Jump to content

Pascal (microarchitecture)

From Wikipedia, the free encyclopedia
(Redirected from Pascal microarchitecture)

Pascal
A GTX 1070 Founders Edition graphics based on the Pascal architecture
LaunchedMay 27, 2016; 8 years ago (2016-05-27)
Designed byNvidia
Manufactured by
Fabrication process
Codename(s)GP10x
Product Series
Desktop
Professional/workstation
Server/datacenter
Specifications
L1 cache24 KB (per SM)
L2 cache256 KB—4 MB
Memory support
PCIe supportPCIe 3.0
Supported Graphics APIs
DirectXDirectX 12 (12.1)
Direct3DDirect3D 12.0
Shader ModelShader Model 6.7
OpenCLOpenCL 3.0
OpenGLOpenGL 4.6
CUDACompute Capability 6.0
VulkanVulkan 1.3
Media Engine
Encode codecs
Decode codecs
Color bit-depth
  • 8-bit
  • 10-bit
Encoder(s) supportedNVENC
Display outputs
History
PredecessorMaxwell
Successor
Painting of Blaise Pascal, eponym of architecture

Pascal is the codename for a GPU microarchitecture developed by Nvidia, as the successor to the Maxwell architecture. The architecture was first introduced in April 2016 with the release of the Tesla P100 (GP100) on April 5, 2016, and is primarily used in the GeForce 10 series, starting with the GeForce GTX 1080 and GTX 1070 (both using the GP104 GPU), which were released on May 27, 2016, and June 10, 2016, respectively. Pascal was manufactured using TSMC's 16 nm FinFET process,[1] and later Samsung's 14 nm FinFET process.[2]

The architecture is named after the 17th century French mathematician and physicist, Blaise Pascal.

In April 2019, Nvidia enabled a software implementation of DirectX Raytracing on Pascal-based cards starting with the GTX 1060 6 GB, and in the 16 series cards, a feature reserved to the Turing-based RTX series up to that point.[3][4]

Details

[edit]
Die shot of the GP100 GPU used in Nvidia Tesla P100 cards
Die shot of the GP102 GPU found inside GeForce GTX 1080 Ti cards
Die shot of the GP106 GPU found inside GTX 1060 cards

In March 2014, Nvidia announced that the successor to Maxwell would be the Pascal microarchitecture; announced on May 6, 2016, and released on May 27 of the same year. The Tesla P100 (GP100 chip) has a different version of the Pascal architecture compared to the GTX GPUs (GP104 chip). The shader units in GP104 have a Maxwell-like design.[5]

Architectural improvements of the GP100 architecture include the following:[6][7][8]

  • In Pascal, a SM (streaming multiprocessor) consists of between 64-128 CUDA cores, depending on if it is GP100 or GP104. Maxwell contained 128 CUDA cores per SM; Kepler had 192, Fermi 32 and Tesla 8. The GP100 SM is partitioned into two processing blocks, each having 32 single-precision CUDA cores, an instruction buffer, a warp scheduler, 2 texture mapping units and 2 dispatch units.
  • CUDA Compute Capability 6.0.
  • High Bandwidth Memory 2 — some cards feature 16 GiB HBM2 in four stacks with a total bus width of 4096 bits and a memory bandwidth of 720 GB/s.
  • Unified memory — a memory architecture where the CPU and GPU can access both main system memory and memory on the graphics card with the help of a technology called "Page Migration Engine".
  • NVLink — a high-bandwidth bus between the CPU and GPU, and between multiple GPUs. Allows much higher transfer speeds than those achievable by using PCI Express; estimated to provide between 80 and 200 GB/s.[9][10]
  • 16-bit (FP16) floating-point operations (colloquially "half precision") can be executed at twice the rate of 32-bit floating-point operations ("single precision")[11] and 64-bit floating-point operations (colloquially "double precision") executed at half the rate of 32-bit floating point operations.[12]
  • More registers — twice the amount of registers per CUDA core compared to Maxwell.
  • More shared memory.
  • Dynamic load balancing scheduling system.[13] This allows the scheduler to dynamically adjust the amount of the GPU assigned to multiple tasks, ensuring that the GPU remains saturated with work except when there is no more work that can safely be distributed to distribute.[13] Nvidia therefore has safely enabled asynchronous compute in Pascal's driver.[13]
  • Instruction-level and thread-level preemption.[14]

Architectural improvements of the GP104 architecture include the following:[5]

