Home > News content

Once again challenged the super king, AMD really wants to kill Intel.

via:博客园     time:2019/5/8 17:32:15     readed:161

再次挑战超算王座,AMD

Note: The title map is not the protagonist of this article, but the current global supercomputer TOP 1, located at the Oak Ridge National Laboratory (ORNL) in the United States (ORNL) "peak" (Summit).

A few hours ago, AMD “low-key” announced a message —— they are about to build the latest supercomputer “Frontier” for the US Department of Energy Oak Ridge National Laboratory with Cray, a well-known supercomputer solution manufacturer. Front line), the goal is to achieve 1.5 exaFlops computing power (exaFlops = 10 billion floating point operations per second, one standard floating point operation is multiplied by two fifteen digits).

data-w=1000

Supercomputer & quot; Frontier & rdquo; (front line)

How exaggerated is this number? Take "Summit" above the supercomputer throne as an example, it can provide 0.2exaFlops computing power. In other words, "Frontier" is about 8 times stronger.

Let's talk about a comparison of images. If you can organize the top 160 computers in the current global supercomputer, their total computing power is almost the same as the target computing power of “Frontier”.

"There is a talented person in Jiangshan," which is not rare in the supercomputer world. The only thing that is magical is that this time it was not Intel, not IBM, but AMD, and it was "all-in-one", CPU, GPU, internal high-speed lines and other key equipment were all from AMD. Create it yourself.

Seeing this, you may also find this thing not so simple.

AMD suddenly took out the "Quick" & rdquo;

data-w=1000

AMD officially stated in the statement: "Frontier will adopt AMD's customized CPU and GPU, and represents AMD's long list of technical achievements in the US Department of Energy's "exascale" program. ”

This sentence is not empty talk, AMD also gives some vague details and figures:

  • “Frontier” will have 24 million times more network bandwidth than regular home network connections, or enough to download 100,000 HD movies in one second.

  • “Frontier” will cover an area of ​​7,300 square feet (approximately 680 square meters), equivalent to two basketball courts.

  • “The internal cable of Frontier” is connected to the head and tail, and the total length can reach 90 miles (approximately 145 kilometers).

Compared to these “milk tea marketing” descriptions that make ordinary people lament, the biggest highlight is the short description:

“Frontier” will use the first GPU and CPU that is “optimized” for supercomputing. Includes a custom AMD Epyc processor with future Zen core architecture for high performance computing (HPC) and artificial intelligence workloads, and a graphics processor processor (GPU) from the AMD Radeon Instinct product line. GPUs feature high-performance computing engines, "widespread" mixed-precision operations, and high-bandwidth memory, and they are connected by AMD's Infinity Fabric and Cray Slingshot high-bandwidth system interconnect architecture (one Epyc processor and four Instinct graphics cards) .

In the top 500 supercomputers in the world, AMD has a CPU share of 0.4% (two AMD Opteron processors) and 0 AMD GPUs. In contrast, the goal of "Frontier" is obviously very ambitious. Not only does AMD's most advanced CPU and GPU combine today, it is more promising to optimize the entire system with its own advantages across two areas.

Not to mention this time AMD is working with professional supercomputer solution provider Cray, which currently has 104 of the top 500 supercomputers in the world. But Cray is not just as simple as building hardware. They also have a range of software for high-performance computing programming environments, supercomputer management, and monitoring. Cray has indicated that it will target AMD's hardware adaptation software.

data-w=584

This is definitely not good news for Intel, NVIDIA, the main computing chip vendors of the current supercomputer systems. Currently, the former provides CPUs for 94.8% of supercomputers in the top 500 supercomputers worldwide, and the latter provides GPUs for 27% of supercomputers.

Supercomputer, AMD's next oil field?

data-w=1000

As we mentioned above, “Frontier” is actually the first result of the US Department of Energy's “exascale” plan. It is actually an allocation specifically designed to accelerate the development of exaFlops supercomputers in the United States. In the case of AMD alone, the contract value exceeded $600 million.

You know, AMD's annual turnover is only 6.48 billion US dollars, the entire "calculation and graphics" business unit's turnover is only 986 million US dollars, the net profit is only 514 million US dollars, 600 million US dollars really not a small number.

In addition to “Frontier”, in the earnings report last year, AMD also revealed that the new supercomputers of Lawrence Livermore National Laboratory, Stuttgart University, and the US Department of Energy will use their hardware. Obviously, “Frontier” is just the beginning, and AMD has “opened” in the field of supercomputers.

We might as well make a bold assumption. If the success of the top supercomputer throne is successful, will it make more supercomputers with updated devices directed to AMD? If 20% of the top 500 supercomputer systems in the world are eaten by AMD in the future, with a total of $100 million per system, the total value will reach $10 billion.

The positive impact of cutting-edge technology is another factor that needs to be considered but is difficult to measure. Military, supercomputer and other fields have always been the most conducive to the development of new technologies. AMD is currently optimized for super and computer based on consumer products, but AMD's new technology for top-level usage scenarios such as supercomputers has Maybe & nbspquo; back to feed rdquo; consumer market?

Su Ma’s imagination of this operation is really big.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments