NV發佈Tesla V100:Volta架構 5120個CUDA - 3C
By Frederica
at 2017-05-11T02:46
at 2017-05-11T02:46
Table of Contents
Nvidia發佈Tesla V100計算卡:Volta架構,恐怖的5120個CUDA
http://www.expreview.com/54059.html
昨晚NVIDIA CEO黃仁勳在GTC 2017大會上發佈了基於Volta架構的旗艦計算卡 Tesla V100
,為了應付日益龐大的AI以及高性能計算需求,Tesla V100擁有超越上一代的5120個CUDA
單元,並且增加能提高深度計算性能的Tensor單元,因此電晶體數目足足有210億之多,
核心頻率為1455MHz ,16GB HBM2顯存,浮點性能更是達到單精確度浮點15 TFLOPS,雙精
度浮點7.5 TFLOPS。
對比上一代Tesla P100,Tesla V100最大變化就是增加了與深度學習高度相關的Tensor單
元,Tensor性能可以達到120 TFLOPS;而且CUDA數目暴增,由3584個增至5120個,增長了
42%;全新的台積電12nm FFN工藝 製造有史以來最大的815mm2 GPU核心(16nm的第四代改
良版本,更高的電晶體密度,更低的功耗;管中窺豹,估計Volta遊戲卡都是使用這個工
藝了);雖然依然是4096bit 16GB的HBM2顯存,但是頻寬已經Tesla P100 720GB/s提升至
900GB/s水準。
除了CUDA單元數量增加,Tesla V100為了更好提升高性能計算,繼續增加二級緩存及寄存
器大小,L2緩存由Tesla P100的4096KB增加到了6144KB,每組SM單元的寄存器檔案大小總
數從14336KB增加到了20480KB。
NVIDIA計算卡專享的NVLink是一種高頻寬的互聯技術,能夠在CPU-GPU和GPU-GPU之間實現
超高速的資料傳輸。 誕生之初就獲得傳統PCIe 3.0 5-12倍的資料傳送速率,還能大幅提
升應用程式的處理速度。 現在NVIDIA將其提升至300GB/s水準,當然了這個技術主要還是
用在高性能計算上,估計不會下放至遊戲卡上。
http://img.expreview.com/news/2017/05/11/Tesla-V1003.jpg
Tesla V100架構:
6組GPC單元,每組GPC單元由14組SM單元構成,滿血版應該是6 x 14 = 84組SM單元,但
Tesla V100只有80組,每組SM單元64個CUDA單元,因此共同構成80 x 64 = 5120個CUDA單
元。 每組SM單元中,FP32:FP64:Tensor單元比例為8:4:1。
http://img.expreview.com/news/2017/05/11/NVIDIA-Volta-GV1001.png
http://img.expreview.com/news/2017/05/11/tesla%20v1002.jpg
那麼問題來了,強大如斯的Tesla V100究竟用在哪裡? NVIDIA表示將會率先更新用於深
度學習超算DGX-1上,內部同樣集成了8張Tesla V100,提供960 TFLOPS Tensor深度計算
性能,能夠在8小時完成TITAN X需時8天的計算量,極大地節約科研人員等待深度計算結
果時間。 就當然啦,擁有8張真·核彈的新DGX-1售價並不便宜,149000美元。
--
老黃已經公布Volta 蘇媽的VEGA仍然保持神祕
--
http://www.expreview.com/54059.html
昨晚NVIDIA CEO黃仁勳在GTC 2017大會上發佈了基於Volta架構的旗艦計算卡 Tesla V100
,為了應付日益龐大的AI以及高性能計算需求,Tesla V100擁有超越上一代的5120個CUDA
單元,並且增加能提高深度計算性能的Tensor單元,因此電晶體數目足足有210億之多,
核心頻率為1455MHz ,16GB HBM2顯存,浮點性能更是達到單精確度浮點15 TFLOPS,雙精
度浮點7.5 TFLOPS。
對比上一代Tesla P100,Tesla V100最大變化就是增加了與深度學習高度相關的Tensor單
元,Tensor性能可以達到120 TFLOPS;而且CUDA數目暴增,由3584個增至5120個,增長了
42%;全新的台積電12nm FFN工藝 製造有史以來最大的815mm2 GPU核心(16nm的第四代改
良版本,更高的電晶體密度,更低的功耗;管中窺豹,估計Volta遊戲卡都是使用這個工
藝了);雖然依然是4096bit 16GB的HBM2顯存,但是頻寬已經Tesla P100 720GB/s提升至
900GB/s水準。
除了CUDA單元數量增加,Tesla V100為了更好提升高性能計算,繼續增加二級緩存及寄存
器大小,L2緩存由Tesla P100的4096KB增加到了6144KB,每組SM單元的寄存器檔案大小總
數從14336KB增加到了20480KB。
NVIDIA計算卡專享的NVLink是一種高頻寬的互聯技術,能夠在CPU-GPU和GPU-GPU之間實現
