Two years after the launch of Turing in September 2020, NVIDIA changed the architecture of their graphics cards to Ampere . AMD did not stand aside and soon after that also updated the RDNA architecture to the second version.
, RX 6000 NVIDIA. , , ( , ), .
, ? ?
. AMD 520 2, Navi. : Instinct MI100 750 2.
AMD Navi 21 Radeon R9 Fury Nano, GCN 3.0 Fiji. 596 2.
2018 AMD 7- TSMC, Vega 20 ( Radeon VII) 331 2. Navi , N7P.
, , NVIDIA, . Ampere, GA102, 628 2. 17% , , TU102: 754 2. NVIDIA GA100: , 7- 826 2. , .
NVIDIA Navi 21 , , — . GA102 28,3 , AMD 5% — 26,8 .
, , , , — , . Navi 21 51,5 , GA102 — 41,1 .
Navi 21 TSMC N7P, N7. GA102 NVIDIA Samsung. NVIDIA 8- ( 8N 8NN). 7 8 : , .
, .
Ampere GA102 RDNA 2 Navi 21
, , , .
, . , , 3D- .
, , . , , , , , .
, , .
AMD Navi 21 NVIDIA GA102 . ,
, NVIDIA : , . , .
, Pascal GP106 ( GeForce GTX 1060) GP104 ( GeForce GTX 1070). , - . .
NVIDIA . ? , PCI Express . , - , , , . , , (ROP).
, , , , . , 3D- — , - . , , , , .
, , AMD Navi 21 Navi 10, Radeon RX 5700 XT. , , , . , Radeon RX 6800 XT 1825 2250 , , TDP, 300 . Radeon RX 5700 XT: 1605 , 1905 225 .
Ampere RDNA 2 , , AMD TSMC - — Radeon RX 6800 Radeon VII .
Radeon VII — AMD TSMC 7- , 64%. : Ampere GA102, NVIDIA TSMC.
, AMD NVIDIA , : , API, , . : , .
. RDNA 2 , : Graphics Command Processor, — (ACE).
, . , , .
AMD Shader Engine (SE), NVIDIA (GPC): , .
: . . , - : , — , — .
RDNA 2 SE — , :
Primitive Setup — , () ;
— ;
(ROP) — , .
Primitive Setup 1 . , , 1,8 2,2 , . Ampere , .
AMD, NVIDIA . NVIDIA Raster Engines. , 1 , — , .
SE Navi 21 128 ROP; GA102 NVIDIA 112 ROP. , AMD , ROP , . - , . SE/GPC.
Shader Engines AMD , (DCU), Navi 21 DCU SE — , Workgroup Processors (WGP). Ampere GA102 (TPC), 6 TPC. 1 , NVIDIA , AMD, TPC, Navi 21 — SE. , GA102 , 42 Primitive Setup, RDNA 2 AMD — 4. Raster Engine TPC, GA102 7 , Navi 21 — . , NVIDIA .
— (CU) RDNA 2 (SM) Ampere — -«».
, , , , . , , - .
, , . , , , .
CU SM AMD NVIDIA . , .
NVIDIA
Turing Pascal, Ampere — , . , , Turing CUDA SM.
Turing ( ), 16x INT32 16x FP32. 32- : INT , FP — .
NVIDIA , SM Ampere 128 CUDA, , , — , Turing . INT32 , . Ampere NVIDIA , FP32. , CUDA SM , .
INT/FP , SM Ampere 128 FP32 64 FP32 64 INT32 . Turing .
, FP32 . , . GeForce RTX 3080, GA102 68 SM.
, FP32 121% GeForce 2080 Ti, 31%. ?
: , FP32.
NVIDIA Turing 2018 , , 36% , , INT32. , .
, FP32 , . SM , 32 , , FP32. SM Ampere , Turing.
, GeForce RTX 3080 11- FP32 2080 Ti INT+FP. , .
? SM , , Turing. (, FP16 FP16), 32 .
Fine-Grained Structured Sparsity. , , . , , , .
: CUDA, , BVH , SM . , , .
, , Optix API NVIDIA.
, , , . , Turing , .
AMD — RDNA 2?
-
, AMD : - SIMD32, SISD, . , . , AMD .
— , Ampere. BVH, RDNA 2 SIMD 32.
, , , , , . : , .
Ray Accelerator , . , , . , , . , NVIDIA SM, Ampere RDNA 2 , .