  • CUDA Compute Capability 6.1.
  • GDDR5X — new memory standard supporting 10Gbit/s data rates, updated memory controller.[15]
  • Simultaneous Multi-Projection - generating multiple projections of a single geometry stream, as it enters the SMP engine from upstream shader stages.[16]
  • DisplayPort 1.4, HDMI 2.0b.
  • Fourth generation Delta Color Compression.
  • Enhanced SLI Interface — SLI interface with higher bandwidth compared to the previous versions.
  • PureVideo Feature Set H hardware video decoding HEVC Main10 (10-bit), Main12 (12-bit) and VP9 hardware decoding.
  • HDCP 2.2 support for 4K DRM protected content playback and streaming (Maxwell GM200 and GM204 lack HDCP 2.2 support, GM206 supports HDCP 2.2).[17]
  • NVENC HEVC Main10 10bit hardware encoding.
  • GPU Boost 3.0.
  • Instruction-level preemption.[14] In graphics tasks, the driver restricts preemption to the pixel-level, because pixel tasks typically finish quickly and the overhead costs of doing pixel-level preemption are lower than instruction-level preemption (which is expensive).[14] Compute tasks get thread-level or instruction-level preemption,[14] because they can take longer times to finish and there are no guarantees on when a compute task finishes. Therefore the driver enables the expensive instruction-level preemption for these tasks.[14]

Overview

[edit]

Graphics Processor Cluster

[edit]

A chip is partitioned into Graphics Processor Clusters (GPCs). For the GP104 chips, a GPC encompasses 5 SMs.

Streaming Multiprocessor "Pascal"

[edit]

A "Streaming Multiprocessor" is analogous to AMD's Compute Unit. An SM encompasses 128 single-precision ALUs ("CUDA cores") on GP104 chips and 64 single-precision ALUs on GP100 chips. While all CU versions consist of 64 shader processors (i.e. 4 SIMD Vector Units, each 16 lanes wide), Nvidia experimented with very different numbers of CUDA cores:

  • On Tesla, 1 SM combines 8 single-precision (FP32) shader processors
  • On Fermi, 1 SM combines 32 single-precision (FP32) shader processors
  • On Kepler, 1 SM combines 192 single-precision (FP32) shader processors and 64 double-precision (FP64) units (on GK110 GPUs)
  • On Maxwell, 1 SM combines 128 single-precision (FP32) shader processors
  • On Pascal, it depends:
    • On GP100, 1 SM combines 64 single-precision (FP32) shader processors and also 32 double-precision (FP64) providing a 2:1 ratio of single- to double-precision throughput. The GP100 uses more flexible FP32 cores that are able to process one single-precision or two half-precision numbers in a two-element vector.[18] This is intended to better serve machine learning tasks.
    • On GP104, 1 SM combines 128 single-precision ALUs, 4 double-precision ALUs (providing a 32:1 ratio), and one half-precision ALU which contains a vector of two half-precision floats which can execute the same instruction on both floats, providing a 64:1 ratio if the same instruction is used on both elements.

Polymorph-Engine 4.0

[edit]

The Polymorph Engine version 4.0 is the unit responsible for Tessellation. It corresponds functionally with AMD's Geometric Processor. It has been moved from the shader module to the TPC to allow one Polymorph engine to feed multiple SMs within the TPC.[19]

Chips

[edit]
GTX 1080 Ti PCB and die
  • GP100: Nvidia's Tesla P100 GPU accelerator is targeted at GPGPU applications such as FP64 double precision compute and deep learning training that uses FP16. It uses HBM2 memory.[20] Quadro GP100 also uses the GP100 GPU.
  • GP102: This GPU is used in the Titan Xp,[21] Titan X Pascal[22] and the GeForce GTX 1080 Ti. It is also used in the Quadro P6000[23] & Tesla P40.[24]
  • GP104: This GPU is used in the GeForce GTX 1070, GTX 1070 Ti, GTX 1080, and some GTX 1060 6 GB's. The GTX 1070 has 15/20 and the GTX 1070 Ti has 19/20 of its SMs enabled; both utilize GDDR5 memory. The GTX 1080 is a fully unlocked chip and uses GDDR5X memory. Some GTX 1060 6 GB's use GP104 with 10/20 SMs enabled and GDDR5X memory.[25] It is also used in the Quadro P5000, Quadro P4000, Quadro P3200 (mobile applications) and Tesla P4.
  • GP106: This GPU is used in the GeForce GTX 1060 with GDDR5[26] memory.[27][28] It is also used in the Quadro P2000.
  • GP107: This GPU is used in the GeForce GTX 1050 and 1050 Ti. It is also used in the Quadro P1000, Quadro P600, Quadro P620 & Quadro P400.
  • GP108: This GPU is used in the GeForce GT 1010 and GeForce GT 1030.
Comparison table of some Kepler, Maxwell, and Pascal chips
GK104 GK110 GM204 (GTX 970) GM204 (GTX 980) GM200 GP104 GP100
Dedicated texture cache per SM 48 KiB — — — — — —
Texture (graphics or compute) or read-only data (compute only) cache per SM — 48 KiB[29] — — — — —
Programmer-selectable shared memory/L1 partitions per SM 48 KiB shared memory + 16 KiB L1 cache (default)[30] 48 KiB shared memory + 16 KiB L1 cache (default)[30] — — — — —
32 KiB shared memory + 32 KiB L1 cache[30] 32 KiB shared memory + 32 KiB L1 cache[30]
16 KiB shared memory + 48 KiB L1 cache[30] 16 KiB shared memory + 48 KiB L1 cache[30]
Unified L1 cache/texture cache per SM — — 48 KiB[31] 48 KiB[31] 48 KiB[31] 48 KiB[31] 24 KiB[31]
Dedicated shared memory per SM — — 96 KiB[31] 96 KiB[31] 96 KiB[31] 96 KiB[31] 64 KiB[31]
L2 cache per chip 512 KiB[31] 1536 KiB[31] 1792 KiB[32] 2048 KiB[32] 3072 KiB[31] 2048 KiB[31] 4096 KiB[31]

