Nvidia GeForce 20 series

Logo of the Nvidia GeForce 20 series

ASUS GeForce RTX 2070 ROG STRIX Advanced - 8GB GDDR6

The GeForce 20 series is a series of graphics chips from Nvidia and the successor to the GeForce 10 series .

In the run-up to Gamescom , Nvidia presented the GeForce 20 series on August 20, 2018. The series, code-named "Turing" (after the British mathematician Alan Turing ), first focused on ray tracing . In order to underline this aspect, Nvidia changed the prefix GTX, which was previously mostly used in sales names, to RTX . Another innovation is the use of GDDR6 memory . Sales started on September 20, 2018.

Turing architecture

The Turing architecture is the direct further development of the Volta architecture, which was never used in the graphics cards of the GeForce series. The Volta architecture was only used in the Titan series , as well as in the professional series, Quadro and Tesla . In the GeForce graphics cards, the Turing architecture replaces the Pascal architecture of the GeForce 10 series .

The graphics processors, which are based on the Turing architecture, are composed of so-called graphics processing clusters (GPC), some of which are also referred to as raster engines. Each graphics processing cluster contains 4 or 6 texture processing clusters (TPC), some of which are not fully activated. A texture processing cluster consists of two shader clusters, which Nvidia calls streaming multiprocessors (SM). These streaming multiprocessors, the function block that includes the most important units, have been significantly modified compared to the Pascal architecture and have been partially reorganized. Each streaming multiprocessor is assigned:

• 64 FP32 units for 32-bit floating point numbers and 2 FP64 units for 64-bit floating point numbers
• 64 INT32 units for 32-bit integers that can work in parallel with the floating point units
• 4 texture units, each consisting of a texture mapping unit and a texture address unit
• 16 load / store units
• 16 special function units
• 8 tensor units
• 1 ray tracing unit

At Pascal, a streaming multiprocessor still consisted of 128 FP32 units which can output either 32-bit floating point numbers (floating point) or 32-bit whole numbers (integers). At Turing, this system was abandoned and instead the FP32 unit was reduced to 64, and 64 new INT32 units were added. This allows both operations to be carried out in parallel. Alternatively, the FP32 units can also calculate 16-bit floating point numbers (half precision) in a ratio of 2: 1 (Pascal 1: 1).

The tensor cores for AI calculations were adopted from the Volta architecture. According to Nvidia, these are FP16 units for matrix calculations. You can also execute FP32, INT4 and INT8 commands, but this is only important for professional applications. According to Nvidia, only the FP16 calculations of the tensor cores are important for 3D applications.

The innovation of the Turing architecture highlighted most by Nvidia Marketing is the hardware-based ray tracing support. For this purpose, each streaming multiprocessor has a ray tracing unit, sometimes also referred to as an RT core. Since Nvidia does not provide any information on how these work, it is currently not possible to reconstruct what exactly an RT unit does. Nvidia only specifies the number of RT units per model and a filling rate of Gigarays per second.

Data overview

Graphics processors

Graphics chip	production			units							L2 cache	API support					Video pro- cessor	Bus interface stelle	NV link
	production process	transis- tors	The - area	ROPs	Unified shaders			Texture units	Tensor cores	RT cores		DirectX	OpenGL	OpenCL	CUDA	volcano
					GPC	SM	ALUs
TU102	12 nm	18.6 billion	754 mm²	96	6th	72	4608	288	576	72	6 MB	12.1	4.6	1.2	7.1	1.1.78	k. A.	PCIe 3.0	2 ×
TU104		13.6 billion	545 mm²	64	6th	48	3072	192	384	48	4 MB
TU106		10.6 billion	445 mm²	64	3	36	2304	144	288	36	4 MB								-