, Gears 5 Radeon RX 6800 ( Navi 21 60 CU) 17%, Shadow of the Tomb Raider — 52%. , NVIDIA RTX 3080 ( 68 SM GA102) 23% 40% .
AMD, , , , .
, RDNA 2 : , INT4 INT8. , AMD (CDNA) , DirectML.
API Microsoft DirectX 12. . NVIDIA, , DLSS. SM — , DirectML, , . Turing, Ampere , FP16.
RDNA 2 — , 32- 16-.
?
AMD SIMD32 , . 32 , , . , , SM Ampere, 32 .
NVIDIA SM 128 FMA FP32 , RDNA 2 64 . FP16 128 FMA , , Ampere FP16.
SM NVIDIA (, 64 FP32 64 INT32) FP16, . AMD SIMD32, , .
, , Ampere : GA102 SM, Navi 21, , , . AMD .
Ampere. , : 2- 50% (Turing TU102 4096 , ), 1- SM .
, - L1 , , . (, ) :
64 ;
48 ;
16 .
, . NVIDIA Micron GDDR6 . , GDDR6, . , 1 , (PAM), GDDR6X :
GDDR6X 2 , . GeForce RTX 3090 24 GDDR6X, 19 /, 936 /. 52% , GeForce RTX 2080 Ti. HBM2, , GDDR6.
Micron, PAM4 , . AMD : , , , - . RDNA 2 — .
. 128 - ? SRAM L3, AMD - 64 . 16 , 64 .
Infinity Cache 1,94 , 1986,6 /. DRAM, . , BVH , Infinity Cache .
, RDNA 2 , Zen 2: , (victim cache) . , , , DRAM.
, , 128 , Infinity Cache 32 L2. GDDR6 DRAM.
AMD — , — .
?
GDDR6X GA102 , . - 3- Navi 21 DRAM, .
AMD GDDR6 , , , GeForce RTX 3080 3090, Micron. GDDR6 , GDDR6X 8 .
RDNA 2, , , , Ampere, SRAM , DRAM, .
. Ampere RDNA 2 mesh- variable rate- DirectX12 Ultimate, NVIDIA .
mesh- , . , .
variable rate-. , . , , , .
(ROP), , variable rate-. NVIDIA ROP 2- .
Turing ROP 512 . ROP , , Ampere ROP GPC. GA102 12 ROP GPC ( 1 ), 112 .
AMD , NVIDIA (. . - L2), ROP / . Navi 21 ROP 8 32- 4 64-.
RTX IO NVIDIA — , , , CUDA.
, : , , , DRAM .
: . NVIDIA , «100- » «20- », , , .
AMD RDNA 2 Radeon RX 6000, Smart Access Memory. RTX IO — , . PCI Express 256 . (BAR), 2008 PCI Express 2.0 , . , , DRAM .
, , , . Windows Ryzen 5000, 500 Radeon RX 6000.
: 15% 4K. , NVIDIA , RTX 3000 .
-
HDMI 2.1 DisplayPort 1.4a. , 4K 240 HDR 8K 60 . 4:2:0, DSC 1.2a. , . HDMI 2.1 6 / 4K 60 .
Ampere RDNA 2 (FreeSync AMD, G-Sync NVIDIA), .
, , 8K AV1, 4K H.264 8K H.265, , . . , - .
,
AMD NVIDIA . , 3D- , .
NVIDIA , Ampere: GA100, GA102 GA104.
— GA102. GA100 — .
CUDA INT32+FP32 — FP64, load/store - L1/L2. , .
GA102/104, , , NVIDIA: , , . Ampere — .
RDNA 2 , , Ampere. AMD GCN .
, RDNA 2 Big Navi, , CDNA Big Vega: Instinct MI100 Arcturus — 50- 128 .
NVIDIA Quadro Tesla, Navi 21 . - , Ampere ?
: .
AMD Radeon RX 6900 XT, Navi 21 ( CU), , GeForce RTX 3090 . GA102 , NVIDIA «»-, Turing .
, , RDNA 2 Xbox Series X/S PlayStation 5, . , GCN Xbox One PlayStation 4, , , , .
, 2013 , , GCN 1.0, . Xbox One X, 2017 , GCN 2.0, 3 .
— , Xbox One PS4, , AMD? . , RDNA 2 , .
, , . NVIDIA AMD , : Ampere , RDNA 2 — .
, . , : Intel Xe. , , , .