New series of GPU cards

[Dirk Broer]

nVidia just seem to have released a GT1030 specially against the Radeon RX 550. What is so special about the Radeon RX 500 series?
Let's start with the RX 550


Model/Type
GPU
Fab
(nm)
Shaders
TMU
ROP
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt


Radeon RX 500 Series






1/16
of SP








Radeon RX 540 (OEM)

Cape Verde PRX

28

512

32

16

1248

78.0

?

65

19.20

1.20

?



Radeon RX 550

Polaris 12

14

512

32

16

1211

75.7

1211

50

24.22

1.51

24.22



Radeon RX 560

Polaris 11

14

1024

64

16

2406

150.4

2406

80

30.01

1.88

30.01



Radeon RX 570

Polaris 20

14

2048

128

32

4784

300.0

4784

150

31.89

2.00

31.89



Radeon RX 580

Polaris 20

14

2304

144

32

5792

362.0

5792

185

31.31

2.13

31.31



Radeon Pro Duo Polaris

2x Polaris 20

14

4608

288

64

11456

716.0

11456

250

45.82

2.86

45.82


The RX 550 is a very capable card in view of its power needs, a mere 50 Watt. Those willing to spend a little more on purchase and power bills can choose for the even more capable RX 560
The TBP (Typical Board Power) of the RX 560 is actually 60-80 Watt, so the GFLOPS/Watt may be even better yet.

What does nVidia offer?


Model/Type
GPU
Fab
(nm)
Shaders
TMU
ROP
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt


Geforce 10 Series






1/32
of SP








GeForce GT 1030

GP108-300-A1

16

384

24

8

952

30

15

30

31.73

1.00

0.50



GeForce GTX 1050

GP107-300-A1

14

640

40

32

1733

54

27

75

23.11

0.72

0.36



GeForce GTX 1050 Ti

GP107-400-A1

14

768

48

32

1981

62

31

75

26.41

0.83

0.41



GeForce GTX 1060 3GB

GP106-300-A1

16

1152

72

48

3470

108

54

120

28.92

0.90

0.45



GeForce GTX 1060 6GB

GP106-400/410-A1

16

1280

80

48

3855

120

60

120

32.13

1.00

0.50



GeForce GTX 1070

GP104-200-A1

16

1920

120

64

5783

181

90

150

38.56

1.21

0.60



GeForce GTX 1080

GP104-400/410-A1

16

2560

160

64

8228

257

128

180

45.71

1.43

0.71



GeForce GTX 1080 Ti

GP102-350-K1-A1

16

3584

224

88

10609

332

166

250

42.44

1.33

0.66



nVidia Titan X

GP102-400-A1

16

3584

224

96

10157

317

159

250

40.63

1.27

0.64



nVidia Titan Xp

GP102-450-A1

16

3840

240

96

12150*

380*

190*

250

48.61

1.52

0.76


*=Normal values unknown yet. For the Titan Xp I only know of the 'Turbo boost' values, at the moment.

Considering the fact that most projects only ask for Single precision, the GT 1030 offers real value. But when we take double precision performance (e.g. MilkyWay) into account, nVidia is fighting a lost battle as they have sacrificed that feature to boost the sales of their professional (Quadro/Tesla) cards. I do not know yet of the effects of the half-precision performance, but nVidia seems to have made the wrong choice there too by offering Half-precision at 1/64th of the Single precision while AMD offers it at the same GFLOPS as the single precision. There ought to be projects where AMD cards smash the nVidia cards, performance-wise.

[NeoGen]

Nice table there Dirk... and indeed it seems in raw computation capacity and computation per watt, which is what matters most in distributed computing, we have the winners in this tier! :)

[vaughan]

Good work Dirk, I'd love to see all the cards GTX1060, GTX1070 through to GTX1080 Ti too :)

[Dirk Broer]

Good work Dirk, I'd love to see all the cards GTX1060, GTX1070 through to GTX1080 Ti too :)

Blistering performance equals blistering power bills:5eek: too, but I'll take care of it

[Dirk Broer]

nVidia just seem to have released a GTX 1650/1660 specially against the Radeon Vega and upcoming Navi. What is so special about the Radeon Vega series? Will Navi be even better?
But let's first start with the RX 500 series


Model/Type
GPU
Fab
(nm)
Shaders
TMUs
ROPs
Compute
Units
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt


Radeon RX 500 Series







1/16
of SP








Radeon RX 540

Lexa

14

512

32

16

8

1151

71.9

1151

50

23.02

1.44

23.02



Radeon RX 550

Lexa

14

512

32

16

8

1247

77.95

1247

50

24.94

1.56

24.94



Radeon RX 560

Baffin

14

1024

64

16

16

2611

163.2

2611

75

34.81

2.18

34.81



Radeon RX 570

Polaris 20XL

14

2048

128

32

32

5095

318.5

5095

120

42.46

2.65

42.46



Radeon RX 580

Polaris 20XTX

14

2304

144

32

36

6175

385.9

6175

185

33.38

2.09

33.38



Radeon RX 590

Polaris 30XT

12

2304

144

32

36

7119

445.0

7119

175

40.68

2.54

40.68



Radeon Pro Duo Polaris

Ellesmere Gemini G

14

2x 2304

2x 144

2x 32

2x 36

11456

716.0

11456

250

45.82

2.86

45.82


Radeon Vega Series







1/16
of SP








Radeon Vega 56

1x Vega 10XL

14

3584

224

64

56

10570

660.4

21140

210

50.33

3.14

100.66



Radeon Vega 64

1x Vega 64

14

4096

256

64

64

12580

786.4

25160

295

42.64

2.67

85.29


Radeon VII Series







1/4
of SP








Radeon VII

1x Vega 20XT

7

3840

240

64

60

13440

3360

26880

295

45.56

11.39

91.12


Radeon 5000 Series







1/16
of SP








Radeon RX 5500

1x Navi 14 XT

7

1408

88

32

22

5196

324.7

10392

150

34.64

2.16

69.28



Radeon RX 5500 XT

1x Navi 14 XTX

7

1408

88

32

22

5196

324.7

10392

130

39.97

2.50

79.94



Radeon RX 5600

1x Navi 10 XE

7

2048

128

64

32

6390

399.4

12780

150

42.60

2.66

85.20



Radeon RX 5600 XT

1x Navi 10 XLE

7

2304

144

64

36

8066

504.0

16131

160

50.41

3.15

100.81



Radeon RX 5700

1x Navi 10XL

7

2304

144

64

36

7949

496.8

15900

180

44.16

2.76

88.33



Radeon RX 5700XT

1x Navi 10XT

7

2560

160

64

40

9754

609.6

19510

225

43.35

2.70

86.71



Radeon RX 5700XT 50th Anniversary

1x Navi 10XTX

7

2560

160

64

40

10138

633.6

20276

235

43.14

2.69

86.28



The RX 550 is a very capable card in view of its power needs, a mere 50 Watt. Those willing to spend a little more on purchase and power bills can choose for the even more capable RX 560
The TBP (Typical Board Power) of the RX 560 is actually 60-80 Watt, so the GFLOPS/Watt may be even better yet. The Radeon VII seems tailor-made for MilkyWay@Home.

What does nVidia offer?


Model/Type
GPU
Fab
(nm)
Shaders
TMUs
ROPs
Compute
Units
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt


