In the past, only once has AMD been able to beat Intel so decisively - introducing dual-core Opteron and Athlon 64 X2 in 2005. This week's launch of Epyc has made this experience even greater. In 2005, AMD's dual cores matched Intel's number of cores, outperforming Intel's clocks and cores, and were quite expensive. This time, AMD is looking for three consecutive wins because it has higher performance, more kernels and lower per-core pricing. It is arguably the company's most serious attack on Intel's high-end Xeon market.

Industry analysts have predicted that AMD'sThe serverMarket share is likely to double in the next 12 months, reaching 10% in the second quarter of 2020. Achieving a larger share in the data center market is a key goal of AMD. Higher corporate and data center market share will not only increase AMD's revenue, but also help stabilize the company's financial performance. A key weakness of AMD in the past two decades has been its dependence on low-end PC and retail sales. These two markets are often sensitive to recession. The low-end PC market also provides the lowest revenue per slot and the lowest profit. The recession has less impact on business cycle. AMD achieved its goal of large enterprise market share between 2005 and 2006, when its server market share exceeded 20%.

Fans like to focus on AMD desktop performance, but outside of games, overall PC sales are declining. The growth of narrow categories such as two-in-one is not enough to offset the overall decline in sales. Although no one expects the PC market to fail, it is clear that the recession of 2011 is not the case. AMD still makes sense in expanding its share of desktop and mobile markets, but more importantly in competing for a share in the server sector, where revenue and unit shipments have increased over the past eight years. The year 2019 may be a downturn in server sales, but the larger trend of shifting workloads to the cloud shows no sign of slowing down.

Why Rome is a Threat to Intel

In the discussion of Rome, we focused on Epyc 7742. This chart from Servethe Home shows Epyc and Xeon in more SKUs:

A pair of AMD Epyc 7742 sells for $13,900. 7502 (32C/64T, 2.5GHz base, 3.35GHz enhancement, $2600) sells for $5200. Intel Xeon Platinum 8260 sells for $4700, but four of them are among the highest scoring systems, with a total cost of $18,800. The $13900 AMD CPU gives you 1.19 times more performance than the $18800 Intel CPU. Four E7-8890v4 models sell for nearly $30,000. A pair of Platinum 8280s is $20,000 and 8676L is priced at $16,600.

However, AMD has not only price advantages, but also cost-effective advantages. Intel aggressively subdivides its product functions and charges more for them, such as price differences between Xeon 8276, 8276M and Xeon Platinum 8276L. The three CPUs are the same, except for the maximum RAM per support.

In this case, "maximum memory" includes Intel Optane. The 4.5TB RAM function assumes that 3TB Optane and 1.5TB RAM are installed. By contrast, all 7nmRome CPUs support up to 4TB of RAM. It is a standard built-in function for all CPUs, simplifying product purchasing and future planning. AMD not only provides chips at lower prices, but also adopts Intel's whole market segmentation method. When AMD is willing to sell you 4TB addressable capacity at a basic price, it will be difficult to get additional RAM support for $8,000.

One of the issues AMD and Epyc are talking about is how it can provide 2S system advantages in 1S configuration. This chart of Servethe Home illustrates these differences very well:

One of AMD's strengths in this area is that it can attack multiple Intel weaknesses at the same time. Need a lot of PCIe channels? AMD is better. Want PCIe 4.0? AMD is better. If your workload is optimally extended with the kernel, no one in each slot will sell more kernels than AMD. Intel still has some advantages - it provides a larger unified L3 cache than AMD (each AMD L3 cache is actually 16MB, and each core has 4MB chips). But these advantages will be limited to specific applications that respond to them. Intel wants suppliers to invest in building support for its Optane DC persistent memory, but it is not clear how many companies will do so. At present, the lowest prices of NAND and DRAM make it more difficult for Optane to compete in the market.

