In the summer of 2017, AMD announced its first generation of EPYC server microprocessors based on the Zen microarchitecture. And although this did not bring tremendous success in terms of sales (Intel was still out of reach), it definitely became a new round of development for the company. Even then, the existing solutions were excellent in almost everything: multi-core and multithreading, automatic overclocking, fast memory, extended capabilities for external peripherals, and much more.
The second generation of Epics was not just a logical continuation of the first series, but also largely solved those few problems that have hung since its announcement. The developers switched to a more advanced technical process, learned how to compose cores (models up to 64 computing units became available), offered even faster memory and an improved PCI Express bus. And after a while, the L3 cache and clock frequencies were additionally raised, updating the line under the code name Rome 7Fx2. Industrial players have finally received a completely competitive alternative for solving a wide range of tasks at an optimal price.
Safe and productive
Processors for server stations are computing locomotives that work almost round the clock and process huge amounts of information, which, among other things, can be confidential. The tasks being solved, be it working with neural networks, database analysis, machine learning or Cloud Services, require maximum performance and scalability over a wide range, while having affordable hardware services and a total cost of ownership. Therefore, the issues of architectural, technological improvements and information security are in the first place here.
AMD EPYC is not just high-performance chips with many cores, but also powerful software support that allows them to resist hardware attacks and MDS vulnerabilities. So, in comparison with the Ryzen 1XXX-2XXX desktop solutions, they are practically unaffected by the bundle of SMM Callout and Collide + Probe and Load + Reload vulnerabilities, which can gain control over UEFI firmware and access to RAM. And timely software correction of the OS and microcode will make it more likely to protect your hardware from unwanted hacker interference. In addition, we have improvements compared to the first version of the architecture. Thus, updated EPYCs are protected from all versions of Specter, no longer need to update the chipset firmware and operating system software instructions.
Drawing a parallel with the blue camp, it should be noted that many lawsuits have been filed against Intel, caused, indeed, by the problems of processor vulnerability. This forced the company's specialists to urgently look for methods to reduce risks together with the implementation of the required software corrections. Ultimately, the global problem was resolved, as evidenced by the announcement of most major cloud providers. But if in terms of information security Xeon still provide due competition, then in terms of crystal architecture and rough performance they are much worse.
Features and benefits of "Epic" processors
Manufacturers of computer equipment annually bring new hardware to the market, striving to keep up with the consumer market and provide GIKs with advanced developments. The release of the second generation of EPYC processors can rightfully be considered AMD's breakthrough in server hardware. These solutions offer impressive performance at a lower cost, without mentioning a completely different level of total ownership cost. It is not surprising that the industry welcomed the new generation of "red" chips with a bang, and such IT giants as Amazon, Microsoft provided software support for them. This allowed EPYC to be rapidly promoted to the general public, despite the conservative views of the server hardware market.
The expediency of choosing AMD solutions becomes obvious when you consider their advantages. Among the main ones that the developers were able to implement:
- New 7nm process technology and microarchitectural improvements (meanwhile, the competitor annually puts new pluses next to its outdated 14nm lithography);
- Chiplet layout, which allows to significantly reduce the cost of multi-core processors;
- A wide range of choices, up to models with 64 cores (128 threads);
- Implementation of a new PCIe bus version 4.0 with significantly increased bandwidth;
- An actively developing ecosystem and support from the industry (both hardware and software manufacturers act as partners).
And now in more detail….
Less is better
The lack of real competition played a cruel joke with Intel. The architecture of their chips has not changed, in fact, for 7 years, and every year we meet "updated Sandy Bridge", just with slightly higher clock frequencies. An increase of 5-10% in performance does not at all attract innovation, especially when you have to buy a new motherboard with the release of new chips.
But let's not talk about the sad: thanks to the implementation of advanced technical solutions, AMD was the first to introduce processors based on 7 nm process technology (we are talking about x86, of course, mobile Apple A12 and Snapdragon 855 do not count here). What does it do? In short - the ability to place more transistors per unit area. From here we get more operations (better performance) with less heating and yield of good crystals at the level of 90%.
Architectural improvements have also affected the connection to RAM. EPYCs of the second generation "out of the box" support the DDR4-3200 MHz standard (Intel runs at 2933 MHz), and enthusiasts can easily overclock it by simply changing the voltage and profile in the BIOS.
Within the framework of this paragraph, one cannot fail to note the work on one of the weakest points of the Zen microarchitecture of the 1st generation, namely, a weak indicator of operations per clock. Zen 2 got rid of this shortcoming, providing the new cores with an almost 15% improvement in computing speed at the same clock speeds. And for multi-threaded performance, typical server operations are 23% faster (don't forget about doubled cores and increased frequency). And all this fits into the declared thermal package, which reaches 250 watts for older models.
The use of a chipset layout to create processors is one of the main arguments in favor of AMD over its eternal competitor. Chiplets are individual crystals that are linked together via a fast data bus. The updated generation of the AMD Infinity microarchitecture is optimized specifically for multi-core computing, in addition to the execution units, a central chiplet with all the necessary controllers is also added, which is responsible for data input and output. The connection of the execution units to the central hub is now provided by high-speed Infinity Fabric tracks. As a result, we find that the flagship Epic 7742 has 8 chiplets, which host 64 hardware cores with the ability to process 128 threads simultaneously. All of this is peppered with an impressive 256MB of fast L3 cache.And the only weak link in the architecture is still the speed of access to RAM (latency). Moreover, Zen 2 is slightly inferior even to its predecessor in the face of the first EPYC.
