James Pearce, TelecomTV (00:02):
Performance and efficiency are key considerations for telecoms operators looking to cash in on the opportunities offered by 5G and AI. But with developing technologies thirstier than ever, how can telcos make sure that energy doesn't become a key constraint on their plans? Well, with me to talk about this and much, much more is Piertro Wagliki, who is the global telco lead for compute and enterprise AI at AMD. Welcome. Let's start by maybe talking about how the telecom industry's definition of performance is evolving in the AI era.

Piotr Weglicki, HPE (00:41):
So I think that telco industry has moved beyond thinking about performance just in the peak throughput. And right now the view on performance is much more nuanced and is looking at how efficiently and how consistently the data is being transferred in the networks. And there are three aspects of this thinking. So the first one is the performance per watt and that comes from the fact that operators need to scale their capacity and throughput without scaling the energy consumption at the same rate. So while the capacity and throughput is increasing, it's inevitable that energy consumption will have to as well. But the key thing is really that the curve of this increase for the energy consumption has to be much flatter than the other one. The second aspect of that is bigger focus on deterministic and low latency performance. And the terminus has always been a cornerstone for the telecom, but low latency is a must in so many applications right now, especially in ultra reliable, low latency communication and at the edge where the consistency is absolutely key.

(02:09)
And the third aspect is the workload efficiency in the virtualized environment. So in other words, how many subscribers and how many sessions could be supported by a single server or by a single rack? And here AMD is bringing a lot of value in this area by integrating more compute and increasing the compute density in a CPU while offering that in a very attractive power envelope. We enable the operators to consolidate the workloads, improve the utilization, and overall get more capacity for every watch that they spend. And then just to sum up, so right now the performance is not as much about how fast, but also how efficiently, reliably, and in a sustainable way the performance is delivered at scale.

James Pearce, TelecomTV (03:11):
That's really, really interesting. Maybe we could talk a bit about what architectural innovations are driving better efficiency in modern telco infrastructure.

Piotr Weglicki, HPE (03:21):
So there are a lot of new innovations that are fundamentally reshaping the efficiency in the telco infrastructure. So the first one is the cloud native and disaggregated architectures. So by virtualizing the network functions and decoupling hardware from software, the operators can consolidate the workloads into fewers but more powerful processor and servers. And that improves the utilization, reduces the idle capacity and lowers both the power consumption but also the footprint. So these are the benefits both in OpEx and in CapEx. Then the second aspect that is really hand in hand is compute pooling and resource sharing that allow the compute capacity to be shared across multiple functions and network functions on the same platform. So what it means that operators instead of over provisioning for each network element for the peak load, they can dynamically allocate the resources in a way that they are needed and how they're needed and when they're needed.

(04:43)
So again, performance per watt is really improving through that. The third aspect of that is the platform architectures. And here it comes both the silicon improvements and also in the platform improvements where we are working together with the OEMs, as the industry, we're bringing more improved memory and IO bandwidth where the tighter connection with Nix is enabling a better performance and lower latency. But Silicon, I want to give a shout out to our industry, to semiconductor industry. We've made a tremendous progress there. So just to give you an example with AMD Epic processors, comparing first generation of Epics, which we introduced in 2017 to the fifth generation Epics, we have improved the performance by a factor of 11 times. So this is a massive, massive improvement. The same time we enable the power efficiency between these generations by a factor of four. So you are getting four times more performance for the same power that it was enabled only a few years ago back in 2017.

(06:08)
So this is a massive, massive improvement there. And the final aspect of innovation is the software optimization. And the fact is putting the more intelligent workloads placement, more topology workloads and numer awareness, these aspects are definitely improving the overall performance. When you add the power management and automation, that really allows the operators to improve the efficiency of their networks in a dramatic way and reduce the power consumption during off peak periods without compromising on the performance for the busy hour performance. So in other words, there is no a single innovation that is driving that, but there is a number of factors, including cloud native design, pull compute and balance platform architectures with the software optimization that's really bring much more capacity, flexibility and better economics to the networks.

James Pearce, TelecomTV (07:25):
Of course, the big topic in the industry at the moment is AI. So maybe you could talk about what impact AI and also edge computing have on energy consumption.

Piotr Weglicki, HPE (07:36):
So this is a great question and there is no hiding that AI with its compute intensity is going to consume more energy, especially when you're taking and training large models or running the inference real-time inference, especially at scale, you will see significant increase in terms of energy consumption. The same thing with edge compute, bringing performance from the data center into the edge is definitely going to increase overall power that it will be required in the network. However, there are a lot of benefits to do that. So for instance, on the edge side, by moving the performance to the edge, will enable processing data much closer to the user and avoid potentially vast volumes of data to be transferred over the network back and forth. And overall, that will reduce the overall power in the network. On AI, on the other hand, the benefits that we are seeing that AI is going to bring to the network will be along the lines of optimizing the radio parameters, scheduling the resources much more intelligently, predict the traffic patterns and being able to scale dynamically the network functions and the resources really to allow operators to reduce the power and energy consumption, especially during off-peak hours without compromising the performance during the busy hours.

(09:29)
Now, the key aspect in that is the platform and there is the infrastructure that we bring. And just to give you an example with AMD, we offer the entire portfolio of technologies that address AI. We bring leading edge CPUs and state-of-the-art GPUs as well as embedded AI, which really allows us to have a very nuanced conversation with the operators to look at what is exactly the problem that they are trying to solve and look for the best technology that can realize that. And it could be a combination of different technologies, of course, as well. So just to sum up, we will definitely see the increase in compute intensity and energy consumption in specific network nodes, but overall the net effect will be that networks will benefit from energy efficiency overall.

James Pearce, TelecomTV (10:35):
That's really, really interesting. Now, if we look ahead to 6G, will energy efficiency become the primary constraint for operators?

Piotr Weglicki, HPE (10:44):
So that's a great question. And you could argue that energy was always, to a certain extent, a constraint in terms of the technology that we are able to deploy in the network. And with each generation, we bring more, more capacity, higher data rates. We bring ultra low latency native AI sensing now and always new spectrum bands. So the computational complexity is increasing and this complexity is increasing much, much faster than the power budgets that are allowed actually by the operators in the network. So today we need to look at how to solve this problem and the 6G success will not be determined by just the peak performance alone. The conversation and the solutions will have to be much more along the lines of performance per watt and the gigabits per watt, for instance, as well as a key consideration there. And this challenge will be not just at the hardware level, but it will be overall system level.

(12:05)
And we mentioned already a number of different technologies like cloud native, AI-driven automation and intelligent workload placing that is really going to help with that. What we believe that in 6G, the energy efficiency is not going to slow down the innovation, but it will shape to a large extent and the most successful 6G architectures will be delivering much more value per watt.

James Pearce, TelecomTV (12:39):
That's really, really interesting. Clearly energy efficiency will be one of the biggest challenges as operators move into the AI and 6G era, but any challenge is also an opportunity to innovate. Thank you so much for joining us today.

Piotr Weglicki, HPE (12:55):
Thank you, James. That was a great conversation.