Nvidia announces Tesla V100 - Volta with 21 billion 12 nanometer transistors

True to tradition, Nvidia holds its annual GPU Technology Conference and during a session, the company’s founder and CEO Jen-Hsun Huang always takes the stage. Last year, the architecture Pascal was achieved with the calculation card Tesla P100 and for the 2017 edition, it’s time for the next step – Time.

Contents hide

1 Specifications: Nvidia Tesla V100

2 Related posts:

Block diagram for Nvidia “Volta” GV100

Screenshot (285) .png

Specifications: Nvidia Tesla V100

	Tesla V100	Tesla P100	Tesla M40	Tesla K40
Technical	12nm TSMC	16nm TSMC	28nm TSMC	28nm TSMC
Circuit	GV100	GP100	GM200	GK110
Circuit surface	815 mm²	610 mm²	601 mm²	551 mm²
Transistors	21.1 billion	15.3 billion	8 billion	7.1 billion
Architecture	Time	Pascal	Maxwell	Kepler
SM cluster	80 st.	56 st.	24 st.	15 st.
CUDA cores	5 120 st.	3 584 st.	3 072 st.	2 880 st.
Texture units	320 st.	224 st.	192 st.	240 st.
Clock frequency	1 370 MHz	1 328 MHz	948 MHz	745 MHz
GPU Boost	1 455 MHz	1 480 MHz	1 114 MHz	875 MHz
Calculating power (SP, FP32)	14 899 GFLOPS	10 608 GFLOPS	6 844 GFLOPS	5 040 GFLOPS
Calculating power (DP, FP64)	7 450 GLOPS	5 304 GFLOPS	214 GFLOPS	1 680 GFLOPS
Memory bus	4 096-bit	4 096-bit	384-bit	384-bit
Minnestyp	HBM2	HBM2	GDDR5	GDDR5
Memory amount	16 GB	16 GB	Up to 24 GB	Up to 12 GB
Memory frequency	~1 750 MHz	1 430 MHz	6 000 MHz	6 000 MHz
Memory bandwidth	900 GB/s	732 GB/s	288 GB/s	288 GB/s
TDP	300 W	300 W	250 W	235 W

Read This Now: Sapphire unveils the Nitro RX Vega 64 and RX Vega 56

The new calculation card is called Tesla V100 and debuts the architecture Volta. The GV100 graphics circuit is based on TSMC’s 12-nanometer technology, which is based on last year’s 16 nanometers but with optimizations for energy efficiency, performance and transistor density. This enables Nvidia to create the most advanced and largest circuit ever.

► Read all articles about #Nvidia Volta

The graphics circuit houses no less than 21 billion transistors, which is comparable to 15.3 billion transistors for the previous GP100 record holder. This with a circuit area of massive 815 mm², considerably larger than single circuits that have historically measured at most closer to 700 mm².

The number of CUDA cores for the Tesla V100 is 5,120 and the computing power is stated to be 15 TFLOPS for single precision (FP32) and 7.5 TFLOPS for double precision (FP64). In addition to this, there is a new type of Tensor cores that can perform calculations specifically for machine learning and with these, the theoretical computing power is as much as 120 TFLOPS.

Read This Now: Asus Radeon R9 280 Strix

The solid graphics circuit is surrounded by four 4 GB HBM2 capsules each with a clock frequency of 875 MHz (1,750 MHz efficient), giving a total of 16 GB of memory and a theoretical bandwidth of 900 GB / s. In terms of connectivity, the Tesla V100 supports second-generation NVLink with a bandwidth of 300 GB / s – almost ten times higher than 32 GB / s for PCI Express 3.0 x16.

Deliveries of the Nvidia Tesla V100 will begin during the third quarter of the year. It is not clear when the new graphics circuit will be released to consumers, but with Pascal in mind, it is likely that Nvidia will create a smaller and more energy-efficient circuit called the GV102 without computationally oriented features.