On Wednesday, August 7, AMD launched the 7nm refresh of its Epyc CPU family. These new cores don’t just one-up Intel in a particular category, they deliver enormous improvements in every category. AMD has cut its per-core pricing, increased IPC, and promises to deliver far more CPU cores than an equivalent Intel socket.
There’s only been one other time that AMD came close to beating Intel so decisively — the introduction of dual-core Opteron and Athlon 64 X2 in 2005. Epyc’s launch this week feels bigger. In 2005, AMD’s dual cores matched Intel on core count, outperformed Intel clock-for-clock and core-for-core, and were quite expensive. This time, AMD is going for the trifecta, with higher performance, more cores, and lower per-core pricing. It’s the most serious assault on Intel’s high-end Xeon market that the company has ever launched.
Industry analysts have already predicted that AMD’s server market share could double within the next 12 months, hitting 10 percent by Q2 2020. Achieving larger share in the data center market is a critical goal for AMD. A higher share of the enterprise and data center market won’t just increase in AMD’s revenue, it’ll help stabilize the company’s financial performance. One of AMD’s critical weaknesses for the last two decades has been its reliance on low-end PCs and retail channel sales. Both of these markets tend to be sensitive to recessions. The low-end PC market also offers the least revenue per-socket and the smallest margins. Enterprise business cycles are less impacted by downturns. AMD briefly achieved its goal of substantial enterprise market share in 2005 – 2006, when its server market share broke 20 percent.
Enthusiasts like to focus on AMD’s desktop performance, but outside of gaming, overall PC sales are declining. Growth in narrow categories like 2-in-1’s has not been sufficient to offset the general sales decline. While no one expects the PC market to fail, it’s clear that the 2011 downturn was not a blip. It still makes sense for AMD to fight to expand its share of the desktop and mobile markets, but it makes even more sense to fight for a share of the server space, where revenue and unit shipments have both grown over the past 8 years. 2019 may be a down year for server sales but the larger trend towards moving workloads into the cloud shows no signs of slowing down.
Why Rome is a Threat to Intel
In our discussions of Rome, we’ve focused primarily on the Epyc 7742. This graph, from ServetheHome, shows Epyc versus Xeon performance across more SKUs. Take a look down the stack:
A pair of AMD Epyc 7742’s is $ 13,900. A brace of 7502’s (32C/64T, 2.5GHz base, 3.35GHz boost, $ 2600) is $ 5200. The Intel Xeon Platinum 8260 is a $ 4700 CPU, but there are four of them in the highest-scoring system, for a total cost of $ 18,800. $ 13,900 worth of AMD CPUs buys you ~1.19x more performance than $ 18,800 worth of Intel CPUs. The comparison doesn’t get better as we drop down the stack. Four E7-8890v4’s would run nearly $ 30,000 at list price. A pair of Platinum 8280s is $ 20,000. The 8676L is a $ 16,600 CPU at list price.
But it’s not just price, or even price/performance where AMD has an advantage. Intel heavily subdivides its product features and charges considerably more for them. Consider, for example, the price difference between the Xeon 8276, 8276M and Xeon Platinum 8276L. These three CPUs are identical, save for the maximum amount of RAM each supports. The pricing, however, is anything but.
In this case, “Maximum memory” includes Intel Optane. The 4.5TB RAM capability assumes 3TB of Optane installed alongside 1.5TB of RAM. For comparison, all 7nm Rome CPUs offer support for up to 4TB of RAM. It’s a standard, baked-in feature on all CPUs, and it simplifies product purchases and future planning. AMD isn’t just offering chips at lower prices, it’s taking a bat to Intel’s entire market segmentation method. Good luck justifying an $ 8000 price increase for additional RAM support when AMD is willing to sell you 4TB worth of addressable capacity at base price.
One of AMD’s talking points with Epyc is how it offers the benefits of a 2S system in a 1S configuration. This chart from ServetheHome lays out the differences nicely:
Part of AMD’s advantage here is that it can hit multiple Intel weaknesses simultaneously. Need lots of PCIe lanes? AMD is better. Want PCIe 4.0? AMD is better. If your workloads scale optimally with cores, no one is selling more cores per socket than AMD. Intel can still claim a few advantages — it offers much larger unified L3 caches than AMD (each individual AMD L3 cache is effectively 16MB, with a 4MB slice per core). But those advantages are going to be limited to specific applications that respond to them. Intel wants vendors to invest in building support for its Optane DC Persistent Memory, but it isn’t clear how many are doing so. The current rock-bottom prices for both NAND and DRAM have made it much harder for Optane to compete in-market.