The transition to 7Nm also gives AMD an advantage in power consumption, especially when considering server decommissioning. STH reports that the single-threaded power consumption of Xeon Platinum 8180 is about 430W (wall power), while that of AMD Epyc 7742 system is about 340W. However, they note that the high number of cores in AMD's latest CPUs will enable them to eliminate 6-8 slots in Intel Xeons (60-80 cores) in 2017 in order to integrate workloads into a single AMD Epyc system. The power savings of eliminating 3-4 dual-slot servers are far greater than the difference of less than 90W between the two CPUs.

Features like DL Boost may give Intel a performance boost in terms of artificial intelligence and machine learning workloads, but the company will face a tough battle. So far, the data we have seen suggests that these factors can help Intel rivals AMD, not defeating it.

What is the cost of Xeon?

The list price we quoted for this story is the official price of 1K units released by Intel for the Xeon CPU. As we all know, they are inaccurate, at least for major OEMs. we do not knowDell,HPThe fees paid by E and other vendors for the Xeon CPU are actually paid, but we know that it is usually much lower than pricing, and pricing is usually only paid by the retail channel.

The gap between Intel's list price and actual price can explain why Threadripper doesn't have much market penetration. Although the Threadripper CPU has provided a much larger core per dollar for two years and higher performance per dollar, OEMs that share sales information (such as MindFactory) report that Threadripper and Skylake-X have very low sales. However, Intel has not shown a particular interest in cutting Core X prices. Although AMD has performed well in this game, it continues to position its 10-core Core i9-9820X ​​as a competitor to chips such as the Threadripper 2950X. This strongly suggests that Intel is not particularly troublesome when it comes to selling 10 core CPUs to OEM partners who need them, although Threadripper is very cost effective and AMD's share of the workstation market is very limited.

Although Intel has cut its HEDT price ($10 Core i7-6950X for $1,723 in 2016 and $900 for the Core i9-9820X ​​today), it has never attempted to match the Threadripper for price/performance. If that fortress will crash, Rome will be the CPU to replace it. If Epyc begins to devour the server market, Ryzen and Threadripper will be considered more trusted workstation CPUs.

AMD-led games

Intel can lower prices in the short term to deal with AMD. In the long run, it must directly challenge AMD. This means more cores are available at a lower price, and each slot supports a higher amount of memory. Built at 14nm Cooper Lake will include additional support for the new artificial intelligence-based AVX-512 command, which will be available in the first half of next year. The chip will help Intel focus on some of the markets it wants to compete with, but it won't change the difference in core numbers between the two companies. Similarly, given that AMD is willing to support up to 4TB of memory per CPU socket, Intel may have difficulty paying an additional $3,000 to $7,000 to support 2TB-4.5TB of RAM.

We still don't know if Intel will increase the number of cores in the Ice Lake server, or what design it will bring to the market, but the ICL in the server will take at least a year. When the ICL server is ready to ship, AMD's 7nm EUV design may also be ready. After opening all the update cycles with Rome, AMD's challenges in the next 12-24 months will demonstrate a smooth and continuous update pace and continuous performance improvements. If this is the case, then its dream of building a stable corporate market for decades will come true.

Don't become arrogant

When AMD launched the dual-core Opteron and its consumer product Athlon 64 X2, there was a clear feeling that the company finally ushered in its own moment. More than a year later, Intel introduced Core 2 Duo. AMD has been lingering in the well-known wilderness for the next 11 years. Later, executives admitted that the company had begun to pay attention and was distracted by ATI's acquisition. Then there is a series of problems.

The simple assumption is that P4 Prescott is a disaster that Intel can't prove. Historically, attacking Intel has often proved similar to hitting a rubber wall with a sledgehammer (pun). Deforming a wall is relatively easy, but completely destroying it is a daunting task. AMD may have the best chance to capture the market share of the enterprise through 7nm Epyc, but building server shares is a slow and cautious process. If AMD wants to keep its goals, it needs to play its part in a different way than it did from 2005 to 2006.

But then again, I will not easily use phrases like the "Golden Age." Although I have not predicted how long it will last, for me, the debut of 7nm Epyc can make me boldly say: Welcome to AMD's second golden age.