The modern graphics processor (GPU) started as an accelerator for Windows video games, but in the last 20 years it has become an enterprise server processor for high-performance computing and artificial intelligence applications.
Now GPUs are at the top of the copy for the performance used in supercomputers, AI training and inferences, drug research, financial modeling and medical imaging. They’ve also been applied to more popular tasks for situations where processors just aren’t fast enough, as with GPUs. relational databases.
As the demand for GPUs grows, so will competition between GPU vendors for servers, and there are only three: Nvidia, AMD and (soon) Intel. Intel has tried and failed twice to come up with an alternative to others’ GPUs, but is making another attempt.
The importance of GPUs in data centers
These three providers recognize the demand for GPUs in data centers as a growing opportunity. This is because GPUs are better suited to processors than many of the computations required by AI and machine learning in enterprise data centers and hyperscaler networks. Processors can do the job; it just takes them longer.
Because GPUs are designed to solve complex mathematical problems in parallel by dividing them into separate tasks that they work on at the same time, they solve them faster. To achieve this, they have multiple cores, much more than general-purpose CPUs. For example, Intel’s Xeon server processors have up to 28 cores, while AMD’s Epyc server processors have up to 64. Unlike Nvidia’s current generation of GPUs, Ampere has 6,912 cores, all running in parallel to do one thing: math, in particular floating-point mathematics.
The performance of GPUs is measured in how many of these math floating point operations they can perform per second or FLOPS. This number sometimes specifies the standardized floating-point format used when making the measure, such as FP64.
So what’s next for GPUs? Very little, as it turns out. Nvidia, AMD and Intel have put their cards on the table for their immediate plans and it looks like this will be fierce competition. Here’s what Nvidia, AMD and Intel have to offer.
Nvidia unveiled its GPU roadmap for the year in March with the announcement of its Hopper GPU architecture, claiming that it could deliver three to six times higher performance than its previous architecture, Ampere, which weighs 9.7 TFLOPS from FP64. Nvidia says the Hopper H100 will reach 60 TFLOPS FP64 performance.
Like previous GPUs, the Hopper H100 GPU can run as a standalone processor running on an add-on PCI Express board in a server. But Nvidia will also pair it with a custom Arm processor called Grace, which it has developed and expects to be available in 2023.
For Hopper, Nvidia has done more than just boost the GPU processor. It also modifies the low-power dual-data rate (LPDDR) 5 memory commonly used in smartphones to create LPDDR5X. It supports error correction code (ECC) and twice the bandwidth of memory than traditional DDR5 memory, for 1TBps bandwidth.
Together with Hopper, Nvidia announced NVLink 4, its latest connection between GPU and GPU. NVLink 4C2C allows Hopper GPUs to talk directly to each other with a maximum total bandwidth of 900 GB – seven times faster than if connected via PCIe Gen5 bus.
“If you’re thinking about data center products, you have three components, and they all need to move forward at the same pace. It’s memory, CPU and communications, “said John Peddie, president of Jon Peddie Research. “And Nvidia did that with Hopper. These three technologies are not in sync, but Nvidia has managed to do so. “
Nvidia plans to ship the Hopper GPU from the third quarter of 2022. OEM partners include Atos, BOXX Technologies, Cisco, Dell Technologies, Fujitsu, GIGABYTE, H3C, Hewlett Packard Enterprise, Inspur, Lenovo, Nettrix and Supermicro.
Due to continued supply pressures at its chipmaker TSMC, Nvidia has opened the door to work with Intel’s foundry business, but warned that such a deal would be in years to come.
AMD has the wind in its back. Sales increased quarter-on-quarter, its market share in the x86 processor grew, and in February it completed the acquisition of Xilinx and its on-the-spot programmable port arrays (FPGAs), adaptive chip systems (SoC), AI engines and software expertise. AMD is expected to release its Zen 4 CPU by the end of 2022.
AMD’s new graphics game processors based on its RDNA 3 architecture are also expected this year. AMD is silent on the specifications of RDNA 3, but gaming bloggers are spreading unconfirmed rumors of a 50% to 60% increase in performance over RDNA 2.