Geforce 10 Series
Pascal






1/32
of SP








GeForce GT 1030

GP108-300-A1

14

384

24

16

3

1127

35.23

17.62

30

37.57

1.17

0.59



GeForce GTX 1050

GP107-300-A1

14

640

40

32

5

1862

58.2

29.1

75

24.83

0.78

0.39



GeForce GTX 1050 Ti

GP107-400-A1

14

768

48

32

6

2138

66.82

33.41

75

28.51

0.89

0.45



GeForce GTX 1060 3GB

GP106-300-A1

16

1152

72

48

9

3935

123

61.5

120

32.79

1.03

0.51



GeForce GTX 1060 6GB

GP106-400/410-A1

16

1280

80

48

10

4375

136.7

68.35

120

36.46

1.14

0.57



GeForce GTX 1070

GP104-200-A1

16

1920

120

64

15

6463

202.0

101.0

150

43.09

1.35

0.67



GeForce GTX 1070 Ti

GP104-300-A1

16

2432

152

64

19

8186

255.8

127.9

180

45.48

1.42

0.71



GeForce GTX 1080

GP104-400/410-A1

16

2560

160

64

20

8873

277.3

138.6

180

49.29

1.54

0.77



GeForce GTX 1080 Ti

GP102-350-K1-A1

16

3584

224

88

28

11340

354.4

177.2

250

45.36

1.42

0.71



nVidia Titan X

GP102-400-A1

16

3584

224

96

28

10970

342.9

171.5

250

43.88

1.37

0.69



nVidia Titan Xp

GP102-450-A1

16

3840

240

96

30

12150

379.7

189.8

250

48.60

1.52

0.76


Geforce 20 Series
Turing






1/32
of SP








GeForce RTX 2060

TU106-200A-KA-A1

12

1920

120

48

30

6451

201.6

12902

160

40.32

1.26

80.64



GeForce RTX 2060 Super

TU106-410-A1

12

2176

136

64

34

7181

224.4

14362

160

44.88

1.40

89.76



GeForce RTX 2070

TU106-400-A1

12

2304

144

64

36

7465

233.3

14930

175

42.66

1.33

85.31



GeForce RTX 2070 Super

TU104-410-A1

12

2560

160

64

40

9062

283.2

18124

215

42.15

1.32

84.30



GeForce RTX 2080

TU104-400-A1

12

2944

184

64

46

10070

314.6

20140

215

46.84

1.46

93.67



GeForce RTX 2080 Super

TU104-450-A1

12

3072

192

64

48

11150

348.5

22300

215

51.86

1.62

103.72



GeForce RTX 2080 Ti

TU102-300-K1-A1

12

4352

272

88

68

13450

420.2

26900

250

53.8

1.68

107.6



nVidia Titan RTX

TU102-400-A1

12

4608

288

96

72

16310

509.8

32620

280

58.25

1.82

116.5


Geforce 16 Series
Turing






1/32
of SP








GeForce GTX 1650

TU117-300-A1

12

896

56

32

14

2984

93.2

5967

75

39.79

1.24

79.56



GeForce GTX 1650 Ti

TU117-400-A1

12

1024

64

32

16

3533

110.4

7066

75

47.11

1.47

94.21



GeForce GTX 1650 Super

TU116-250-KA-A1

12

1280

80

32

20

4416

138.0

8832

100

44.16

1.38

88.32



GeForce GTX 1660

TU116-300-A1

12

1408

88

48

22

5027

157.1

10054

120

41.89

1.31

83.78



GeForce GTX 1660 Super

TU116-300-A1

12

1408

88

48

22

5027

157.1

10054

125

40.22

1.26

80.43



GeForce GTX 1660 Ti

TU116-400-A1

12

1536

96

48

24

5437

169.9

10874

120

45.31

1.42

90.62



Volta






1/2
of SP








nVidia Titan V

GV100-400-A1

12

5120

320

96

80

14900

7450

29800

250

59.6

29.8

119.2



nVidia Titan V CEO edition

GV100

12

5120

320

128

80

14900

7450

29800

250

59.6

29.8

119.2


Considering the fact that most projects only ask for Single Precision, the GT 1030 offers real value. But when we take half or double precision performance (e.g. MilkyWay) into account, nVidia is fighting a lost battle as they have sacrificed those features to boost the sales of their professional (Quadro/Tesla) or semi-professional (Titan V) cards. I do not know yet of the effects of the half-precision performance, but nVidia seemed to have made the wrong choice in the GeForce 10-series by offering Half-Precision at 1/64th of the Single Precision while AMD offers it at the same or later even at double the GFLOPS as compared to the Single Precision. There ought to be projects where the AMD cards smash the nVidia GeForce 10-series cards, performance-wise. nVidia hasn't made that mistake with the GeForce 20- and GeForce 16-series though. The GeForce 16xx 'Super' models feature GDDR6 RAM (like the GeForce 20 series), with increased bandwidth. The old Titan V's can still be used as a Radeon VII killer and offer an even better DP performance.

[vaughan]

My eVGA GTX1660 ti runs really well, considering it is running 24/7.

