Today, Intel officially released the third-generation Max Expandable Processor, code-named “Cooper Lake”, with an increase in frequency, core number, memory, and more, focusing on DLBoost’s deep learning capabilities, the only mainstream data center processor with built-in AI in the world today. Cooper Lake is mainly for the four- and eight-way market, and later this year there will be a 10nm Ice Lake-based Zhiqiang, for the single-way, two-way market, the two complement the formation of a complete third generation family.
As to why this is divided, Intel’s explanation is that more customers want AI hardening on more road systems.
Cooper Lake is still manufactured using a 14nm process with up to 28 core 56 threads (eight passes is up to 224 core448 threads), some models increase the number of cores, while the frequency is higher, the reference frequency is up to 3.1GHz, the single core turbo speed is up to 4.3GHz, the three-stage cache is up to 38.5MB (1.375MB per core), and the thermal design 150-250W power consumption.
Memory supports six-channel DDR4, a maximum frequency of 3200MHz, up to 12, and supports 16Gb of high density particles, a maximum capacity of 4.5TB per way, 36TB on eight,36TB, and of course Intel’s own Optane PMem Auton sustainable memory, with the release of the new 200 series.
Technically, some models add Intel Speed Select (SST) technology, including STT Core Power (SST-CP), SST Turbo Frequency (SST-TF), to optimize processing resources, improve workload performance, improve resource utilization, and optimize platform TCO costs, while supporting intel AVX-512 instruction set, VNNI/BFloat16 deep learning acceleration instruction set.
With a four- and eight-way system, Cooper Lake has six UPI interconnect buses built in, with a maximum transmission of 10.4GT/s.
In terms of input and output, the processor supports up to 48 PCIe 3.0 channels, and the C620A series chipsets (including C621A, C627A, C629A) also provide up to 20 PCIe 3.0, 10 USB 3.0, 14 SATA 6Gbps, and the processor and chipset interconnect channels are DMI 3.0 x4.
The Cooper Lake package interface has since been changed to the new Socket P-Plus, also known as LGA 4189, and the future Ice Lake is the new interface, not compatible with the current LGA3647 platform.
Eight-way system structure diagram
AI is undoubtedly the greatest feature of these three generations of scalable strength, and it has been moving rapidly.
The first point in 2017 supports the AVX-512 FP32 floating-point format, has the advantage of high precision, has 24 optimizations, Jinshan Yun, the United States, Nanjing University, Yinggu and other Chinese technology enterprises, institutions have been deployed.
The second generation in 2018 joined the DLBoost INT8 integer format, 44 optimizations, we also saw more domestic institutions name: Science and Technology, Shanghai Jiaotong University, JD.com, Cloud Big Data, Zhixin Original, HuiWei Hui shadow, Qingyun QingCloud, Neusoft and so on.
In the latest third generation, we have ushered in the new BFloat16 (BF16) for AI computing, deep learning and other applications, on the one hand, the appropriate reduction of accuracy (still close to FP32), improve energy efficiency, software applications do not need much adjustment, on the other hand, a significant improvement in AI performance, the official claim compared to the previous generation of FP32 training performance increased by up to 1.93 times, reasoning performance up to 1.9 times.
At present, the third generation of the support of the BF16 Tostrengthen Extensible processor has been supported by Alibaba Cloud, Ant Financial, Tencent Cloud, Haixin Kejin, Neusoft and other companies, the measured performance improvement is less than 54%, more than 97%.
The BF16 standard has also been widely adopted by Google, Microsoft, ARM and other AI companies, and Intel X86 is the first and only CPU processor to support it.
According to Intel’s official data, if the strong E7 v4 series is the benchmark, the AI training performance is 82% higher if the three generations of extensible X-Strong avX-512, and the BF16 Plus DLBoost can provide a 2.14 times improvement;
Strong ecological support has always been Intel’s invincible advantage, this time is no exception, in the official support list can be seen a series of famous, including many global technology enterprises, scientific research institutions, software giants, universities, there are many Chinese Internet and technology enterprises, institutions of higher learning, communications operators, and even can see UnionPay, Yunda Express.
Cooper Lake’s three-generation extensible prosandment is divided into three sub-series, with a total of 11 different models, all with H or HL numbers ending, with the difference that HL supports up to 4.5TB of memory per way and H’s 1.12TB.
The highest-end Xeon Platinum 8300 series has six models, namely 8380HL, 8380H, 8376HL, 8376H, 8354H, 8353H, the first four 28 core 56 Threads, the latter two 18 core 36 threads, thermal design power consumption of 150250W, memory frequency of up to 3200MHz (one per channel) or 2933MHz (two per channel), support four-way, eight-way expansion.
The Gold 6300 series has three models, namely 6348H, 6328HL, 6328H, the previous 24 core 48 threads, the last two 16 core 32 threads, thermal design power consumption is 165W, the memory frequency of up to 2933MHz, only four ways to support.
The Xeon Gold 5300 series has two models, namely 5320H, 5318H, a 20 core 40 thread, a 18 core 36 thread, thermal design power consumption of 150W, memory frequency down to 2666MHz, only four-way, the series of RAS capabilities have been streamlined.
Among them, the gold medal 6328HL, 6328H, 5320H, 5318H four support SST technology.
Intel says that compared to the old platform five years ago, which should be Xeon E7 v3?, the average performance of the three generations of Scalable Tobe5 has increased by 90 percent, and the average database performance has improved by up to 98 percent.
If the customer does not know which model to choose, Intel also made targeted recommendations according to the actual application, such as the master AI training, choose more core 8380H, 8376H, 8354H;