Author: Q above text
Recently, AMD officially released the second generation EPYC 霄龙骁龙 processor, which provides a big meal for the data center market. The specifications are far ahead and the ecological construction is also booming. So what is the actual performance?
AnandTech had the privilege of doing some actual measurements and coming together to see.
The specifications of the second generation Xiaolong are undoubtedly quite dazzling and there are no rivals: 7nm new technology, Zen 2 new architecture, Chiplet chiplet design, up to 64 core 128 threads, maximum 256MB L3 cache, first support PCIe 4.0 and 128 Channel, single channel up to 4TB DDR4-3200 memory, 18GT/s high speed low latency second generation Infinity Fabric interconnect bus, SME Secure Memory Encryption, SEV Secure Encryption Virtualization … …
What AnandTech got wasThe flagship model Snapdragon 7742, 64 core 128 threads, the reference frequency 2.25GHz, the highest acceleration of 3.40GHz, the third-level cache 256MB, the thermal design power consumption 225W, the price is 6950 US dollars.
The top level of Intel is currentlyXeon Platinum 8280/8280M, 14nm process, 28 core 56 threads, reference frequency 2.7GHz, maximum acceleration 4.0GHz, 3rd level cache 38.5MB, thermal design power consumption 205W, price 10009/13012 USD.
In fact, Intel also has 56 core 112 thread Xeon Platinum 9282, but in the LGA package, integrated on the motherboard, thermal design power consumption of up to 400W, the price is said to be $ 50,000 or 60,000.
Just before the release of the second generation of Xiaolong, Intel just announced a new core package of 56 core 118 threads, but it has not been officially released, and the specific specifications have not been announced, and even if the core quantity is still inferior, how can the price be two? Three thousand dollars.
The second generation Xiaolong double product line
The second generation of Xiaolong VS second generation can be extended to strong
Second generation Xiaolong single product line
In the actual test, two Snapdragon 7742 were used to form a total of 128 core 256 threads. The motherboard is a reference design board. The memory is Micron's DDR4-3200 512GB (32GB× 16). The hard disk boot disk is Samsung MZ7LM240, data. The disk is Micron 9300 3.84TB, power supply 1200W.
AlsoThe first generation of Xiaolong flagship 7601, 32 core 64 threads, the highest frequency 2.2-3.2GHz, three-level cache 64MB, thermal design power consumption 180W, also with 16 32GB memory.
One of the Intel’s participation in the war is what I just said.Xeon Platinum 8280Another is the first generation scalableXeon Platinum 8176It is also a 28 core 56 thread with a frequency of 2.1-3.8GHz, a third level cache of 38.5MB, a thermal design power of 165W and a price of $8,725.
Sometimes will joinXeon E5-2699 v4, Broadwell four generations of Core with architecture, 14nm process, 22 core 44 threads, frequency 2.2-3.6GHz, L3 cache 55MB, thermal design power consumption 145W, price 4115 US dollars.
Since the server test items are more professional and complicated, we don't do too much expansion here to see the difference between the two.
In addition, the memory delay problem of Xiaolong is complicated because it will be introduced separately.
In the single-threaded test of SPEC CPU2006, the individual cores of Snapdragon 7742 and 7601 can be accelerated up to the specified 3.4GHz and 3.2GHz respectively, and the Xeon 8176 can reach 3.8GHz, but the data of Xeon 8280 cannot be obtained. If it can reach 4GHz The performance can be about 3-5% higher than the strong 8176.
The single-threaded performance of the Xiaolong 7742 is significantly higher than that of the previous generation Xiaolong 7601 in almost every project, up to 36%, and an average of 18%.If one is excluded and one is reversed by 1%, the average increase of other projects is 22%.
Xiaolong 7742 Contrast Xeon 8176 is high and low, the best lead is 28%, the worst is 39% behind, and the average is 7% behind. If the contrast is strong, 8280 may be about 10% behind.
Also note that the version of the GCC compiler is very important. The newer the better, the performance of GCC 8.3 is slightly higher than that of GCC 7.4 Snapdragon 7742, and 456.hmmer even doubled.
In the multi-threaded test of SPEC CPU2006, the Snapdragon 7742 doubled the frequency and the frequency was higher than that of the Snapdragon 7601. The leading range was up to 153% of the horror, and the average was 109%, more than doubled.
霄龙 7742 Contrast Xeon 8176 is the general presence of rolling, leading up to 188%, leading at least 36%, with an average of 121%!Even the Xeon 8280 can lead more than 110%.
In some tests such as libquantum, Snapdragon 7742 can run to 3.2GHz for all core threads, while 2.5GHz for other tests such as h264ref.
In the 7-Zip compression test, Xiaolong 7742 leads Xiaolong 7601 78%, leading Xeon 817654%, the decompression test leads 1.27 times respectively.1. 51 times.
In the Java Max-jOPS test, Xiaolong 7742 leads Xiaolong 7601 60%, leading Xeon 828038-48%.
If there are four Java virtual machines per node, the official data provided by the Opteron 7742 can lead the Lenovo system to measure the Xeon 8280 73%.
In the Java Critical-jOPS test, the big pages (huge pages), the dragon 7742 lead Xeon 817633%Small pages are terrible2. 57 times.
According to two official data, if the virtual machine is configured with more memory, the Snapdragon 7742 can lead.66%.
NAMD high-performance computing test, Xeon 8280 finally pulled back a game, but even if you open the AVX-512 command set, you can only lead Snapdragon 7742 2%, otherwise Snapdragon 7742 can lead43%The comparison of the Snapdragon 7601 increased by 71%.
Although this test is not comprehensive because of the time, especially the lack of the highest load test, butObviously, we can see that the second-generation Xiaolong's powerful advantage is 50-100% higher than the opponent's performance, and the price is 40% lower. No matter the performance, cost performance and energy consumption ratio, it is ruthlessly crushed.
What's more, the second-generation Snapdragon has newer processes, more cores, more memory channels and capacity, more PCIe lanes and the first PCIe 4.0.
It's no wonder that many hardware and software giants are rushing to AMD, and it's no wonder that AMD has set a goal of achieving a double-digit data center market share (currently 3.4%).
AnandTech is also full of praise for the second generation of dragons, and believes that AMD has achieved a wonderful stellar achievement, it is worth applauding.
AMD will follow the Zen 3 and Zen 4 architectures on schedule, and Intel will launch a new Xeon Ice Lake new technology new architecture next year. It claims that IPC is up 18% and supports eight channels of memory. The number of cores is almost certainly more than 56, but I don't know if the chiplet chiplet design will be used, and how will the frequency and power consumption be.