Anyone working in IT knows that many technologies improve incrementally.
With product announcements, it’s not unusual to see:
- a 10% performance gain
- a 20% efficiency improvement
- a 17% reduction in TCO.
But for many years, processors bucked that trend.
Gordon Moore, co-founder of Intel®, in 1965, famously said that integrated circuitry density doubled every two years, with a consequent increase in processor performance. And though Moore’s Law seems dead (though some would disagree), as density improvements have NOT doubled every two years in recent memory, the x86 processor manufacturers, Intel® and AMD, continue to try and drive toward more-than-incremental performance and efficiency gains with every generation.
Back in November, AMD released the fourth generation of their 4th gen AMD EPYC™ 9004 Series processors. The new processor family codenamed “Genoa,” is the fastest ever from AMD. It integrated a mix of revolutionary design improvements, including up to 96 cores per socket, DDR5 memory, and PCIe® Gen 5, which are critical for AI and machine learning applications. It also supports CXL® 1.1+ for memory expansion, helping customers to meet demands for larger in-memory workload capacity.
This isn’t just another incremental performance gain. 4th Gen AMD EPYC™ processors offer up to 2.8x more performance with unprecedented performance per watt, and industry experts have concluded that 4th Gen AMD EPYC™ processors will transform what cloud and edge providers are capable of — and what they’ll have to do to remain competitive.
Here at Hypertec, as we stare into our crystal ball, we can see four ways Genoa will force technologists to rethink edge and cloud.
#1: 4th Gen AMD EPYC will drive massive uplift in cloud and edge speed and scale
Researchers and analysts have demonstrated, across a truly diverse range of workloads, that Genoa-powered servers outperform any other server platform ever made.
- For cloud workloads, 107% higher performance
- For HPC workloads, 123% higher performance
- For enterprise workloads, 94% higher performance.
- Source: Amd.com
Whether in the cloud or the edge, organizations that choose Genoa-powered servers will see substantial improvements in overall performance. Workloads that once took minutes of precious processing time, like in-memory analytics, will now take seconds. In addition, Genoa will support larger-scale workloads (larger datasets) than any other server platform, thanks to increases in memory size and overall bandwidth.
These capabilities make Genoa wonderful from the perspective of responsiveness and scale, but they also have implications for innovation.
Suddenly, organizations will be able to architect completely new applications which, on previous servers, would have been processor, memory, or throughput bound. At the edge especially, new applications will emerge to take advantage of Genoa-powered performance. Services like augmented reality become much closer to reality due to the power of this platform.
#2: 4th Gen AMD EPYC drives massive server consolidation opportunity
Of course, there are many ways to leverage performance. Thanks to the radical improvements in performance across all manner of workloads, organizations that choose Genoa-powered servers will be able to rethink the number of physical servers they need to do the job.
Let’s take a look at a set of simple VMmark 3.1.1 server virtualization benchmarks that AMD uses to present the opportunity (they can be found here, under footnote 8), comparing a Genoa-based server with a similar two-socket Intel®-based server and a previous-generation AMD EPYC™-based server.
|Server Configuration||Tiles||Virtual Machines|
|2-node, 2P 40-core 3rd Gen Intel® Xeon® Platinum 8380, codenamed Ice Lake, released April 2021||14 tiles||266 VMs|
|2-node, 2P 64-core 3rd Gen AMD EPYC™ 7763-powered server, launched in 2019||24 tiles||456 VMs|
|2-node, 2P 96-core 4th Gen AMD EPYC™ 9654-powered server||44 tiles||836 VMs|
To put it simply, the latest AMD EPYC™-powered server can support 3.14x more virtual machines than a two-year-old Intel® server.
What does this mean for organizations?
AMD shows us, that on a particular workload, 15 last-generation Intel® servers can be consolidated onto 5 Genoa-powered servers. That’s a 3-for-1 consolidation, consistent with the VMmark numbers we saw. For organizations seeking greater operational efficiency at all levels, from power and cooling to server-level monitoring, maintenance, repair, and update, 4th Gen AMD EPYC™ servers are a game-changer.
