Speaking of subsidies, here is what I can gather from the current state of grift. - 8.5B direct grants from CHIPS act. - 2.5B from military money earmarked to reshore armament and intelligence semiconductor sources. - 11B in federal loans. - 25% investment tax credit. Intel is behind on a lot of things. Arrow Lake should catch Intel up around the start of 2025 but the fruit of that labor won't start hitting the balance sheet until Q1 reporting and won't be strong until Q2 reporting.. They might match AMD's Zen 5 on performance. It is expected to be very similar but they will be usurped by AMD again with Zen 6. Of course, Intel should have Lunar Lake online by mid next year. Intel is promising, "leadership in compute, graphics and AI," so perhaps my pessimism is misplaced. To put the above numbers into context, TSMC's Fab 21 (located in Arizona) is expected to cost about $40B when the second phase is complete.
This two week old graphic shows that Intel is leading the industry in PowerPoint slides that show it is leading the industry. For clarity, Intel 7 is a 10nm process node. Intel 4 is a 7nm process. Much like TDP which no longer means "total dissipated power" but has quietly been redefined as "Our advertising team's optimistic guess at Typical Dissipated Power, process nodes no longer represent actual feature size. This is true across the industry, not just Intel.
This seems like a buy opportunity, for people who believe in the company. I wish I did. To be clear, I do not predict that Intel will fail. They may succeed. I just don't know. For people who know Intel is going to succeed, this is a clear buy signal. For the rest of this year, I see a lot more down side than upside.
Intel is too big to fail. Through all the problems, I see a bright future with the new ASML machines and their plan to manufacture < 3 Nm CPU like TSMC does. Best time to buy is now.
For what it's worth, I completely ignore rumors of Nvidia buying Intel. While Nvidia can afford it, they were blocked from buying ARM so this would be pretty open corruption on the part of the regulators. These days, corruption is daily and glaring so it is possible but I chose to believe the odds are low. I'm more interested in the 2nm equipment Intel ordered from Applied Materials. They claim they don't need all of AM's tech to achieve this node; primarily, just the robotics. If this turns out to be hubris, Intel will be a total loss. They can just buy the process technology and carry on but it will be indicative the company's process IP is worthless.
Intel is claiming 45 TOPS from Lunar Lake. If this is real, it will put them on about the same footing as AMD Zen 5. There is some speculation Intel will not make the 40 TOPS speculated performance limit that Microsoft is speculated to impose on local AI processing for copilot.
Intel has shown a slide of the Lunar Lake mobile CPU complex that shows RAM on board. I've heard 16 and 32GB versions will be available. This makes a whole lot of sense to me. If we want laptops to be cheaper with lower power use, we're going to need more integration.
Moore's Law is Dead just talked about a Dell roadmap that suggests M.2 graphics cards are coming. It makes sense, since M.2 is 4x PCIe lanes. Throw in PCIe 5 and you have a lot of bandwidth. I'm posting this under INTC because M.2 was introduced by Intel. DELL uses Intel exclusively (for now). Also, there are rumors that Intel has taken a huge step forward on their graphics platform. It's plausible. Laptops with expandable graphics would be a nice design win for Intel. AMD and Nvidia would probably follow shortly but Intel seems to be breaking this ground and will have first mover advantage for a while.
We can stop pretending Arc is a big threat to Nvidia or AMD. AMD is going to dominate CPU for the rest of 2024. They will dominate all niches. The question is, can Intel come back strong in 2025? I believe they can. At the very least, AMD dominance may be short lived. It will depend on TSMC N3p node coming online with decent yields by October, if they want processors in the retail chain by the start of 1Q25. I believe the odds of this process being ready are roughly 70%. They can produce product with N3p now but they need to improve yields and have four months to do it. The wildcard threat here is the Qualcomm Snapdragon X Elite. A whole lot of handheld users would trade 20% performance for 40% battery life. 10h battery life will be a standard feature, in 2025. I believe we are entering the post-performance era of compute. Any system made in the last 3 years is plenty fast enough for 95% of the population. Intel is pricing their products aggressively to compete with AMD. They will lose the high end for the second half of 2024. Arrow Lake will be very competitive single core performance but it will max out at 8P cores. We are starting to see that people don't care about E cores. I think E cores are a dead end, although probably could be extremely relevant in the server space. Intel is working on a 2 nano node. If they can stand that up and move to all performance cores, they could dominate again. The N2 node is, by a vast amount, the biggest hurdle to achieve that.
It will be interesting to see if Intel can convince the YouTube influencer world that RAM on the CPU complex is a good thing. Influencers are nerds. They want to be able to change and upgrade stuff. 10 years ago, RAM on complex would have been dead on arrival. Today, I think Intel has a shot and I think the idea is brilliant. This may be the first time in compute history we will be able to lower the pin count on a CPU socket. I can imagine an SoC with 400 pins/balls. All they need are maybe eight PCIe v5 lanes wifi, SSD, and expansion, two USB v4.0, HDMI out (I believe DisplayPort is going away due to committee politics), and a bunch of power. To the extent they can get the industry to accept this architecture, Intel can effectively eat the lunch of system integrators, DRAM vendors, etc. If you think about it, Intel's real problem is being able to compete with Qualcomm on power management (Note: I did not write power consumption).
I want to mention Intel's move away from HyperThreading (tm). I believe I understand it but discussion might shine some light on this for all of us. Hyperthreading was originally a way to more efficiently use die space to enable multi-threaded applications without adding cores. It originated in the days of single and double core counts. In some situations, hyperthreading is a mess that chokes on diverse compute loads but it is effective in other situations. Over the years, HT has evolved to be pretty good under a lot of situations. That evolution has added a lot of complexity. HT was never 2x speed bump, though. In the early days, it was closer to 1.2x average and evolved to 1.5x, or so. I have no idea how it scales on recent silicon. As I understand Intel's dropping of HT, it comes down to them saying die space is no longer a constraint (thanks to both process and chiplets). Improving HT will add further complexity. They are at the point, it makes more sense to just add an entire second core than it does to add transistors for the purpose of HT. So, they can strip out HT and add in as many full cores as they want. Multi-core doesn't scale with perfect efficiency, but it is worlds closer than HT. Apparently, this is also related to the thrust behind efficiency cores. They are small and light so they can toss in a lot of them (14900 has 8P cores and 16E cores). As I understand it, efficiency cores perform similar to performance cores for integer tasks. Efficiency cores fall short when doing complex tasks, like floating point math or SIMD. When you think about it, the average person surfing the net does not use the FPU, SIMD, or AVX (wider extensions of SIMD). They are using a tiny part of the CPU with most of the die consuming power without being used. I'm inclined to think HyperThreading doesn't matter. Efficiency cores don't matter, either. Intel needs to post great benchmark results to get people to buy their parts. How they get there is their business. Literally. It is also clear the strategy of efficiency cores is not going to be effective against heavy work loads such as bulk video transcoding, math simulators, etc. These cases simply need as many heavy cores as they can have. Efficiency cores look like a very effective strategy against RISC designs, like ARM. Since this is where most of the industry is headed, I'm inclined to buy into Intel's strategy and am considering buying back into INTC.
I don't see much talk of why i9 13900 is faster than i9 14900 in some games. The emperor is clothed with Intel's marketing budget. From what I can tell, changes to CEP, process, and some other minor optimizations will provide a significant step forward for Intel. This, even without design improvements. I don't expect them to pass AMD in the next year, at the least, but I do expect them to remain somewhat close.
