workitems]
share an instruction stream
• Up to 40 groups are simultaneously
interleaved
Execution Execution Execution Execution
context context context context
64 KB 64 KB 64 KB 64 KB • Up to 2560 individual contexts can be
“Shared” memory stored
(64 KB)
Source: http://developer.amd.com/afds/assets/presentations/2620_final.pdf
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute
56. AMD Radeon HD 7970 “core”
Fetch/ = SIMD function unit,
Decode
control shared across 64 units
(1 MUL-ADD per clock)
Fetch/
Decode
• The core contains 64 functional units
Fetch/
Decode • Four clocks to execute an instruction
for all fragments in a group
Fetch/
Decode
• Four groups are executed in parallel
Execution
context
Execution
context
Execution
context
Execution
context
each clock (decode, fetch, and
64 KB 64 KB 64 KB 64 KB execute four instruction streams in
“Shared” memory parallel)
(64 KB)
Source: http://developer.amd.com/afds/assets/presentations/2620_final.pdf
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute
57. AMD Radeon HD 7970 “Compute Unit”
Fetch/ = SIMD function unit,
Decode
control shared across 64 units
(1 MUL-ADD per clock)
Fetch/
Decode
• Groups of 64 [fragments/vertices/etc.]
Fetch/ are a wavefront
Decode
• Four wavefronts are executed each
Fetch/
Decode clock
Execution
context
Execution
context
Execution
context
Execution
context
• Up to 40 wavefronts are interleaved,
64 KB 64 KB 64 KB 64 KB totaling 2560 Threads
“Shared” memory
(64 KB)
Source: http://developer.amd.com/afds/assets/presentations/2620_final.pdf
http://www.anandtech.com/show/4455/amds-graphics-core-next-preview-amd-architects-for-compute
58. AMD Radeon HD 7970
There are 32 of these things on the HD 7970: That’s 81,920 fragments!
59. The talk thus far: processing data
Part 3: moving data to processors
60. Recall: “CPU-style” core
OOO exec logic
Branch predictor
Fetch/Decode
Data cache
ALU (a big one)
Execution
Context
61. “CPU-style” memory hierarchy
OOO exec logic
L1 cache
Branch predictor (32 KB)
25 GB/sec
Fetch/Decode to memory
L3 cache
ALU (8 MB)
shared across cores
Execution
contexts
L2 cache
(256 KB)
CPU cores run efficiently when data is resident in cache
(caches reduce latency, provide high bandwidth)
62. Throughput core (GPU-style)
Fetch/
Decode
ALU 1 ALU 2 ALU 3 ALU 4
280 GB/sec
ALU 5 ALU 6 ALU 7 ALU 8
Memory
Execution
contexts
(256 KB)
More ALUs, no large traditional cache hierarchy:
Need high-bandwidth connection to memory
63. Bandwidth is a critical resource
– A high-end GPU (e.g. Radeon HD 7970) has...
• Over twenty times (4.1 TFLOPS) the compute performance of quad-core
CPU
• No large cache hierarchy to absorb memory requests
– GPU memory system is designed for throughput
• Wide bus (280 GB/sec)
• Repack/reorder/interleave memory requests to maximize use of memory
bus
• Still, this is only ten times the bandwidth available to CPU
64. Bandwidth thought experiment
A
Task: element-wise multiply two long vectors A and B ×
1.Load input A[i] B
+
2.Load input B[i] C
=
3.Load input C[i] D
4.Compute A[i] × B[i] + C[i]
5.Store result into D[i]
Four memory operations (16 bytes) for every MUL-ADD
Radeon HD 7970 can do 2048 MUL-ADDS per clock
Need ~32 TB/sec of bandwidth to keep functional units busy
Less than 1% efficiency… but 6x faster than CPU!
65. Bandwidth limited!
If processors request data at too high a rate,
the memory system cannot keep up.
No amount of latency hiding helps this.
Overcoming bandwidth limits are a common challenge
for GPU-compute application developers.
66. Reducing bandwidth requirements
• Request data less often (instead, do more math)
– “arithmetic intensity”
• Fetch data from memory less often (share/reuse data
across fragments
– on-chip communication or storage
67. Reducing bandwidth requirements
• Two examples of on-chip storage
– Texture caches
– OpenCL “local memory” (CUDA shared memory)
1 2 3 4
Texture caches:
Capture reuse across
Texture data fragments, not temporal
reuse within a single
shader program
68. Modern GPU memory hierarchy
Fetch/
Decode
Texture cache
ALU 1 ALU 2 ALU 3 ALU 4 (read-only)
ALU 5 ALU 6 ALU 7 ALU 8 L2 cache
(~1 MB) Memory
Shared “local”
Execution storage
contexts or
(256 KB) L1 cache
(64 KB)
On-chip storage takes load off memory system.
Many developers calling for more cache-like storage
(particularly GPU-compute applications)
69. Don’t forget about offload cost…
• PCIe bandwidth/latency
– 8GB/s each direction in practice
– Attempt to pipeline/multi-buffer uploads and downloads
• Dispatch latency
– O(10) usec to dispatch from CPU to GPU
– This means offload cost if O(10M) instructions
69
70. Heterogeneous devices to the rescue ?
• Tighter integration of CPU and GPU style cores
– Reduce offload cost
– Reduce memory copies/transfers
– Power management
• Industry shifting rapidly in this direction
– AMD Fusion APUs – NVIDIA Tegra 3
– Intel SNB/IVY – Apple A4 and A5
–… – QUALCOMM Snapdragon
– TI OMAP
–… 70