A few more important advantages
The updated processors have 128 PCIe 4.0 bus lanes each. Not only the number in the name has changed, but also the real data transfer rate, which at its peak can reach a record half a terabyte per second. True, doubling the number of lines does not apply to two-socket motherboards. The Infinity Fabric bus that connects them takes half. But splitting the processor lines into 8 groups of 16 pieces provides them with simultaneous support for up to 32 solid-state drives connected via SATA or NVMe interfaces.
The introduction of the new PCIe standard is likely to go unnoticed for desktop hardware consumers, but for the owners of server stations it promises quite real advantages. For example, a bus doubles the speed of solid state drives, scaling linearly for file read and write operations. It is also an opportunity to transfer data at a higher speed over the network or to improve communication with graphic and tensor accelerators when servicing neural networks.
An important plus is the advanced encryption capabilities of RAM, in particular AES-128, and improved software virtualization. For this, the developers had to "sew" a separate 32-bit AMD Secure Processor controller into the chip to expand the cryptographic functionality. Interestingly, such impressive capabilities of the second generation EPYC are likely a consequence of work on external projects: game consoles from Microsoft and Sony. Since the customers insisted on software isolation of their products with protection against hacking with hardware encryption methods.
Each server according to needs
The architecture, process technology and buses are, of course, good, but the end user needs specifics. The main thing that distinguishes AMD server processors from each other is the difference in the number of cores, the rest is, as a consequence, from the first derivative. There is where to roam, 14 models are available for all occasions:
- 8 cores / 16 threads - EPYC 7252 (7262);
- 12 cores / 24 threads - EPYC 7272;
- 16 cores / 32 threads - EPYC 7282 (7302);
- 24 cores / 48 threads - EPYC 7352 (7402);
- 32 cores / 64 threads - EPYC 7452 (7502, 7452);
- 48 cores / 96 threads - EPYC 7552 (7642);
- 64 cores / 128 threads - EPYC 7702 (7742).
The EPYC processors from the Rome 7Fx2 line will be rightly added to the above list. In fact, this is the same architecture from which the developers were able to squeeze out an additional 500 MHz per core and added a cache. It turned out quite convincingly, according to AMD themselves, these are processors with the best performance per core on the market. True, the number of these same cores is limited to three models: with 8, 16 and 24. Immediately, we note that this kind of "pebbles" are perfect for data centers and work with databases.
AMD EPYC chips range in price from $ 450- $ 6900, depending on the model. Building systems with 64 cores will provide the required level of performance for most cloud services, and if this is not enough, we will install systems from two sockets. It's hard to imagine tasks that would cause problems with 128 cores and 256 threads running concurrently.
The rest of the parameters affecting processor performance change linearly with the increase in the number of physical cores. Each chiplet contains 32 MB of top-level cache and 4 MB of L2 cache. Uniprocessor modifications (with the letter P in the name) are not limited by anything in comparison with dual-processor brothers, they can also have from 8 to 64 cores and an upper cache up to 256 MB.
Compare and choose the best
Presenting server solutions based on the updated architecture, the company focused on direct comparisons of computing power and price / performance ratio with competitors. So, in a pair of top-end solutions EPYC 7742 from AMD and Intel's Xeon 8280L, the brainchild of the "red" camp has almost a two-fold advantage. Moreover, junior 32-cores are also not lagging behind, showing similar, and in some places better performance than the competitor's flagship chips. As a result, we get a situation in which a server with one socket on EPYC will be no worse than a top-end two-socket server on Intel. For the consumer, this means less energy consumption, savings on licenses, and hence the total cost of owning server equipment.
Another example of comparison is the "popular" 8-core Xeon Silver 4215 with a price tag of $ 800-900 against the EPYC 7282 with twice as many cores. The latter has a lower cost, but even in such layouts provides significantly better performance. And if we compare the 7Fx2 model, then the difference in the speed of operations becomes simply colossal, which once again proves the advantage of the 7nm process technology and improved architecture.
Another interesting competing niche for the consumer is solutions based on EPYC 7452 and Xeon 6226 processors. Yes, the cost here is not in AMD's favor, but the price / performance ratio is again on the EPYC side. If we consider the best option for buying a server machine for the office, then you should pay attention to the 32-core EPYC 7502. Here is a combo of high performance and a relatively affordable price. In fact, it can be called the golden mean, which will suit most consumers.
And finally
The release of AMD EPYC has seriously shaken the position of Xeons in the server processor segment. And the improved architecture of Zen 2 only strengthened the taken vector of "red" to monopolize the market. Subsequently, this forced many industrial players to seriously think about the prospects for buying chips made using the 7nm process technology. In view of the fact that there are literally no answers to Intel's solutions, AMD has chosen the optimal time for its triumphant return to the throne of the king of microprocessor technology.
The average consumer and owners of large IT infrastructures are presented with options for any budget and performance level. Moreover, in each segment EPYC looks preferable to its competitor. AMD processors offer solid hardware protection, a robust ecosystem, fast performance, savings on license purchases and total cost of ownership. The advantages will be enough for the majority, and the prospects for mastering 5-nanometer production only increase the interest in the company's products from potential consumers.
We have appreciated all the advantages of AMD EPYC processors and running epic servers . Hurry to try and you!
