Crunching numbers while staying energy efficient
The increasing importance of computer efficiency can be quantified by the amount of computation performed per watt of power converted to heat. The Green500 List serves as a benchmark for this metric.
The EdgeCluster uses processors optimized for mobile devices, which are designed with thermal efficiency in mind due to their compact size and limited cooling capabilities. Despite this constraint, these processors typically deliver impressive processing performance, making them well-suited for applications requiring fast computation and low power consumption.
AMD Ryzen 8840U is completed by an integrated GPU and a NPU unit making it a versatile yet powerfull computing plateform with a TDP of only 28W.
While not being the fastest CPU currently available, it is nonetheless one of the most efficient currently available for x64 architecture.
Single thread performances
Recent server CPU raised the number of available cores to insane amount. Latest CPU generation pushed this number above 100 cores. To allow such high number of core inside the same die, tradeoff must has been made to keep the TDP at an acceptable level. This is usually done by reducing the max frequency impacting the single thread performances or by using large cooling solution to extract the heat from the die.
Also, running all the cores at their maximum single thread performances on a server class CPU usually trigger thermal throttling. Staying in the 'turbo' frequency area with every core is simply challenging.
Not every workload can be parallelized, thus, single thread performance is still a key metric for crunching numbers.
| CPU | Single Thread Performances1 | Est. Price2 | # of core | Price per core |
|---|---|---|---|---|
| AMD EPYC 9124 | 27291 | ~US$ 1,0003 | 16 | ~US$ 62 |
| AMD Ryzen 5800X | 34451 | ~US$ 2003 | 8 | ~US$ 25 |
| AMD Ryzen 8840U | 35661 | ~US$ 5504 | 8 | ~US$ 69 |
2 As of november 2024
3 Does not include motherboard
4 Include motherboard
Multi Threading Performances
Comparing the multithreading performances between a 16 cores CPU and a 8 cores CPU isn't really fair as the 16 cores CPU has the advantage of the Infinity Fabric that manage data exchange between each cores.
Still, we can compare them by dividing the result of a benchmark by the number of core. That will help us figure out how much work a core is doing by itself.
| CPU | Multi Thread Performances1 | # of core | Score per core |
|---|---|---|---|
| EPYC 9124 | 43963 | 16 | 2747 |
| AMD Ryzen 5800X | 27841 | 8 | 3480 |
| AMD Ryzen 8840U | 23892 | 8 | 2986 |
We can see that the 8840U has a higher score per core in multithreading than EPYC class CPU. This is might be explained by the memory bandwidth available per core.
This also hints us that if the intented workload can be distributed over many computing nodes, the 8840u will perform substantially better than an EPYC class CPU in such case.
Spreading the heat
By using smaller CPU on multiple small boards, heat is spread more evenly allowing an higher average TDP. EdgeCluster embbeds 4 small motherboards packing all the needed features to run your workload.
| CPU | TDP | # of cores | Power budget per core |
|---|---|---|---|
| EPYC 9124 | 200W | 16 | 12.5W |
| AMD Ryzen 5800X | 105W | 8 | 13.5W |
| AMD Ryzen 8840U | 28W | 8 | 4.4W (3.5W)1 |
The EdgeCluster enclosure is equiped with multiple 40mm fans to help with cooling even if each 8840U board comes with its own cooling fan. Moving the heat out of the enclosure is important to avoid any thermal throttling and let you grab all the available computing power.
Power Delivery
The EdgeCluster features 4 nodes each rated at 120W maximum power consumption. The installed power supply is rated for 500W of sustained power delivery allowing it to work without sweating.
Extremly Compact Enclosure
The EdgeCluster is a 1U 19" enclosure with 255mm depth that can easily fit half depth 19" cabinet. Those cabinets are usually used for edge network appliances.
With only 500W of power, the EdgeCluster can sit next to your network switch in your office.
To figure out how big is EdgeCluster, it matches the size of an ordinary 19" 1U network switch.
Network Bandwidth
Each EdgeCluster node is equiped with two ethernet interfaces: 1GbE and 2.5GbE. These two ports can be used simultaneously for transfering data from/to each node.
Additionnaly to the accessible ethernet ports, EdgeCluster features an extremly fast private LAN running over USB4NET allowing 10GbE network class bandwith between nodes. This private LAN is arranged as a ring network and require some routing configuration on the network stack to be exploited.
More processing power under the hood
AMD Ryzen 8840U is a technological marvel. It also features an embedded GPU allowing the acceleration of certain class of computation. Even if this iGPU isn't the most performant of the market, it can still provide some more punch for crunching numbers. All this without breaking the power budget!
| GPU | PassMark - Direct compute1 | TDP |
|---|---|---|
| NVidia RTX40902 | 27999 | 450W |
| NVidia GTX 1650 (Mobile)2 | 4115 | 50W |
| NVidia RTX 2060 (Mobile)2 | 3989 | 80W |
| 780M3 | 3958 | 15W4 |
2 Requires additional hardware
3 Free, embbeded inside the CPU!
4 This is the TDP of the GPU section of the APU. It adds up with the CPU TDP of 28W.
As we can see, the 8840U is still a very capable GPU matching dedicated silicon.
NPU, the new kid on the block
With the current AI trend, CPU manufacturer are all in with adding dedicated silicon for crunching neural network matrix.
The 8840U is no exception and features a small yet powerfull NPU processor rated at 16TOPS (INT8).
Interested? Want to know more?
Drop your email in the form below. We will get back to you as soon as possible.