#3: Cutting required data center rack capacity
Imagine for a moment that a cloud provider has a Tier 2 data center, built on a modular design using five containerized modules. Each module contains 200 three-year-old, 2U, dual-socket servers, spread across ten racks, giving the organization a total of 1000 servers.
Now imagine ripping and replacing those servers with Genoa. It would be possible to support the same number of workloads on 200 servers, in ten racks — in a single container.
This shift has massive implications
- Suddenly data center operators and cloud providers have room to grow.
- They suddenly cut their power, cooling, and monitoring costs, improving profitability.
But if you’re a colocation provider, this cut in required rack capacity might not be a good thing. If you’re charging your customer per unit of rack space, and they’ve gone from needing ten racks to two, your cost assumptions, pricing models, and revenue forecasts start going out the window.
This shift will, over time, also have massive implications for budgeting and procurement. Instead of purchasing five modular containers, with requisite power and cooling, a company might only purchase one. Instead of hiring enough staff to monitor five containers and 1000 servers, they only need resources to monitor and manage a single modular container.
#4: Data centers and edge compute nodes will have to be redesigned
Have to be redesigned?
Yes. To put it simply, these 4th Gen AMD EPYC™-based servers will force a radical rethink among data center designers.
The good news is that, in many respects, data center operators and designers will be empowered to do more with less. We’ve already seen that organizations could cut their rack requirements by 50% or more while supporting the same number of workloads, with greater performance.
This shift opens all sorts of avenues for rethinking data center and edge design
- Data center planners can suddenly consider smaller sites. Instead of 50,000 square feet for a new data center, an organization might only need 20,000 square feet.
- Edge workloads can be supported in tiny, two or three-server enclosures instead of half-rack enclosures.
- Compute can be added to a vastly diverse set of locations than ever before.
- Compute capacity can be tripled at a given site simply by swapping out old servers for new ones.
However, these changes will require another shift in edge and cloud data center design.
In real-world testing, researchers see 1kW-1.2kW power consumption while testing dual-socket 4thGen AMD EPYC™ servers. That’s a substantial increase from the 800W-1kW power envelopes seen in the previous generation of servers.
On the one hand, thanks to the substantial performance per watt gains we’ve seen, if organizations plan to use Genoa-powered servers to consolidate their server estates, they’re likely going to be in luck — shifting from 15 1000W servers to 5 1.2kW servers is a power consumption win.
Those racks won’t stay empty forever. And a rack that’s fully populated with 1U, 2-socket Genoa-powered servers will consume more than 50kW at peak load.
Most data centers and edge locations cannot sustain that power consumption level, and they certainly cannot sustain the air cooling requirements needed, especially since power density increases air cooling requirements in a non-linear fashion. Old CRACs will simply not be able to cool a data center filled with these servers.
How will you cope with this shift?
In the short term, we’ll see organizations spacing out Genoa-powered servers, intermingling them with cooler servers, or simply not fully populating racks
However, particularly for new builds at the edge or in the cloud, organizations will be evaluating ways to move past the limits of air cooling, especially since the power of Genoa will offer them substantial operational and competitive advantages. Our expectation is that, from 2023-2025, we will see organizations start to come to terms with the power and cooling demands of this latest generation of servers and begin to think past air-cooled data center designs.
To put it simply, Genoa will drive a shift toward immersion cooling — the only cost-effective way to scale racks full of Genoa-powered servers. Immersion cooling is proven to
- Support a 10x increase in server density
- Cut build costs by 50%
- Reduce cooling OPEX by 95%
- Support cooling densities of 50kW per rack in production environments.
So, to conclude…
The 4th Gen AMD EPYC™ -powered servers will force organizations to rethink how they’re supporting initiatives in the cloud and at the edge. These platforms offer unprecedented potential for cost-savings, operational simplicity, performance gains, and support for innovation, but will require designers and developers to evaluate, test, and implement liquid cooling to ensure optimal performance and value.
Learn more about how Hypertec supports this need for immersion cooling