超高速的資料傳輸。 誕生之初就獲得傳統PCIe 3.0 5-12倍的資料傳送速率,還能大幅提
升應用程式的處理速度。 現在NVIDIA將其提升至300GB/s水準,當然了這個技術主要還是
用在高性能計算上,估計不會下放至遊戲卡上。
http://img.expreview.com/news/2017/05/11/Tesla-V1003.jpg
Tesla V100架構:
6組GPC單元,每組GPC單元由14組SM單元構成,滿血版應該是6 x 14 = 84組SM單元,但
Tesla V100只有80組,每組SM單元64個CUDA單元,因此共同構成80 x 64 = 5120個CUDA單
元。 每組SM單元中,FP32:FP64:Tensor單元比例為8:4:1。
http://img.expreview.com/news/2017/05/11/NVIDIA-Volta-GV1001.png
http://img.expreview.com/news/2017/05/11/tesla%20v1002.jpg
那麼問題來了,強大如斯的Tesla V100究竟用在哪裡? NVIDIA表示將會率先更新用於深
度學習超算DGX-1上,內部同樣集成了8張Tesla V100,提供960 TFLOPS Tensor深度計算
性能,能夠在8小時完成TITAN X需時8天的計算量,極大地節約科研人員等待深度計算結
果時間。 就當然啦,擁有8張真·核彈的新DGX-1售價並不便宜,149000美元。
--
老黃已經公布Volta 蘇媽的VEGA仍然保持神祕
--
Tags:
3C
All Comments
By Isabella
at 2017-05-13T05:08
at 2017-05-13T05:08
By Frederica
at 2017-05-16T13:31
at 2017-05-16T13:31
By Aaliyah
at 2017-05-18T20:19
at 2017-05-18T20:19
By Donna
at 2017-05-22T11:55
at 2017-05-22T11:55
By Audriana
at 2017-05-25T13:17
at 2017-05-25T13:17
By Robert
at 2017-05-29T11:45
at 2017-05-29T11:45
By Franklin
at 2017-06-02T09:45
at 2017-06-02T09:45
By Ida
at 2017-06-03T07:08
at 2017-06-03T07:08
By Catherine
at 2017-06-05T02:28
at 2017-06-05T02:28
By Lauren
at 2017-06-07T22:59
at 2017-06-07T22:59
By Suhail Hany
at 2017-06-10T06:54
at 2017-06-10T06:54
By Jessica
at 2017-06-15T00:15
at 2017-06-15T00:15
By Lydia
at 2017-06-17T10:19
at 2017-06-17T10:19
By Susan
at 2017-06-22T02:46
at 2017-06-22T02:46
By Leila
at 2017-06-24T15:57
at 2017-06-24T15:57
By Linda
at 2017-06-27T15:34
at 2017-06-27T15:34
By Margaret
at 2017-06-30T22:41
at 2017-06-30T22:41
By Isabella
at 2017-07-03T13:00
at 2017-07-03T13:00
By Cara
at 2017-07-05T02:20
at 2017-07-05T02:20
By Aaliyah
at 2017-07-09T10:07
at 2017-07-09T10:07
By Joe
at 2017-07-13T10:44
at 2017-07-13T10:44
By Ida
at 2017-07-15T01:02
at 2017-07-15T01:02
By Carol
at 2017-07-18T12:55
at 2017-07-18T12:55
By Hardy
at 2017-07-20T02:53
at 2017-07-20T02:53
By Megan
at 2017-07-20T06:20
at 2017-07-20T06:20
By Kelly
at 2017-07-24T06:21
at 2017-07-24T06:21
By Kumar
at 2017-07-25T11:24
at 2017-07-25T11:24
By Tristan Cohan
at 2017-07-28T00:37
at 2017-07-28T00:37
By Charlie
at 2017-07-28T16:57
at 2017-07-28T16:57
By Jake
at 2017-07-28T20:25
at 2017-07-28T20:25
By Ida
at 2017-08-01T21:45
at 2017-08-01T21:45
