“AMD is back, and it’s here to stay,” enthuses Mark Papermaster, senior vice president of technology and engineering at the US chip maker. Not surprisingly, after years in which the company has been unable to keep pace with products aimed at professionals and large corporations that demand maximum performance from their top rival, AMD has managed to launch a large portfolio aimed at professionals with its new family of Ryzen processors and EPYC for data centers.
Maintaining a great deal in the data center business is key for the next decade. Storing information is becoming cheaper, everything is saved, but now they need algorithms and a great computational power to make sense of them and serve the companies or governments. That’s where data centers come in, graphics cards with great computing power and neural networks that already have Google or Facebook underway.
Despite the company’s great victories in recent weeks (Xbox One X processor and graphics across the Mac range that has dedicated cards), AMD is clear that this year’s professional market will be its biggest asset. “We are going to double our offer before the end of the year”
Epyc is a family of processors whose objective is to offer a large number of cores , more bandwidth than the competition and an innovative protocol of channels of input and outputs in machines of double socket, the most used, and in those of a single socket , a purpose that attracts a lot to AMD and where they believe they can make a “big difference”.
Architecture, performance and strengths
The family consists of 4 ranges from the most basic having 8 physical cores to the highest performing range with 32 cores. All have the same DDR channels, 128 PCI express lines and the security features that AMD has included. The company has made a lot of emphasis that, unlike its competitors, the customer only has to pay for the performance they need.
According to Papermaster, Epyc can serve 95% of customers who demand servers for their business in the cloud. And the performance per watt and its price does not make it compete against Intel’s Broadwell, but “against what they get next,” say from AMD.
Both per-numeric and floating-point computation yields are significant, according to AMD’s meager metrics at its Austin, Texas-based headquarters Monday. Bandwidth is one of its keys, especially facing the future where it is increasingly important to feed the cores constantly and very fast to save on costs.
With 8 memory channels in each EPYC processor, a dual-socket server can have 32 DIMMS slots with 16 channels. Simplest: 4 terabytes of RAM.
These processors have been designed from scratch for almost five years, since AMD could not compete with its old architecture. They started from the architecture and core design, then intercom to make it as fast as possible and finished with a new FinFET design to make a larger dense silicon chip in components. More silicon in the same socket size.
By the way, the socket will serve, at least, for the next two generations of Zen processors, says AMD. It is one of the keys to inviting big corporations to invest with them.
Most surprising is its entry and exit management system. Of paramount importance in the new era in which we enter centralized intelligence and massive processing of data using graphics chips, which are basically machines to do accounts. They have achieved a bandwidth of 256GB / s per socket and a latency, they say, minimal with each GPU. They have also managed to get PCIe switches removed.
Finally, while the market is geared towards dual-socket solutions, AMD has put great emphasis on its range of single-socket products that they call “no compromise” since they have almost all the fundamental advantages of dual socket management of many PCIe lines.
In security, the most striking feature is its constantly volatile memory encryption system through a dedicated chip completely independent of the processor and invisible to the operating system. Since AMD point out its importance, since most “hacks” are from within and this avoids physical reading of data that have instance in RAM. Xataka asked about the impact on performance, and AMD, which could not give exact figures, indicated that it lost just over 1%. It will depend on several factors, and the company did not give specific details.
You may also like to read: The Snapdragon 835 promises in almost everything, but it has an ace well kept in the sleeve
Now, for giant companies that have hundreds of computers running, consumption is just as important as performance. Here the star measurement is computing power per watt. According to AMD studies, processors are half the consumption in normal data centers, making it a crucial element to gain energy efficiency.
AMD engineers have come up with several interesting features. For example, the processor has thousands of temperature sensors that communicate with the power management system and the processor frequency to distribute workloads between cores constantly.
The best is its dynamic voltage regulation system by core. Now the processor will not operate at the voltage that indicates the kernel with higher load and speed of work, but each core will operate at a frequency and voltage. The more frequency you need, the more voltage you need to consume.
Customers will be offered various control tools to save energy while maintaining certain requirements . A company can select a consistent performance, so it will adjust the energy savings when it can. But you can also set a maximum consumption and operate at maximum performance depending on each situation.
At the input and output hub, now that there are no PCIe switches, the processor can lower the full voltage of the parts you do not need. This way, it is possible to have 128 channels without the consumption of electricity.
Finally, AMD has hinted at the future of ARM’s future work machines: “Making low-power servers at all costs is very easy, it’s difficult to lower consumption when you want to have maximum performance.”
EPYC is the first stone of machines ready for the future in AI
Now, the future is focused on data processing and neural network training that can make sense to billions of entries. For those calculations you need graphics cards, which is the only element that continues to grow in linear or exponential performance.
So, from June, they will distribute these beasts of computing to their partners. They are graphical for servers and have passive cooling and a great calculation capacity.
AMD has said that artificial intelligence is so important that only three or four companies can not afford to take control. They also want to offer compatibility with all popular frameworks and create an open source platform. Ogi Brkic, of Radeon Technology, says that before the end of the year companies will be using AMD servers with EPYC and Radeon very compact and with computing capacities close to 100 teraflops.
Its ROCm 1.6 platform allows developers to directly access lower-level operations within the graph, something that has always been demanded to perform certain operations in less time. “It’s like going back to the old school, programming assembler,” said Gregory Stoner, director of Radeon Open Compute. “It’s fantastic”. They have several tools of direct communication between code Python or C ++ and the GPU. Some reporters asked what their offer was compared to Google or Nvidia, clear leaders in the sector. AMD said they do not offer the complete platform from software to hardware, but focus on just one point. For now, it is working.
AMD is back
It’s not just a little creative slogan. It is true. After many years stalled as a cheap alternative to Intel, it seems that AMD has achieved two major milestones with Ryzen and EPYC in the professional market and large corporations. It is key for the company, and for Intel, to have relevance in the sector, as the consumer market has turned towards portable solutions of low consumption. However, mobile devices will increasingly rely on companies that have to deal with an inabarcable amount of data. And there have to be AMD to offer an alternative to Intel, which has been dictating the pace and price for years.