Desktop model data

model	Official launch	Graphics processor (GPU)									Graphics memory			performance						Power consumption
		Type	Active units						Chip clock		Memory size	Storage cycle	Storage interface	Memory bandwidth (GB / s)	Computing power ( TFlops )		Fill rate			MGCP	Readings
			ROPs	SM	ALUs	Texture units	Tensor cores	RT cores	Base	Boost					ALUs ( 16  /  32  /  64  bit)	Tensor cores ( 4  /  8  /  16  bit)	Rays (GR / s)	Pixels (GP / s)	Texel (GT / s)
																					Idle	3D
GeForce RTX 2060	Jan. 7, 2019	TU106-200-A1	48	30th	1920	120	240	30th	1365 MHz	1680 MHz	06 GB GDDR6	7000 MHz	192 bits	336	12.90 /  06.45 / 0.202	206/103/520	0≈5	80.6	201.6	160 W	k. A.	k. A.
GeForce RTX 2060 Super	9 Jul 2019	TU106-410-A1	64	34	2176	136	272	34	1470 MHz	1650 MHz	08 GB GDDR6		256 bit	448	14.36 /  07.18 / 0.224	230/115/570	0≈6	105.6	224.4	175 W
GeForce RTX 2070	17th Oct 2018	TU106-400-A1 TU106-400A-A1	64	36	2304	144	288	36	1410 MHz	1620 MHz					14.93 /  07.46 / 0.233	239/119/600	0≈6	103.7	233.3	175 W 185 W.
GeForce RTX 2070 Super	9 Jul 2019	TU104-410-A1	64	40	2560	160	320	40	1605 MHz	1770 MHz					18.12 /  09.06 / 0.283	290/145/720	0≈7	113.3	283.2	215 W
GeForce RTX 2080	Sep 20 2018	TU104-400-A1 TU104-400A-A1	64	46	2944	184	368	46	1515 MHz	1710 MHz					20.14 / 10.07 / 0.315	322/161/810	0≈8	109.4	314.6	215 W 225 W
GeForce RTX 2080 Super	23 Jul 2019	TU104-450-A1	64	48	3072	192	384	48	1650 MHz	1815 MHz					22.30 / 11.15 / 0.348	357/180/890	0≈8	116.2	348.5	250 W
GeForce RTX 2080 Ti	27 Sep 2018	TU102-300-K1-A1 TU102-300A-K1-A1	88	68	4352	272	544	68	1350 MHz	1545 MHz	11 GB GDDR6		352 bits	616	26.90 / 13.45 / 0.420	430/215/108	≈10	136	420.2	250 W 260 W
Nvidia Titan RTX	Dec 18, 2018	TU102-400-A1	96	72	4608	288	576	72	1350 MHz	1770 MHz	24 GB GDDR6		384 bits	672	32.63 / 16.31 / 0.510	522/261/131	≈11	129.6	388.8	280 W

Remarks

1. ↑ The time of the public presentation is given, not when the models are available.
2. ↑ All performance values ​​given are theoretical maximum values ​​for the GPU's boost clock. They are not directly comparable with other architectures. In practice, the performance is limited by the ability to parallelize the calculation and by the first limiting resource (shader, texture units, memory bandwidth, chip temperature, power consumption).
3. ↑ ^{a b} The specified clock rates are the reference data recommended or specified by Nvidia, the effective clock is specified for the memory clock. However, the exact clock rate can deviate by a few megahertz due to different clock generators, and the final definition of the clock rates is in the hands of the respective graphics card manufacturer. It is therefore entirely possible that there are or will be graphics card models that have different clock rates.
4. ↑ The MGCP value given by Nvidia corresponds to the maximum power consumption under standard conditions defined by Nvidia. It cannot be directly compared with values ​​from other manufacturers. With overclocking and good cooling, the values ​​can also be exceeded significantly.
5. ↑ The power consumption of graphics cards with Nvidia reference design is stated. The values ​​represent mean values ​​from various measurements on various websites that require special equipment. The values ​​fluctuate somewhat due to differences in the measurement loads and measurement methods and due to fluctuations in series.

Web links

Commons : Nvidia GeForce 20 series  - collection of images, videos and audio files

• Product overview on the manufacturer's website

Individual evidence

1. ↑ ^{a b c d} Turing TU102, -104, -106: The technology of the Nvidia GeForce RTX 2080 Ti, 2080 & 2070 (page 2). Computerbase, September 14, 2018, accessed July 4, 2019 . 
2. ↑ NVIDIA GeForce RTX 2060. Nvidia Corporation, January 7, 2019, accessed January 7, 2019 . 
3. ↑ NVIDIA GeForce RTX 2060 Super. Nvidia Corporation, July 2, 2019, accessed July 3, 2019 . 
4. ↑ NVIDIA GeForce RTX 2070. Nvidia Corporation, August 20, 2018, accessed September 28, 2018 . 
5. ↑ NVIDIA GeForce RTX 2070 Super. Nvidia Corporation, July 2, 2019, accessed July 3, 2019 . 
6. ↑ NVIDIA GeForce RTX 2080. Nvidia Corporation, August 20, 2018, accessed September 28, 2018 . 
7. ↑ NVIDIA GeForce RTX 2080 Super. Nvidia Corporation, July 2, 2019, accessed July 3, 2019 . 
8. ↑ NVIDIA GeForce RTX 2080 Ti. Nvidia Corporation, August 20, 2018, accessed September 28, 2018 . 
9. ↑ NVIDIA TITAN RTX now available. Retrieved on May 9, 2020 (German).