The green tech giant has just completed the 2022 edition of its traditional GTC, and there was some very heavy stuff on the menu.
Each year, the Nvidia GPU Technology Conference is one of the unmissable highlights on the calendar for professionals in AI and other disciplines related to high-performance computing (HPC). And 2022 will be no exception; the firm unveiled last night the first products of its new Hopper architecture. On the menu: a brand new H100 GPU chip and a CPU “superchip” which are already giving the professionals concerned cold sweats.
During its conference, the green team reminded everyone why it is in the top 10 of the most valuable companies in the world. Indeed, Nvidia has perfectly negotiated the shift to AI over the past decade by identifying the potential of this technology very early on; it was therefore able to offer excellent products which certainly contributed to its development.
But machine learning continues to progress. The firm explains that mega-neural networks, like those that produced the groundbreaking Alpha Fold database, become immensely complex. In the case of GPT-2, Google’s dedicated language AI, the network had about 1.5 billion distinct parameters in 2019; last year, with the GPT-3 version, this figure exceeded 1.5 trillion (millions of millions)!
The problem is that the more this complexity increases, the longer the models become to train. Unless there is a real algorithmic revolution, there is therefore no secret: you have to work hard on the hardware side. And that, Nvidia has understood, because it is precisely the number one objective of this Hopper architecture.
Indeed, the firm claims that this architecture is specially designed to accelerate the training of so-called “Transformer” networks. Nothing to do with the films of Michael Bay; very briefly, this term simply refers to a machine learning technology on which some of the most complex AI systems in the world are based, such as GPT-3.
Nvidia H100: an AI-optimized GPU behemoth
According to Nvidia, this new Hopper architecture would allow the new H100 GPUs to drive these models up to… six times faster. In addition, H100 will be compatible with the brand new version of NVLink, the successor to SLI. It will therefore be possible to mount up to 256 H100 GPUs in series! Simply hair-raising, knowing that this chip is already a little monster in itself.
Indeed, from the height of its 80 billion transistors, H100 is the very first GPU in the world to support the PCIe Gen 5 standard. In itself, this is already important information since the bandwidth of this new standard has been more or less doubled compared to the PCIe 4 standard. In addition, it will also use the HBM3 interface, and will therefore achieve a memory bandwidth of approximately 3 TB/s.
All of these elements put together, the H100 GPU would be approximately three times faster thanA100 for basic operations. Impressive knowing that the latter is a monster that we find today in heaps of supercomputers peak. Also, on giant Transformers models like GPT-3, the performance would be 9 times better. A veritable torrent of data which will have to be managed.
Grace: a double “superchip” that promises to be breathtaking
For this, Nvidia also presented Grace, its new “CPU superchip”. And it is quite a fascinating object for hardware enthusiasts. In essence, they are two separate CPUS, but linked by technology analogous to the NVLink GPU; this is called NVLink-C2C.
Does this remind you of something ? This is normal: recently, Apple unveiled its M1 Ultra processor which is, technically, a couple of M1 Max chips agglomerated on the same support. The technique is quite different, but it is clear that this approach, which consists of accumulating existing chips, is on the rise.
But Grace is boxing in a whole different category than Apple’s chip; this chip embeds no less than 144 Arm cores with a bandwidth of 1 TB/s. In other words, you will not find this machine in a standard computer. We’re not talking about a consumer CPU here, but a war machine designed to serve as part of “massive-scale” servers and HPC machines.
Grace can be used in two different scenarios. It will be compatible with servers based on a 100% CPU architecture; but it will also be able to drive an armada of H100 chips in GPU-accelerated machines.
The future king of AI supercomputers is on the way
To put all this beautiful hardware to use, Nvidia also announced that it was going to build a new “AI supercomputer”. They are quite different machines from traditional supercomputers; as their name suggests, they are optimized specifically for applications related to artificial intelligence.
She says that once completed, this machine called Eos will be “the fastest supercomputer in the world”. And to achieve this, the green team intends to put its new Hopper architecture to work.
Indeed, the machine will carry no less than 4600 copies of this new H100 GPU. Enough to obliterate the famous symbolic bar of the exaflop; such a machine would even be able to offer 18.4 exaflops of “AI performance”, or 18.4 billion billion operations every second. Enough to shake the most convoluted neural networks in the world.
With all this new state-of-the-art equipment available, we can therefore legitimately expect new spectacular breakthroughs in the world of artificial intelligence, and this in the fairly near future.