By Heather
at 2017-08-04T18:57
at 2017-08-04T18:57
By Wallis
at 2017-08-06T00:13
at 2017-08-06T00:13
By Zanna
at 2017-08-06T13:46
at 2017-08-06T13:46
By Jacob
at 2017-08-06T21:26
at 2017-08-06T21:26
By Iris
at 2017-08-11T03:10
at 2017-08-11T03:10
By Hazel
at 2017-08-16T02:21
at 2017-08-16T02:21
By Hazel
at 2017-08-20T04:05
at 2017-08-20T04:05
By Freda
at 2017-08-24T10:44
at 2017-08-24T10:44
By Quanna
at 2017-08-28T15:11
at 2017-08-28T15:11
By Mary
at 2017-08-31T01:51
at 2017-08-31T01:51
By Bethany
at 2017-09-03T10:58
at 2017-09-03T10:58
By Caroline
at 2017-09-04T21:41
at 2017-09-04T21:41
By Iris
at 2017-09-09T18:03
at 2017-09-09T18:03
By Caroline
at 2017-09-13T06:31
at 2017-09-13T06:31
By Ina
at 2017-09-14T09:43
at 2017-09-14T09:43
By Mary
at 2017-09-17T18:19
at 2017-09-17T18:19
By Tom
at 2017-09-22T11:13
at 2017-09-22T11:13
By Genevieve
at 2017-09-26T07:17
at 2017-09-26T07:17
By Lucy
at 2017-09-28T02:51
at 2017-09-28T02:51
By Ida
at 2017-10-02T17:16
at 2017-10-02T17:16
By Oscar
at 2017-10-06T01:20
at 2017-10-06T01:20
By Elma
at 2017-10-10T09:06
at 2017-10-10T09:06
By Olive
at 2017-10-12T23:10
at 2017-10-12T23:10
By Catherine
at 2017-10-15T05:36
at 2017-10-15T05:36
By Mason
at 2017-10-16T20:49
at 2017-10-16T20:49
By Joe
at 2017-10-20T19:06
at 2017-10-20T19:06
By Anthony
at 2017-10-21T19:54
at 2017-10-21T19:54
By Annie
at 2017-10-22T20:23
at 2017-10-22T20:23
By Dorothy
at 2017-10-24T12:27
at 2017-10-24T12:27
By Joseph
at 2017-10-24T20:32
at 2017-10-24T20:32
By Charlotte
at 2017-10-25T03:11
at 2017-10-25T03:11
By Delia
at 2017-10-28T19:07
at 2017-10-28T19:07
By Skylar DavisLinda
at 2017-11-02T11:49
at 2017-11-02T11:49
By Puput
at 2017-11-05T10:10
at 2017-11-05T10:10
By Robert
at 2017-11-10T00:56
at 2017-11-10T00:56
By Charlotte
at 2017-11-10T17:49
at 2017-11-10T17:49
By Yedda
at 2017-11-13T15:20
at 2017-11-13T15:20
By Bennie
at 2017-11-16T22:51
at 2017-11-16T22:51
Related Posts
R7 1700簡單開箱+坎坷歷程(圖少傷眼)
By Hedwig
at 2017-05-11T02:43
at 2017-05-11T02:43
30-35K 攝影PS修圖
By Dora
at 2017-05-11T01:30
at 2017-05-11T01:30
30k 模擬機
By Skylar Davis
at 2017-05-10T23:15
at 2017-05-10T23:15
G4560推的動擷取卡Elagato HD60 Pro嗎
By Caroline
at 2017-05-10T22:53
at 2017-05-10T22:53
35K遊戲實況機
By Daph Bay
at 2017-05-10T22:48
at 2017-05-10T22:48