NEC Vector Engine | |
Developer | NEC |
Manufacturer | TSMC |
Type | Vector processors |
Introduction | 2017 (announced) |
Production | 2018 |
µarch | SX-Aurora |
Word size | 64 bit 8 octets
16 nibbles |
Process | 16 nm 0.016 μm
1.6e-5 mm |
Technology | CMOS |
Vector Engine (VE) is a family of vector processors designed as PCIe accelerator cards designed by NEC.
Overview
NEC introduced the Vector Engine (VE) in 2017 as the successor to the SX line of supercomputers. VEs depart from all prior generations by departing from the traditional self-hosted vector processor and nodes. With the introduction of the Vector Engine, NEC moved to an accelerator card architecture whereby a Vector Engine (VE) PCIe card are installed onto standard x86 server which serves as the Vector Host (VH).
Members
Type 10
- See also: SX-Aurora
Vector Engine Type 10 (VE10) are first-generation Vector Engines. Those processors are based on the SX-Aurora microarchitecture and are fabricated on TSMC 16 nm process. Type 10 features eight vector cores along with six HBM2 stacks.
- Proc 16 nm process
- Mem 24 GiB / 48 GiB
- HBM 4-Hi (750 GB/s) / 8-Hi (1.2 TB/s) HBM2
- Perf 2.150-2.458 teraFLOPS
List of Vector Engine Type 10 Processors | |||||||
---|---|---|---|---|---|---|---|
Vector Processor | HBM2 | ||||||
Model | Launched | Cores | L3$ | Frequency | Performance | Memory | Bandwidth |
Type 10A | 2018 | 8 | 16 MiB 16,384 KiB 16,777,216 B 0.0156 GiB | 1.6 GHz 1,600 MHz 1,600,000 kHz | 2.46 teraFLOPS 2,460,000,000,000 FLOPS 2,460,000,000 KFLOPS 2,460,000 MFLOPS 2,460 GFLOPS 0.00246 PFLOPS | 48 GiB 49,152 MiB 50,331,648 KiB 51,539,607,552 B 0.0469 TiB | 1,229 GB/s 1,144.595 GiB/s 1,172,065.735 MiB/s 1,229,000 MB/s 1.118 TiB/s 1.229 TB/s |
Type 10B | 2018 | 8 | 16 MiB 16,384 KiB 16,777,216 B 0.0156 GiB | 1.4 GHz 1,400 MHz 1,400,000 kHz | 2.15 teraFLOPS 2,150,000,000,000 FLOPS 2,150,000,000 KFLOPS 2,150,000 MFLOPS 2,150 GFLOPS 0.00215 PFLOPS | 48 GiB 49,152 MiB 50,331,648 KiB 51,539,607,552 B 0.0469 TiB | 1,229 GB/s 1,144.595 GiB/s 1,172,065.735 MiB/s 1,229,000 MB/s 1.118 TiB/s 1.229 TB/s |
Type 10C | 2018 | 8 | 16 MiB 16,384 KiB 16,777,216 B 0.0156 GiB | 1.4 GHz 1,400 MHz 1,400,000 kHz | 2.15 teraFLOPS 2,150,000,000,000 FLOPS 2,150,000,000 KFLOPS 2,150,000 MFLOPS 2,150 GFLOPS 0.00215 PFLOPS | 24 GiB 24,576 MiB 25,165,824 KiB 25,769,803,776 B 0.0234 TiB | 768 GB/s 715.256 GiB/s 732,421.875 MiB/s 768,000 MB/s 0.698 TiB/s 0.768 TB/s |
Count: 3 |
Type 10E
- See also: SX-Aurora
Type 10 E was announced in late 2019 and entered production in early 2020. The new cards offer similar specifications to the prior generation but offer higher memory bandwidth (E for enhanced memory bandwidth).
- Proc 16 nm process
- Mem 24 GiB / 48 GiB
- HBM 4-Hi (1 TB/s) / 8-Hi (1.35 TB/s) HBM2
- Perf 2.150-2.433 teraFLOPS
List of Vector Engine Type 10E Processors | |||||||
---|---|---|---|---|---|---|---|
Vector Processor | HBM2 | ||||||
Model | Launched | Cores | L3$ | Frequency | Performance | Memory | Bandwidth |
Type 10AE | January 2020 | 8 | 16 MiB 16,384 KiB 16,777,216 B 0.0156 GiB | 1.584 GHz 1,584 MHz 1,584,000 kHz | 2.433 teraFLOPS 2,433,000,000,000 FLOPS 2,433,000,000 KFLOPS 2,433,000 MFLOPS 2,433 GFLOPS 0.00243 PFLOPS | 48 GiB 49,152 MiB 50,331,648 KiB 51,539,607,552 B 0.0469 TiB | 1,382 GB/s 1,287.088 GiB/s 1,317,977.905 MiB/s 1,382,000 MB/s 1.257 TiB/s 1.382 TB/s |
Type 10BE | January 2020 | 8 | 16 MiB 16,384 KiB 16,777,216 B 0.0156 GiB | 1.408 GHz 1,408 MHz 1,408,000 kHz | 2.163 teraFLOPS 2,162,700,000,000 FLOPS 2,162,700,000 KFLOPS 2,162,700 MFLOPS 2,162.7 GFLOPS 0.00216 PFLOPS | 48 GiB 49,152 MiB 50,331,648 KiB 51,539,607,552 B 0.0469 TiB | 1,382 GB/s 1,287.088 GiB/s 1,317,977.905 MiB/s 1,382,000 MB/s 1.257 TiB/s 1.382 TB/s |
Type 10CE | January 2020 | 8 | 16 MiB 16,384 KiB 16,777,216 B 0.0156 GiB | 1.4 GHz 1,400 MHz 1,400,000 kHz | 2.15 teraFLOPS 2,150,000,000,000 FLOPS 2,150,000,000 KFLOPS 2,150,000 MFLOPS 2,150 GFLOPS 0.00215 PFLOPS | 24 GiB 24,576 MiB 25,165,824 KiB 25,769,803,776 B 0.0234 TiB | 998.4 GB/s 929.832 GiB/s 952,148.437 MiB/s 998,400 MB/s 0.908 TiB/s 0.998 TB/s |
Count: 3 |
Type 20
Type 20 is planned for the 2020-21 timeframe. NEC says it will feature higher memory bandwidth as well as higher core count and frequency.
Type 30
Type 30 is planned for the 2022 timeframe. NEC says Type 30 will feature a new architecture as well as higher memory bandwidth and higher core count and frequency.
See also
designer | NEC + |
first announced | 2017 + |
full page name | nec/vector engine + |
instance of | integrated circuit family + |
main designer | NEC + |
manufacturer | TSMC + |
microarchitecture | SX-Aurora + |
name | NEC Vector Engine + |
process | 16 nm (0.016 μm, 1.6e-5 mm) + |
technology | CMOS + |
word size | 64 bit (8 octets, 16 nibbles) + |