What Did NVIDIA Announce About the Tesla V100 at GTC 2017? NVIDIA Keynote Part 6

TL;DR

NVIDIA announced the Tesla Volta V100 at GTC 2017 as a 12-nanometer FinFET processor with 21 billion processors, 5,000 processor cores, and a new tensor core delivering 120 teraflops of tensor operations. Its architecture also includes a 20-megabyte register file, 900 gigabytes-per-second memory, and second-generation NVLink. Read on for its specifications and performance comparisons with Pascal.

Transcript

I would like to introduce you to something that has taken several thousand Engineers several thousand Engineers several years to create it is a masterpiece on many levels it is the most complex project that has ever been undertaken arguably the most expensive computer project the world's ever done ladies and gentlemen the Tesla Volta v00 this is ma... Read More

Key Insights

• "every single transistor that is possible to make by today's physics was crammed into this processor" (1:15)
• "a brand new type of processor called tensor core which results in a 120 Tera flops of tensor operations" (2:09)
• "we've been able to achieve 900 gabyt per second it is just so fast" (3:33)
• "it literally does the 4x4 multiply plus C at the same time and it dumps into result 20 times increase throughput" (5:20)
• "Volta is one and a half times the floating Point performance general purpose Computing 12 times the tensor operations compared to Pascal for deep learning training and six times for inferencing" (5:49)

Questions & Answers

Q: What did NVIDIA announce about the Tesla V100 at GTC 2017?

NVIDIA introduced the Tesla Volta V100, describing it as one of its most complex and expensive computer projects. The processor uses TSMC's 12-nanometer FinFET process and reaches the reticle limits of photolithography.

Q: What are the Tesla Volta V100's core specifications?

The presentation states that the V100 has 21 billion processors, almost 100 billion vias, an 800-square-millimeter die, and approximately 5,000 processor cores. It delivers 7.5 teraflops of 64-bit floating-point performance and 15 teraflops of 32-bit floating-point performance.

Q: What is the tensor core in the NVIDIA Tesla V100?

The tensor core is a new CUDA tensor operation instruction that functions as both an instruction and a data format. It uses a 4x4 matrix processing array to perform the A times B plus C operation used in deep learning.

Q: How much tensor performance does the Tesla V100 provide?

The Tesla V100 provides 120 teraflops of tensor operations. Its tensor core performs the 4x4 multiplication plus C in parallel, producing a 20-times increase in throughput.

Q: How does the Tesla V100 compare with the Pascal P100?

The presentation says Volta delivers 1.5 times the general-purpose floating-point performance of Pascal. It also provides 12 times the tensor operations for deep learning training and six times the performance for inferencing.

Q: How is the Tesla V100 memory system organized?

The architecture has a 20-megabyte register file that keeps memory close to the processors, along with 16 megabytes of cache. Its Samsung memory achieves 900 gigabytes per second.

Q: How fast is the Tesla V100's second-generation NVLink?

The second-generation NVLink provides 300 gigabytes per second. The presentation describes this as approximately 10 times the speed of the fastest PCI Express available at the time.

Q: What is the difference between deep learning training and inferencing?

Training the network is the first step and is described as very computationally intensive. Inferencing is the second step, involving the production application of the network, and is also computationally intensive but not as intensive as training.

Summary & Key Takeaways

• The Tesla Volta V100 is a masterpiece project that incorporates the most advanced technology, with a chip made at the limits of photolithography.

• It contains 21 billion processors, almost 100 billion vas connectors, and 800 millim squared die size, making it an impressive feat of engineering.

• The Volta V100 also introduces a brand new type of processor called tensor core, capable of 120 teraflops of tensor operations.