Performance

[edit]

The theoretical single-precision processing power of a Pascal GPU in GFLOPS is computed as 2 × operations per FMA instruction per CUDA core per cycle × number of CUDA cores × core clock speed (in GHz).

The theoretical double-precision processing power of a Pascal GPU is 1/2 of the single precision performance on Nvidia GP100, and 1/32 of Nvidia GP102, GP104, GP106, GP107 & GP108.

The theoretical half-precision processing power of a Pascal GPU is 2× of the single precision performance on GP100[12] and 1/64 on GP104, GP106, GP107 & GP108.[18]

Successor

[edit]

The Pascal architecture was succeeded in 2017 by Volta in the HPC, cloud computing, and self-driving car markets, and in 2018 by Turing in the consumer and business market.[33]

P100 accelerator and DGX-1

[edit]

Comparison of accelerators used in DGX:[34][35][36]

Model Architecture Socket FP32
CUDA
cores		 FP64 cores
(excl. tensor)		 Mixed
INT32/FP32
cores		 INT32
cores		 Boost
clock		 Memory
clock		 Memory
bus width		 Memory
bandwidth		 VRAM Single
precision
(FP32)		 Double
precision
(FP64)		 INT8
(non-tensor)		 INT8
dense tensor		 INT32 FP4
dense tensor		 FP16 FP16
dense tensor		 bfloat16
dense tensor		 TensorFloat-32
(TF32)
dense tensor		 FP64
dense tensor		 Interconnect
(NVLink)		 GPU L1 Cache L2 Cache TDP Die size Transistor
count		 Process Launched
B200 Blackwell SXM6 N/A N/A N/A N/A N/A 8 Gbit/s HBM3e 8192-bit 8 TB/sec 192 GB HBM3e N/A N/A N/A 4.5 POPS N/A 9 PFLOPS N/A 2.25 PFLOPS 2.25 PFLOPS 1.2 PFLOPS 40 TFLOPS 1.8 TB/sec GB100 N/A N/A 1000 W N/A 208 B TSMC 4NP Q4 2024 (expected)
B100 Blackwell SXM6 N/A N/A N/A N/A N/A 8 Gbit/s HBM3e 8192-bit 8 TB/sec 192 GB HBM3e N/A N/A N/A 3.5 POPS N/A 7 PFLOPS N/A 1.98 PFLOPS 1.98 PFLOPS 989 TFLOPS 30 TFLOPS 1.8 TB/sec GB100 N/A N/A 700 W N/A 208 B TSMC 4NP
H200 Hopper SXM5 16896 4608 16896 N/A 1980 MHz 6.3 Gbit/s HBM3e 6144-bit 4.8 TB/sec 141 GB HBM3e 67 TFLOPS 34 TFLOPS N/A 1.98 POPS N/A N/A N/A 990 TFLOPS 990 TFLOPS 495 TFLOPS 67 TFLOPS 900 GB/sec GH100 25344 KB (192 KB × 132) 51200 KB 1000 W 814 mm2 80 B TSMC 4N Q3 2023
H100 Hopper SXM5 16896 4608 16896 N/A 1980 MHz 5.2 Gbit/s HBM3 5120-bit 3.35 TB/sec 80 GB HBM3 67 TFLOPS 34 TFLOPS N/A 1.98 POPS N/A N/A N/A 990 TFLOPS 990 TFLOPS 495 TFLOPS 67 TFLOPS 900 GB/sec GH100 25344 KB (192 KB × 132) 51200 KB 700 W 814 mm2 80 B TSMC 4N Q3 2022
A100 80GB Ampere SXM4 6912 3456 6912 N/A 1410 MHz 3.2 Gbit/s HBM2e 5120-bit 1.52 TB/sec 80 GB HBM2e 19.5 TFLOPS 9.7 TFLOPS N/A 624 TOPS 19.5 TOPS N/A 78 TFLOPS 312 TFLOPS 312 TFLOPS 156 TFLOPS 19.5 TFLOPS 600 GB/sec GA100 20736 KB (192 KB × 108) 40960 KB 400 W 826 mm2 54.2 B TSMC N7 Q1 2020
A100 40GB Ampere SXM4 6912 3456 6912 N/A 1410 MHz 2.4 Gbit/s HBM2 5120-bit 1.52 TB/sec 40 GB HBM2 19.5 TFLOPS 9.7 TFLOPS N/A 624 TOPS 19.5 TOPS N/A 78 TFLOPS 312 TFLOPS 312 TFLOPS 156 TFLOPS 19.5 TFLOPS 600 GB/sec GA100 20736 KB (192 KB × 108) 40960 KB 400 W 826 mm2 54.2 B TSMC N7
V100 32GB Volta SXM3 5120 2560 N/A 5120 1530 MHz 1.75 Gbit/s HBM2 4096-bit 900 GB/sec 32 GB HBM2 15.7 TFLOPS 7.8 TFLOPS 62 TOPS N/A 15.7 TOPS N/A 31.4 TFLOPS 125 TFLOPS N/A N/A N/A 300 GB/sec GV100 10240 KB (128 KB × 80) 6144 KB 350 W 815 mm2 21.1 B TSMC 12FFN Q3 2017
V100 16GB Volta SXM2 5120 2560 N/A 5120 1530 MHz 1.75 Gbit/s HBM2 4096-bit 900 GB/sec 16 GB HBM2 15.7 TFLOPS 7.8 TFLOPS 62 TOPS N/A 15.7 TOPS N/A 31.4 TFLOPS 125 TFLOPS N/A N/A N/A 300 GB/sec GV100 10240 KB (128 KB × 80) 6144 KB 300 W 815 mm2 21.1 B TSMC 12FFN
P100 Pascal SXM/SXM2 N/A 1792 3584 N/A 1480 MHz 1.4 Gbit/s HBM2 4096-bit 720 GB/sec 16 GB HBM2 10.6 TFLOPS 5.3 TFLOPS N/A N/A N/A N/A 21.2 TFLOPS N/A N/A N/A N/A 160 GB/sec GP100 1344 KB (24 KB × 56) 4096 KB 300 W 610 mm2 15.3 B TSMC 16FF+ Q2 2016

See also

[edit]

