NVIDIA unveils Hopper, its new {hardware} structure to remodel knowledge facilities into AI factories

NVIDIA did it once more, however this time with a twist — showing to borrow a web page from the competitors’s playbook. At NVIDIA GTC, which has formed into one of many AI trade’s most essential occasions, the corporate introduced the newest iteration of its {hardware} structure and merchandise. This is a breakdown of the bulletins and what they imply for the ecosystem at giant.

Hopper: NVIDIA’s new GPU structure

GTC, which started Monday and runs by Thursday, options 900+ periods. Greater than 200,000 builders, researchers, and knowledge scientists from 50+ international locations have registered for the occasion. At his GTC 2022 keynote, NVIDIA founder and CEO Jensen Huang introduced a wealth of reports in knowledge heart and high-performance computing, AI, design collaboration and digital twins, networking, automotive, robotics, and healthcare.

Huang’s framing was that “corporations are processing, refining their knowledge, making AI software program … changing into intelligence producers.” If the objective is to remodel knowledge facilities into ‘AI Factories,’ as NVIDIA places it, then inserting Transformers on the coronary heart of this is smart.

The centerfold within the bulletins has been the brand new Hopper GPU Structure, which NVIDIA dubs “the subsequent technology of accelerated computing.” Named for Grace Hopper, a pioneering U.S. pc scientist, the brand new structure succeeds the NVIDIA Ampere structure, launched two years in the past. The corporate additionally introduced its first Hopper-based GPU, the NVIDIA H100.

NVIDIA claims that Hopper brings an order of magnitude efficiency leap over its predecessor, and this feat is predicated on six breakthrough improvements. Let’s undergo them, holding fast notes of how they examine to the competitors.

First, manufacturing. Constructed with 80 billion transistors utilizing a cutting-edge TSMC 4N course of designed for NVIDIA’s accelerated compute wants, H100 options main advances to speed up AI, HPC, reminiscence bandwidth, interconnect, and communication, together with practically 5 terabytes per second of exterior connectivity. On the manufacturing stage, upstarts similar to Cerebras or Graphcore have been additionally pushing the boundaries of what is doable.

Second, Multi-Occasion GPU (MIG). MIG know-how permits a single GPU to be partitioned into seven smaller, totally remoted cases to deal with several types of jobs. The Hopper structure extends MIG capabilities by as much as 7x over the earlier technology by providing safe multitenant configurations in cloud environments throughout every GPU occasion. Run:AI, a accomplice of NVIDIA, presents one thing comparable as a software program layer, going by the title of fractional GPU sharing.

Third, confidential computing. NVIDIA claims H100 is the world’s first accelerator with confidential computing capabilities to guard AI fashions and buyer knowledge whereas they’re being processed. Prospects also can apply confidential computing to federated studying for privacy-sensitive industries like healthcare and monetary companies, in addition to on shared cloud infrastructures. This isn’t a function now we have seen elsewhere.

Fourth, 4th-Technology NVIDIA NVLink. To speed up the biggest AI fashions, NVLink combines with a brand new exterior NVLink Swap to increase NVLink as a scale-up community past the server, connecting as much as 256 H100 GPUs at 9x increased bandwidth versus the earlier technology utilizing NVIDIA HDR Quantum InfiniBand. Once more, that is NVIDIA-specific, though opponents typically leverage their very own specialised infrastructure to attach their {hardware} too.

Fifth, DPX directions to speed up dynamic programming. Dynamic programming is each a mathematical optimization methodology and a pc programming methodology, initially developed within the Fifties. By way of mathematical optimization, dynamic programming normally refers to simplifying a choice by breaking it down right into a sequence of determination steps over time. Dynamic programming is principally an optimization over plain recursion.

NVIDIA notes that dynamic programming is utilized in a broad vary of algorithms, together with route optimization and genomics, and it will probably pace up execution by as much as 40x in contrast with CPUs and as much as 7x in contrast with previous-generation GPUs. We’re not conscious of a direct equal within the competitors, though many AI chip upstarts additionally leverage parallelism.

The sixth innovation is the one we deem a very powerful: a new Transformer engine. As NVIDIA notes, transformers are the usual mannequin alternative for pure language processing, and probably the most essential deep studying fashions ever invented. The H100 accelerator’s Transformer Engine is constructed to hurry up these networks as a lot as 6x versus the earlier technology with out shedding accuracy. This deserves additional evaluation.

The Transformer Engine on the coronary heart of Hopper

Trying on the headline for the brand new transformer engine on the coronary heart of NVIDIA’s H100, we had been reminded of Intel architect Raja M. Koduri’s remarks to ZDNet’s Tiernan Ray. Koduri famous that the acceleration of matrix multiplications is now an important measure of the efficiency and effectivity of chips, which implies that each chip might be a neural web processor.

