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 business’s most vital occasions, the corporate introduced the most recent 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 via Thursday, options 900+ periods. Greater than 200,000 builders, researchers, and information scientists from 50+ nations have registered for the occasion. At his GTC 2022 keynote, NVIDIA founder and CEO Jensen Huang introduced a wealth of reports in information middle 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 information, making AI software program … changing into intelligence producers.” If the purpose is to rework information 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 following era of accelerated computing.” Named for Grace Hopper, a pioneering U.S. laptop 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, maintaining 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 almost 5 terabytes per second of exterior connectivity. On the manufacturing stage, upstarts comparable to Cerebras or Graphcore have been additionally pushing the boundaries of what is potential.
The NVIDIA H100 GPU, the primary to make the most of the brand new Hopper structure NVIDIA
Second, Multi-Occasion GPU (MIG). MIG expertise permits a single GPU to be partitioned into seven smaller, absolutely remoted situations to deal with several types of jobs. The Hopper structure extends MIG capabilities by as much as 7x over the earlier era by providing safe multitenant configurations in cloud environments throughout every GPU occasion. Run:AI, a associate of NVIDIA, gives one thing related as a software program layer, going by the identify 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 information whereas they’re being processed. Prospects also can apply confidential computing to federated studying for privacy-sensitive industries like healthcare and monetary providers, in addition to on shared cloud infrastructures. This isn’t a function we now have seen elsewhere.
Fourth, 4th-Era NVIDIA NVLink. To speed up the most important AI fashions, NVLink combines with a brand new exterior NVLink Change 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 era 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 often refers to simplifying a call 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 may possibly pace up execution by as much as 40x in contrast with CPUs and as much as 7x in contrast with previous-generation GPUs. We aren’t conscious of a direct equal within the competitors, though many AI chip upstarts additionally leverage parallelism.
The sixth innovation is the one we deem an important: a new Transformer engine. As NVIDIA notes, transformers are the usual mannequin selection for pure language processing, and one of the vital vital 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 era with out dropping accuracy. This deserves additional evaluation.
The Transformer Engine on the coronary heart of Hopper
Wanting on the headline for the brand new transformer engine on the coronary heart of NVIDIA’s H100, we have 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 a vital measure of the efficiency and effectivity of chips, which signifies that each chip might be a neural web processor.
Koduri was spot on after all. In addition to Intel’s personal efforts, that is what has been driving a brand new era 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 any case, they simply 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 stunning, 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, instructed ZDNet, “Fashions hold getting greater and larger, they’re actually, actually massive, and actually costly to coach.” The most important fashions today typically can’t be saved fully within the reminiscence circuits that accompany a GPU. Hopper comes with reminiscence that is quicker, extra, and shared amongst SMs.
One other enhance comes from NVIDIA’s new fourth-generation tensor cores, that are as much as 6x quicker chip-to-chip in comparison with A100. Tensor cores are exactly what’s used for matrix multiplications. In H100, a brand new FP8 information sort is used, leading to 4 occasions quicker compute in comparison with earlier era 16-bit floating-point choices. On equal information sorts, there nonetheless is a 2x speedup.
H100 compute enchancment abstract NVIDIA
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 expertise 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, robotically dealing with re-casting and scaling between FP8 and 16-bit in every layer to ship as much as 9x quicker AI coaching and as much as 30x quicker AI inference speedups on giant language fashions in comparison with the prior era A100.
So whereas this isn’t a radical redesign, the mixture of efficiency and effectivity enhancements end in a 6x speedup in comparison with Ampere, as NVIDIA’s technical weblog elaborates. NVIDIA’s concentrate on bettering efficiency for transformer fashions is by no means misplaced.
Transformer fashions are the spine of language fashions used extensively at the moment, comparable to BERT and GPT-3. Initially developed for pure language processing use instances, their versatility is more and more being utilized to laptop imaginative and prescient, drug discovery, and extra, as we now have been documenting in our State of AI protection. In keeping with a metric shared by NVIDIA, 70% of revealed analysis in AI within the final 2 years is predicated on transformers.
The software program aspect of issues: excellent news for Apache Spark customers
However what in regards to the software program aspect of issues? In earlier GTC bulletins, software program stack updates have 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 vital 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 focus on rising areas comparable to accelerating quantum circuit simulation (cuQuantum common availability) and 6G physical-layer analysis (Sionna common 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 modifications.
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 modifications for Apache Spark customers, with 80 % of the Fortune 500 utilizing Apache Spark in manufacturing, isn’t any small information. It is not the primary time NVIDIA exhibits Apache Spark customers some love both.
Total, NVIDIA appears to be sustaining its momentum. Whereas the competitors is fierce, with the headstart NVIDIA has managed to create, radical redesigns could probably not be known as for.