References

[edit]
  1. ^ "NVIDIA 7nm Next-Gen-GPUs To Be Built By TSMC". Wccftech. June 24, 2018. Retrieved July 6, 2019.
  2. ^ "Samsung to Optical-Shrink NVIDIA "Pascal" to 14 nm". Retrieved August 13, 2016.
  3. ^ "Accelerating The Real-Time Ray Tracing Ecosystem: DXR For GeForce RTX and GeForce GTX". NVIDIA.
  4. ^ "Ray Tracing Comes to Nvidia GTX GPUs: Here's How to Enable It". April 11, 2019.
  5. ^ a b "NVIDIA GeForce GTX 1080" (PDF). International.download.nvidia.com. Retrieved September 15, 2016.
  6. ^ Gupta, Sumit (March 21, 2014). "NVIDIA Updates GPU Roadmap; Announces Pascal". Blogs.nvidia.com. Retrieved March 25, 2014.
  7. ^ "Parallel Forall". NVIDIA Developer Zone. Devblogs.nvidia.com. Archived from the original on March 26, 2014. Retrieved March 25, 2014.
  8. ^ "NVIDIA Tesla P100" (PDF). International.download.nvidia.com. Retrieved September 15, 2016.
  9. ^ "Inside Pascal: NVIDIA's Newest Computing Platform". April 5, 2016.
  10. ^ Denis Foley (March 25, 2014). "NVLink, Pascal and Stacked Memory: Feeding the Appetite for Big Data". nvidia.com. Retrieved July 7, 2014.
  11. ^ "NVIDIA's Next-Gen Pascal GPU Architecture to Provide 10X Speedup for Deep Learning Apps". The Official NVIDIA Blog. Retrieved March 23, 2015.
  12. ^ a b Smith, Ryan (April 5, 2015). "NVIDIA Announces Tesla P100 Accelerator - Pascal GP100 Power for HPC". AnandTech. Retrieved May 27, 2016. Each of those SMs also contains 32 FP64 CUDA cores - giving us the 1/2 rate for FP64 - and new to the Pascal architecture is the ability to pack 2 FP16 operations inside a single FP32 CUDA core under the right circumstances
  13. ^ a b c Smith, Ryan (July 20, 2016). "The NVIDIA GeForce GTX 1080 & GTX 1070 Founders Editions Review: Kicking Off the FinFET Generation". AnandTech. p. 9. Retrieved July 21, 2016.
  14. ^ a b c d e Smith, Ryan (July 20, 2016). "The NVIDIA GeForce GTX 1080 & GTX 1070 Founders Editions Review: Kicking Off the FinFET Generation". AnandTech. p. 10. Retrieved July 21, 2016.
  15. ^ "GTX 1080 Graphics Card". GeForce. Retrieved September 15, 2016.
  16. ^ Carbotte, Kevin (May 17, 2016). "Nvidia GeForce GTX 1080 Simultaneous Multi-Projection & Async Compute". Tomshardware.com. Retrieved September 15, 2016.
  17. ^ "Nvidia Pascal HDCP 2.2". Nvidia Hardware Page. Retrieved May 8, 2016.
