In addition to rapidly advancing 10nm, 7nm processes, Ice Lake processors, Xe GPU graphics cards and computing cards in the consumer space, Intel Xeon Xeon server will also accelerate the pace, 10nm is not far away.
Intel disclosed at a recent investor conference,The renewal cycle of the Xeon product line will be shortened from the current 5-7 quarters (15-21 months) to the next 4-5 quarters (12-15 months).
Intel just released the second generation of scalable Xeon a month ago, codenamed Cascade Lake-X, largeAbout a year later, at the beginning of the second quarter of 2020, Intel's first 10nm process, the Xeon Ice Lake, will be available, bringing new architecture and stronger single-core performance, and the number of cores is expected to be more.
But before 10nm Ice Lake,There will be the last generation of 14nm Xeon, codename Copper LakeThe architecture continues, the core number should not continue to increase (single chip 28 core / dual chip 56 core), the focus is on technology enhancements, such asExclusive support AVX512-BF16 (bfloat16) instruction setCan accelerate machine learning and near sensor computing applications.
Intel said,Cooper Lake and Ice Lake have been sampled.
Curiously, the release time of Cooper Lake is not clear, and Cooper Lake and Ice Lake have always been placed on the same platform. Does this mean that they will abandon the current LGA3647 and replace the new interface?
In 2021, Intel will introduce the new generation Xeon, codenamed "Sapphire Rapids", which should be based on the enhanced 10nm++ process. In 2022, it will be the next generation with undetermined codename. It is expected to use the 7nm process for the first time.
The Intel 7nm process will be available in 2021, with the first product being a Xe GPU compute card, and the processor's first release is not yet announced.
As the company's transformation is data-centric, the Xeon series will become more and more important in the Intel product line, along with Xe GPUs, FPGAs and various dedicated accelerators to form a complete AI computing platform.
The first 7nm EUV process is a GPU rather than an expected CPU. Why?
Intel explained that GPUs are less risky and easier to manufacture when using new processes.
10nm will go into battle this year, and 10nm+, 10nm++ will appear in succession in the next year; 7nm will appear in 2021, and in 2022 and 2023, 7nm+ and 7nm++ will be continuously introduced.
So, what is Intel's first 7nm process product in 2021? It's not a CPU processor, it's a GPU graphics card, specifically a GPGPU general-purpose computing accelerator card based on the Xe architecture, EMIB 2D integrated package and Foveros 3D hybrid package, for data center AI and high performance computing.
As a CPU-based company, why did Intel launch GPUs instead of CPUs in a new generation of processes? This question must be very curious. Intel senior vice president Venkata Renduchintala explained their choice afterwards. The reason is that when using a new generation of processes, especially EUV lithography, the first GPU is because of lower risk and more defects. Easy to solve, even if you sacrifice a bit of performance.
From the point of view of Intel, they also learned the lesson of the 10nm process, 7nm EUV is a new generation of technology, will not take the new process to produce more difficult CPU, but choose GPU products to test water, control new The risk of mass production in order to gain experience and lessons for the CPU to launch the 7nm EUV process.
In addition, in terms of verification, Intel also has actions.
The second generation of 10nm process processor Tiger Lake, launched in 2020, will be the successor to this year's 10nm process Ice Lake processor. Tiger Lake's CPU core is not expected to change much, but the GPU architecture will be upgraded to the latest Xe architecture, and Intel mentions that its performance is four times that of existing processors because Tiger Lake's GPU integrates 96 EU units compared to Ice Lake. Increased by 50%, it is 3 times larger than the Core 8/9 of the current 24 EU units.
From the information released yesterday, the Tigger Lake processor will not only make innovations in display technology and I/O technology, but also the GPU architecture will undergo major changes. With the new X graphics engine, the guess should be with Xe. Independently architecture a system.
How strong is Tiger Lake's GPU performance? Intel officials also announced some details yesterday, such as 3 times wireless network speed, 4 times graphics performance, up to 3 times AI performance, 2 times productivity, 4 times encoding performance, the increase is still very scary.
WiFi, AI performance, these are good, how is the graphics and encoding performance 4 times? In fact, Intel has detailed explanations in the PPT comments, such as GPU performance:
"Approx. 4x Graphics Performance: Estimated by Intel as of April 2019, based on the 3DMark 11 and Firestrike scores of TGL U42 96EU 15W as compared to WHL U42 24EU 15W.”
This 4x GPU performance comparison is a 15W processor, the codename WHL (Whisky Lake) processor is 24 EU units, and the same 15W TDP power TGL (Tiger Lake) processor is 96 EU units, just the right size It is 4 times the former.
Even if you don't consider Tiger Lake's GPU architecture is a new generation of Xe architecture, the scale is enough to crush the current processor, 96 EU units are already four times the current level, compared to the Ice Lake processing of 10nm process Up to 64 EU units have also been increased by 50%, and if it is the 48 EU units of the mainstream model, the increase from Ice Lake to Tiger Lake is also doubled.
According to this performance level, the existing 24 EU unit Intel processors play some high-end game frames or single digits within 10, but 4 times the performance of the number of frames is very likely, originally run like a slide The game can also reach a frame rate of 30-40fps, which means that the basic game experience can be guaranteed, and the nuclear display is no longer waste.