Meanwhile, AMD has begun shipping the Instinct MI250 line of graphics accelerators for enterprise computing, significantly faster than the previous MI100 series. The memory bus has doubled from 4096 bits to 8192 bits, the memory bandwidth has doubled to 3.2 TBps from 1.23 TBps, and performance has more than quadrupled from 11.5 TFLOPS of FP64 performance to 47 , 9 TFLOPS. It’s slower than AMD’s Hopper 60TFLOPS, but still competitive.
Daniel Newman, chief analyst at Futurum Research, said AMD’s ability to grab market share will come with the growth of the AI market. And he said he believes AMD’s success in the CPU market can help GPU sales. “What AMD has really created in the last five, seven years is a pretty strong loyalty that can be passed on,” he said. “The question is, can they significantly increase AI / HPC’s market share?”
He said the answer could be yes, because the company has been extremely good at finding market opportunities and managing its supply chain to achieve its goals. And with CEO Lisa Su at the helm, “It’s very difficult for me to exclude AMD in any area where they’ve decided to compete at the moment,” he said.
Jonathan Kassel, chief analyst for advanced computing, AI and IoT at Omdia, said he believes AMD’s success with its Epyc server processors will open up the Instinct processor.
“I think over time we can see that AMD is using its microprocessor success in the data center and using that to get companies looking. [Instinct]. “I think we’ll see that AMD is trying to use its customer relationships to try to expand its presence there,” he said.
Instinct has been available since Q1 2022. So far, its most popular use is the Oak Ridge National Labs supercomputer, which features a lot of performance in very little space. But the labs are also building a supercomputer entirely from AMD, called Frontier, to be released later this year. OEM partners that ship instinct products include ASUS, ATOS, Dell Technologies, Gigabyte, Hewlett Packard Enterprise (HPE), Lenovo, Penguin Computing and Supermicro.
Intel has long struggled to do more than basic integrated graphics processors for its desktop processors. For desktops, it has its new Intel Xe line, while the server equivalent is known as the Intel Server GPU.
Now the company says it will enter the data center GPU field this year with a processor codenamed Ponte Vecchio, which reportedly provides 45 TFLOPS on the FP64 – almost the same as AMD’s MI250 and 25% behind Nvidia’s Hopper.
“It’s really going to hurt the environment,” Paddy said. “From what we’ve been told – and we’ve heard rumors and other leaks – it’s very scalable. Ponte Vecchio is due out later this year.
Newman has also heard positive things about Ponte Vecchio, but said that the real opportunity for Intel is with its oneAPI software strategy.
oneAPI is a unifying software development platform that the company is working on, which is designed to choose the most appropriate type of silicon that Intel produces – x86, GPU, FPGA, AI processors – when compiling applications, instead of forcing the developer to choose one silicon type and code to it. It also provides a number of API libraries for features such as video processing, communications, analysis, and neural networks.
This abstraction eliminates the need to determine the best targeting processor, as well as the need to work with different tools, libraries and programming languages. So instead of coding to a specific processor in a specific language, developers can focus on business logic and write in Data Parallel C ++ (DPC ++), an open source version of C ++ designed specifically for data concurrency and heterogeneous programming.
One factor that separates Intel from Nvidia and AMD is where it makes its chips. While others use Taiwanese chipmaker TSMC, Intel makes many of its own chips in the United States, with other factories in Ireland, Malaysia and Israel. And there are big plans to build more in the United States. That gives him certain benefits, Castle said. “Control [it has] his own production gives him control over his destiny in a certain way, “he said. “I see these things as assets for the company.”
Ultimately, Newman said, competition between Nvidia, AMD and Intel could boil down to a software race. “If you ask [Nvidia’s] top engineers, they will say we are not a chip company. We are a software company. I really believe that Intel has not thought of an AI software company so far, but if they succeed [oneAPI] That’s right, I see a real opportunity there, “he said.
Copyright © 2022 IDG Communications, Inc.