  18. ^ a b Smith, Ryan (July 20, 2016). "The NVIDIA GeForce GTX 1080 & GTX 1070 Founders Editions Review: Kicking Off the FinFET Generation". AnandTech. p. 5. Retrieved July 21, 2016.
  19. ^ Smith, Ryan (July 20, 2016). "The NVIDIA GeForce GTX 1080 & GTX 1070 Founders Editions Review: Kicking Off the FinFET Generation". AnandTech. p. 4. Retrieved July 21, 2016.
  20. ^ Harris, Mark (April 5, 2016). "Inside Pascal: NVIDIA's Newest Computing Platform". Parallel Forall. Nvidia. Retrieved June 3, 2016.
  21. ^ "NVIDIA TITAN Xp Graphics Card with Pascal Architecture". NVIDIA.
  22. ^ "NVIDIA TITAN X Graphics Card with Pascal". GeForce. Retrieved September 15, 2016.
  23. ^ "New Quadro Graphics Built on Pascal Architecture". NVIDIA. Retrieved September 15, 2016.
  24. ^ "Accelerating Data Center Workloads with GPUs". NVIDIA. Retrieved September 15, 2016.
  25. ^ Zhiye Liu (October 22, 2018). "Nvidia GeForce GTX 1060 Gets GDDR5X in Fifth Makeover". Tom's Hardware. Retrieved February 2, 2024.
  26. ^ "NVIDIA GeForce 10 Series Graphics Cards". NVIDIA.
  27. ^ "NVIDIA GeForce GTX 1060 to be released on July 7th". VideoCardz.com. June 29, 2016. Retrieved September 15, 2016.
  28. ^ "GTX 1060 Graphics Cards". GeForce. Retrieved September 15, 2016.
  29. ^ Smith, Ryan (November 12, 2012). "NVIDIA Launches Tesla K20 & K20X: GK110 Arrives At Last". AnandTech. p. 3. Retrieved July 24, 2016.
  30. ^ a b c d e f Nvidia (September 1, 2015). "CUDA C Programming Guide". Retrieved July 24, 2016.
  31. ^ a b c d e f g h i j k l m n o Triolet, Damien (May 24, 2016). "Nvidia GeForce GTX 1080, le premier GPU 16nm en test !". Hardware.fr (in French). p. 2. Retrieved July 24, 2016.
  32. ^ a b Smith, Ryan (January 26, 2015). "GeForce GTX 970: Correcting The Specs & Exploring Memory Allocation". AnandTech. p. 1. Retrieved July 24, 2016.
  33. ^ "NVIDIA Turing Release Date". Techradar. February 2, 2021.
  34. ^ Smith, Ryan (March 22, 2022). "NVIDIA Hopper GPU Architecture and H100 Accelerator Announced: Working Smarter and Harder". AnandTech.
  35. ^ Smith, Ryan (May 14, 2020). "NVIDIA Ampere Unleashed: NVIDIA Announces New GPU Architecture, A100 GPU, and Accelerator". AnandTech.
  36. ^ "NVIDIA Tesla V100 tested: near unbelievable GPU power". TweakTown. September 17, 2017.
Retrieved from "https://en.wikipedia.org/w/index.php?title=Pascal_(microarchitecture)&oldid=1253162519"