From what I read somewhere, I think it was at Serve The Home the AMD vega cards pump out too much heat.

[Dirk Broer]

My eVGA GTX1660 ti runs really well, considering it is running 24/7.

From what I read somewhere, I think it was at Serve The Home the AMD vega cards pump out too much heat.

One can only hope that the NAVI cards, produced at 7nm, will solve that problem

[Nflight]

Thank you Dirk, this mesmerizing chart has gone and caused me to miss an appt. in my busy schedule. :( Well at least I am going to follow in Vaughan's footsteps with my next GPU choice! Thank you Vaughan and Dirk!

[plonk420]

here's a "review thread" for the 1650. sadly, there's no TechReport review. i loooove their frametimes/frames spent past x milliseconds. GamersNexus has a few frametime graphs, tho

https://www.reddit.com/r/hardware/comments/bgg8v0/gtx_1650_launch_thread/

the RX570 slaps it down nicely, tho. (tho i might suggest a used Fury for ~$110-130 on eBay if you've got the power)

on the other hand, looks like Navi might be GCN (https://www.phoronix.com/scan.php?page=news_item&px=Open-Source-Navi-GFX1010). again. i'll reserve judgement for after reviews, but at least this tempers what i'm going to expect from it

[Dirk Broer]

I have updated the table for even newer GPU's (and adjusted the values to those given at https://www.techpowerup.com/ ) Some big changes!

[plonk420]

not impressed with compute performance (tho not surprised), but i AM impressed by its gaming performance!

[Dirk Broer]

not impressed with compute performance (tho not surprised), but i AM impressed by its gaming performance!

The GeForce GTX 1650 Ti, or the GeForce GTX 1660 Ti?

[Dirk Broer]

Updated the charts, Radeon RX 5600 XT is our present champion

Here he/she is, compared to some former champions:



Model/Type
GPU
Fab
(nm)
Shaders
TMUs
ROPs
Compute
Units
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt



Radeon HD 5970

Hemlock XT
=2x Cypress XT

40

2x 1600

2x 80

2x 2

20

4640

928.0

-

294

15.78

3.16

-



Radeon HD 7990

Malta XT
=2x Tahiti XT2

28

2x 2048

2x 128

2x 32

32

8192

2048.0

-

375

21.85

5.46

-



Radeon RX 5600 XT

1x Navi 10 XLE

7

2304

144

64

36

8066

504.0

16131

160

50.41

3.15

100.81

[plonk420]

can't remember. probably both.

has the 5700/5700XT/5600XT gotten any better at compute since launch?

[Dirk Broer]

can't remember. probably both.

has the 5700/5700XT/5600XT gotten any better at compute since launch?

I don't know for the 5600XT, which is new in the charts. Wikipedia reports higher values than I had before for the 5700/5700XT.

[Nflight]

Here is a recent article on the competition between GPU's https://www.tomshardware.com/features/amd-radeon-rx-5600-xt-vs-nvidia-geforce-rtx-2060?utm_source=notification

[Nflight]

And then someone posts this ideal solution, ARM server involved: https://www.anandtech.com/show/15621/marvell-announces-thunderx3-96-cores-384-thread-3rd-gen-arm-server-processor?utm_source=notification

[Dirk Broer]

And then someone posts this ideal solution, ARM server involved: https://www.anandtech.com/show/15621/marvell-announces-thunderx3-96-cores-384-thread-3rd-gen-arm-server-processor?utm_source=notification

Erh.....where's the GPU in the Cavium ThunderX3 server? It is a fantastic ARM CPU, 96 Cores/384 Threads, ARMv8.3 instruction set, but does the ThunderX3 server have PCIe slots?

Edit: I found a Cavium ThunderX2 server with a nVidia GT 710 (https://www.anandtech.com/show/12571/gigabyte-thunderxstation-cavium-thunderx2-socs).

BOINC-wise the only caveat is the screaming lack of 64-bit Linux ARM applications.
With the few projects that have such apps you can sure make a big dent using an dual ThunderX3 machine.....768 threads pumping WUs direction project server.
You would need 128 Odroid-N2's to do the same, or 192 Raspberry Pi 4's!