Koduri was spot on after all. Apart from Intel’s personal efforts, that is what has been driving a brand new technology of AI chip designs from an array of upstarts. Seeing NVIDIA consult with a transformer engine made us wonder if the corporate made a radical redesign of its GPUs. GPUs weren’t initially designed for AI workloads in spite of everything, they only occurred to be good at them, and NVIDIA had the foresight and acumen to construct an ecosystem round them.

Going deeper into NVIDIA’s personal evaluation of the Hopper structure, nonetheless, the notion of a radical redesign appears to be dispelled. Whereas Hopper does introduce a brand new streaming multiprocessor (SM) with many efficiency and effectivity enhancements, that is so far as it goes. That is not shocking, given the sheer weight of the ecosystem constructed round NVIDIA GPUs and the large updates and potential incompatibilities a radical redesign would entail.

Breaking down the enhancements in Hopper, reminiscence appears to be a giant a part of it. As Fb’s product supervisor for PyTorch, the favored machine studying coaching library, informed ZDNet, “Fashions hold getting greater and larger, they’re actually, actually huge, and actually costly to coach.” The largest fashions lately typically can’t be saved fully within the reminiscence circuits that accompany a GPU. Hopper comes with reminiscence that is sooner, extra, and shared amongst SMs.

One other enhance comes from NVIDIA’s new fourth-generation tensor cores, that are as much as 6x sooner chip-to-chip in comparison with A100. Tensor cores are exactly what’s used for matrix multiplications. In H100, a brand new FP8 knowledge kind is used, leading to 4 occasions sooner compute in comparison with earlier technology 16-bit floating-point choices. On equal knowledge varieties, there nonetheless is a 2x speedup.

As for the so-called “new transformer engine,” it seems that is the time period NVIDIA makes use of to consult with “a mix of software program and customized NVIDIA Hopper Tensor Core know-how designed particularly to speed up transformer mannequin coaching and inference.”

NVIDIA notes that the transformer engine intelligently manages and dynamically chooses between FP8 and 16-bit calculations, mechanically dealing with re-casting and scaling between FP8 and 16-bit in every layer to ship as much as 9x sooner AI coaching and as much as 30x sooner AI inference speedups on giant language fashions in comparison with the prior technology A100.

So whereas this isn’t a radical redesign, the mix of efficiency and effectivity enhancements lead to a 6x speedup in comparison with Ampere, as NVIDIA’s technical weblog elaborates. NVIDIA’s deal with bettering efficiency for transformer fashions is under no circumstances misplaced.

Transformer fashions are the spine of language fashions used extensively at the moment, similar to BERT and GPT-3. Initially developed for pure language processing use circumstances, their versatility is more and more being utilized to pc imaginative and prescient, drug discovery, and extra, as now we have been documenting in our State of AI protection. In line with a metric shared by NVIDIA, 70% of printed analysis in AI within the final 2 years is predicated on transformers.

The software program facet of issues: excellent news for Apache Spark customers

However what concerning the software program facet of issues? In earlier GTC bulletins, software program stack updates had been a key a part of the information. On this occasion, whereas NVIDIA-tuned heuristics that dynamically select between FP8 and FP16 calculations are a key a part of the brand new transformer engine internally, updates to the external-facing software program stack appear much less essential as compared.

NVIDIA’s Triton Inference Server and NeMo Megatron framework for coaching giant language fashions are getting updates. So are Riva, Merlin, and Maxin — a speech AI SDK that features pre-trained fashions, an end-to-end recommender AI framework, and an audio and video high quality enhancement SDK, respectively. As NVIDIA highlighted, these are utilized by the likes of AT&T, Microsoft, and Snapchat.

There are additionally 60 SDK updates for NVIDIA’s CUDA-X Libraries. NVIDIA selected to spotlight rising areas similar to accelerating quantum circuit simulation (cuQuantum normal availability) and 6G physical-layer analysis (Sionna normal availability). Nevertheless, for many customers, the excellent news might be within the replace within the RAPIDS Accelerator for Apache Spark, which speeds processing by over 3x with no code adjustments.

Whereas this was not precisely distinguished in NVIDIA’s bulletins, we expect it must be. An in a single day 3x speedup with out code adjustments for Apache Spark customers, with 80 % of the Fortune 500 utilizing Apache Spark in manufacturing, isn’t any small information. It isn’t the primary time NVIDIA reveals Apache Spark customers some love both.

General, NVIDIA appears to be sustaining its momentum. Whereas the competitors is fierce, with the headstart NVIDIA has managed to create, radical redesigns might probably not be known as for.