Tesla V100 is the most powerful accelerator, which will accelerate the development of high performance computing and artificial intelligence.
Месяц назад Nvidia представила graphics card Tesla V100, первый ускоритель нового поколения Volta – тогда еще в виде карты типа Mezzanine (SXM2). Теперь в ассортименте производителя появилась его версия под стандартный разъем PCI-Express x16.
Tesla V100 PCIe also uses a graphics processor with Volta GV100 5120 stream processors and 640 kernels and tensor 16 ГБ памяти HBM2 4096-bit. Изменилась тактовая частота ядра, because it works with a maximum frequency of about 1370 MHz (in this version SXM2 1455 MHz).
The nucleus consists of Volta GV100 80 blocks SM, which combine to give 5120 stream processors. Новинкой же являются 640 tensor core units, which are used for machine learning and building neural networksDespite the frequency change, Map offers a similar computing power - 28 Half precision teraflops, 14 Teraflops single-precision and 7 ТЕРАФЛОПС двойной точности (in this version SXM2 respectively 30, 15 and 7,5 TeraFLOPS). Вычислительная мощность при глубоком обучении в свою очередь составляет 112 instead 120 TeraFLOPS. Пропускная способность памяти осталась без изменений и составляет до 900 GB / sec.
Tesla V100 PCIe uses the PCI-Express interface 3.0 x16, so when connecting multiple cards bandwidth is "only" 32 GB / sec (SXM2 version can be used with NVLink bus bandwidth 300 Gbit / s). Но более низкие частоты повлияли на низкое потребление электроэнергии, since TDP rate is only 250 instead 300 AT.
| Model | Tesla P100 (SXM2) | Tesla P100 (PCIe) | Tesla V100 (SXM2) | Tesla V100 (PCIe) |
| Generation | Nvidia Pascal | Nvidia Pascal | Nvidia Volta | Nvidia Volta |
| Lithograph | TSMC 14 nm FinFET |
TSMC 14 nm FinFET |
12 nm TSMC FFN |
12 nm TSMC FFN |
| core area | 610 mm2 | 610 mm2 | 815 mm2 | 815 mm2 |
| graphics processor | Pascal GP100 | Pascal GP100 | Volta GV100 | Volta GV100 |
| Core Frequency | 1480 MHz | 1300 MHz | 1455 MHz | ~ 1370 MHz |
| Computing power FP16 | 21,2 TeraFLOPS | 18,7 TeraFLOPS | 30 TeraFLOPS | 28 TeraFLOPS |
| Computing power FP32 | 10,6 TeraFLOPS | 9,3 TeraFLOPS | 15 TeraFLOPS | 14 TeraFLOPS |
| Computing power FP64 | 5,3 TeraFLOPS | 4,7 TeraFLOPS | 7,5 TeraFLOPS | 7 TeraFLOPS |
| Computing power tensor (Deep Learning |
– | – | 120 TFLOPS | 112 TeraFLOPS |
| video memory | 16 HBM2 GB 4096-bit | 16 HBM2 GB 4096-bit | 16 HBM2 GB 4096-bit | 16 HBM2 GB 4096-bit |
| Memory Bandwidth | 720 Gbit / s | 720 Gbit / s | 900 GB / sec | 900 GB / sec |
| type of card | Mezzanine (SXM2) | PCIe 3.0 x16 | Mezzanine (SXM2) | PCIe 3.0 x16 |
| Cooling | passive | passive | passive | passive |
| TDP | 300 AT | 250 AT | 300 AT | 250 AT |
Tesla V100 PCIe card should be available later this year – both in the range of Nvidia, and partner companies (Hewlett-Packard Enterprise, eg, He announced three systems, working on the basis of this design).