[Dirk Broer]

Updated the charts, Radeon RX 5600 XT is our present champion

Here he/she is, compared to some former champions:



Model/Type
GPU
Fab
(nm)
Shaders
TMUs
ROPs
Compute
Units
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt



Radeon HD 5970

Hemlock XT
=2x Cypress XT

40

2x 1600

2x 80

2x 2

20

4640

928.0

-

294

15.78

3.16

-



Radeon HD 7990

Malta XT
=2x Tahiti XT2

28

2x 2048

2x 128

2x 32

32

8192

2048.0

-

375

21.85

5.46

-



Radeon RX 5600 XT

1x Navi 10 XLE

7

2304

144

64

36

8066

504.0

16131

160

50.41

3.15

100.81



And here are the new 6000-series:


Model/Type
GPU
Fab
(nm)
Shaders
TMUs
ROPs
Compute
Units
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt



Radeon RX 6400

1x Navi 24XL

6

768

48

32

12

3,565

222.8

7,130

53

67.26

4.20

134.52



Radeon RX 6500XT

1x Navi 24XT

6

1024

64

32

16

5,765

360.3

11,530

113

51.02

3.19

102.04



Radeon RX 6600

1x Navi 23XL

7

1792

112

64

28

8,928

558.0

17,860

132

67.64

4.23

135.30



Radeon RX 6600XT

1x Navi 23XT

7

2048

128

64

32

10,605

662.8

21,209

160

66.28

4.14

132.56



Radeon RX 6700

1x Navi 22 XTL

7

2304

144

64

36

10,370

648

20,740

180

57.61

3.60

115.22



Radeon RX 6700XT

1x Navi 22 XT

7

2560

192

64

40

11,520

720

23,040

200

57.60

3.60

115.20



Radeon RX 6800

1x Navi 21 XL

7

3840

240

96

60

16,166

1,010

32,333

250

64.66

4.04

129.33



Radeon RX 6800XT

1x Navi 21 XT

7

4608

288

128

72

20,736

1,296

41,472

300

69.12

4.32

138.24



Radeon RX 6900XT

1x Navi 21 XTX

7

5120

320

128

80

23,040

1,440

46,080

300

76.80

4.80

153.60


What would the 6600XT be capable of, I wonder?

And here are the competitors from the green camp:


Model/Type
GPU
Fab
(nm)
Shaders
TMUs
ROPs
Compute
Units
GFLOP SP
GFLOP DP
GFLOP HP
TBP
Gflop(SP)
/Watt
Gflop(DP)
/Watt
Gflop(HP)
/Watt



GeForce RTX 3050

GA106-150

8

2560

80

32

20

9,098

142

9,098

130

69.98

1.09

69.98



GeForce RTX 3060

GA106-300-A1

8

3584

112

64

28

12,740

199

12,740

170

74.94

1.17

74.94



GeForce RTX 3060 Ti

GA104-200-A1

8

4864

152

80

38

16,197

253

16,197

200

80.99

1.27

80.99



GeForce RTX 3070

GA104-300-A1

8

5888

184

96

46

20,372

318

20,372

220

92.60

1.45

92.60



GeForce RTX 3080

GA102-200-K1-A1

8

8704

272

96

68

29,768

465

29,768

320

93.03

1.45

93.03



GeForce RTX 3090

GA102-300-A1

8

10496

328

112

82

35,686

558

35,686

350

101.96

1.59

101.96


If it is single precision you are after, nVidia is the card. For double or half precision: go AMD. nVidia is also the brand to go for in case of Tensor-based computation at the moment.

[Jason1478963]

Great work on the charts. It is interesting and helpful to see the performance per watt for out next investments.

[Dirk Broer]

Great work on the charts. It is interesting and helpful to see the performance per watt for out next investments.

One can only hope for a RX 6500 and a RTX 3050 for the sake of crunching on a budget, I can't afford 300 Watt cards running 24/7...

[NeoGen]

Those double precision values (as well as per watt) on the 6900 XT are making me drool... :)

[Dirk Broer]

I added the GeForce RTX 3600 and the Radeons RX 6700 and 6700XT

[Dirk Broer]

Radeons RX 6400, RX 6500XT and RX 6600 added