The move to 7nm has given AMD an advantage in power consumption as well, particularly when you consider server retirements. STH reports single-threaded power consumption on a Xeon Platinum 8180 at ~430W (wall power), compared to ~340W of wall power for the AMD Epyc 7742 system. What they note, however, is that the high core count on AMD’s newest CPUs will allow them to retire between 6-8 sockets worth of 2017 Intel Xeons (60-80 cores) in order to consolidate the workloads into a single AMD Epyc system. The power savings from retiring 3-4 dual-socket servers is much larger than the ~90W difference between the two CPUs.
Features like DL Boost may give Intel a performance kick in AI and machine learning workloads, but the company is going to be fighting a decidedly uphill battle and thus far, the data we’ve seen suggests these factors can help Intel match AMD as opposed to beating it.
How Much Do Xeon’s Really Cost?
The list prices we’ve been quoting for this story are the formal prices that Intel publishes for Xeon CPUs in 1K units. They are also widely known to be inaccurate, at least as far as the major OEMs are concerned. We don’t know what Dell, HPE, and other vendors actually pay for Xeon CPUs, but we do know it’s often much less than list price, which is typically paid only by the retail channel.
The gap between Intel list prices and actual prices may explain why Threadripper hasn’t had much market penetration. Despite the fact that Threadripper CPUs have offered vastly more cores per $ and higher performance per dollar for two years now, the OEMs that share sales information, like MindFactory, report very low sales of both Threadripper and Skylake-X. Intel, however, has also shown no particular interest in slashing Core X prices. It continues to position a 10-core Core i9-9820X as appropriate competition for chips like the Threadripper 2950X, despite AMD’s superior performance in that match-up. This strongly implies that Intel is having no particular trouble selling 10-core CPUs to the OEM partners that want them, despite Threadripper’s superior price/performance ratio and that AMD’s share of the workstation market is quite limited.
While Intel has trimmed its HEDT prices (the 10-core Core i7-6950X was $ 1723 in 2016, compared to $ 900 for a Core i9-9820X today), it has never attempted to price/performance match against Threadripper. If that bulwark is going to crumble, Rome will be the CPU that does it. Ryzen and Threadripper will be viewed as more credible workstation CPUs if Epyc starts chewing into the server market.
Intel is Playing AMD’s Game Now
Intel can cut its prices to respond to AMD in the short-term. Long-term, it’s going to have to challenge AMD directly. That’s going to mean delivering more cores at lower prices, with higher amounts of memory supported per socket. Cooper Lake, which is built on 14nm and includes additional support for new AI-focused AVX-512 instructions, will arrive in the first half of next year. That chip will help Intel focus on some of the markets it wants to compete in, but it won’t change the core count differential between the two companies. Similarly, Intel may have trouble putting a $ 3000 – $ 7000 premium on support for 2TB – 4.5TB of RAM given that AMD is willing to support up to 4TB of memory on every CPU socket.
We don’t know yet if Intel will increase core counts with Ice Lake servers, or what sorts of designs it will bring to market, but ICL in servers is at least a year away. By the time ICL servers are ready to ship, AMD’s 7nm EUV designs may be ready as well. Having kicked off the mother of all refresh cycles with Rome, AMD’s challenge over the next 12 – 24 months will be demonstrating ongoing smooth update cadences and continued performance improvements. If it does, it has a genuine shot at building the kind of stable enterprise market it’s desired for decades.
Don’t Get Cocky
When AMD launched dual-core Opteron and its consumer equivalent, the Athlon 64 X2, there was a definite sense that the company had finally arrived. Just over a year later, Intel launched the Core 2 Duo. AMD spent the next 11 years wandering in the proverbial wilderness. Later, executives would admit that the company had taken its eye off the ball and become distracted with the ATI acquisition. A string of problems followed.
The simplistic assumption that the P4 Prescott was a disaster Intel couldn’t recover from proved incorrect. Historically, attacking Intel has often proven akin to hitting a rubber wall with a Sledgehammer (pun intended). Deforming the wall is comparatively easy. Destroying it altogether is a far more difficult task. AMD has perhaps the best opportunity to take market share in the enterprise that it has ever had with 7nm Epyc, but building server share is a slow and careful process, not a wind sprint. If AMD wants to keep what it’s building this time around, it needs to play its cards differently than it did in 2005 – 2006.
But with that said, I don’t use phrases like “golden age” lightly. I’m using it now. While I make no projections on how long it will last, 7nm Epyc’s debut has made it official, as far as I’m concerned: Welcome to the second golden age of AMD.