We have a bit of an update from the folks over at Intel, they actually had a slick presentation and they gave us a little bit of insight on some stuff, including new processors, upcoming architectures, future manufacturing processes, XE graphics, their plans for datacenters, artificial intelligence, and a bunch more stuff. Instead of bombarding you with everything, I’m going to break this article into a few sections, each one using Intel’s catchy phrases.
New Manufacturing Ideology
There is SuperFin – which is a new manufacturing process – and the Willow Cove architecture that uses it. Then there is Tiger Lake CPUs – incorporating Willow Cove alongside a new XE graphics engine – and I will also go over exactly what XE LP is. Finally, I wanted to cover Intel first discrete graphics card in two decades, which is going to be rolled out this year. And finally, the gaming GPUs codenamed XE HPG. I know that it is a lot to process, so I’m going to break it down into highlights. However, I will say this: The end of 2020 and 2021 is definitely going to be an interesting space for both the CPU and GPU market. Now let’s discuss Intel’s plans.
Let’s start off with the problem that Intel found themselves in the last few years. Basically their architecture evolution has been tied to the hip of new manufacturing processes. This that means if a process didn’t work out, future designs would be delayed or canceled altogether. That is what happened with 10nm after 14nm and it completely broke Intel’s Tick-Tock model. As a result, going forward we can forget about refreshes followed by new architectures. Not only that, they are behind in the process node battle, and they know it so something had to give. We are coming to a point where moving to a different node, like 7nm or below, will give diminishing performance returns, both AMD and Intel have admitted that. Therefore, enhancements need to be made to the architecture itself for it to take full advantage of existing process technologies.
Those can have much more beneficial returns than a node change. What is going to replace Tick-Tock? Well right now it looks like Intel will be sticking to updating the 10nm process by combining something called SuperMIM technology with an enhanced FinFET process. They are calling it 10nm SuperFin, rather than 10+++++. There is quite a bit going on behind the scenes rather than just a normal refresh though. There could be an entire article dedicated to what Intel has done to create SuperFin, but in a nutshell they have completely revised the 10nm process with top to bottom updates. It impacts everything from the transistor gates that can affect how current is managed, to the addition of a SuperMIM capacitor stage for helping efficiency at higher frequencies. Supposedly, the 10nm we knew from 2018 and 2019 was deconstructed and rebuilt, which sounds like the only thing Intel could have done after all its failures. Either way, they are claiming the largest performance jump we have seen between two architectures on the same process.
In the future it looks like 7nm has been scrubbed from the roadmap and replaced with a 10nm Enhanced SuperFin process. Will this work versus AMD’s aggressive movement to new manufacturing processes? Well we are entering a new era where Intel believes they can get better results from optimizing their architectures rather than tying them to a new process, so it could get really interesting. Like all the “Coves”, Willow Cove is considered Intel’s high performance architecture, while the “Monts” like Tremont and upcoming Gracemont are ultra-low-power cores. When you combine the high and low power architectures together, you get Intel’s hybrid CPUs like Lakefield and another revision called Alder Lake in 2021.
Willow Cove Cores
This comes together with the Willow Cove architecture, which will feature all of the updates baked into 10nm SuperFin, along with its own enhancements for caching and other elements. Consider it a huge update to the Sunny Cove cores found in today’s Ice Lake chips. However, because of all the changes, the best way to view it is Intel building a whole new house on Sunny Cove’s foundation. As for the actual performance of Willow Cove, well Intel didn’t really say anything. They did show a little graph that explains that the new architecture can reach higher sustained clock speeds at the same voltage as Sunny Cove. Not only that, but due to the 10nm SuperFin updates, it also has a wider voltage range. Now I want to pause here for a few seconds and focus on this part. First of all, the clock speeds are obviously based on a completely theoretical single core Turbo workload. The number of times I have seen the Sunny Cove architecture in an Ice Lake CPU actually hitting 4GHz is zero. Either way, Intel claims Willow Cove can supposedly hit 5GHz, but there was absolutely no mention about multi-core frequencies, and that is where previous 10nm products like Ice Lake got their butts kicked. All of this basically means that Willow Cove can use less voltage to achieve the same or possibly better clocks than the previous generation, and that means improved efficiency… maybe.
Tiger Lake Processors
This brings us to Tiger Lake, which will be a notebook-focused CPU series that is going to be the star of Intel’s mobile lineup starting next month. That is because it incorporates the Willow Cove architecture, huge updates to the memory subsystem, their next generation graphics engine named XE, and a bunch of other things. In practical terms, this means Tiger Lake will natively support DDR4-3200 and LP4x-4267, but it can also incorporate DDR5-5400 in the future, which is just crazy. Meanwhile, memory bottlenecks were eliminated by moving the controllers to an 8×16 layout rather than the previous generations 4×32. They have also doubled the last level cache to 24MB. Tiger Lake natively incorporates Thunderbolt 4 and USB 4.0, which offers a massive 40Gbps of bandwidth. However, remember it is ultimately up to the notebook manufacturers to enable these technologies, so don’t expect them to see them on every single device. And yes, PCIe Gen4 is finally arriving on Tiger Lake for both storage and discrete graphics cards. Another thing Intel did is made all the onboard devices like USB and display controllers more efficient, which frees up more onboard power for the high performance CPU cores. The interesting thing here is Intel is significantly expanding Tiger lake’s power envelope. Remember Ice Lake mobile CPUs were available at multiple performance levels from 9W to 28W, but this new lineup can run from 10W all the way up to 65W.
Another significant step forward is Intel’s next generation graphics architecture called XE being added to Tiger Lake. In this case, it’s the XE LP or Low Power version, but it’s supposed to offer almost double the performance as Gen11 graphics found on Ice Lake. Some of that improvement is due to the fact that XE will have up to 96 Execution Units (EUs) instead of 64, but most of it comes from Intel’s ground up redesign of the execution units themselves. The new XE design introduces shared a thread control per pair of use, and more logic units, which increases overall graphics horsepower by a significant amount. There is a whole lot more too, but I will talk about that in a separate article. Now those new stages are integrated into a much larger graphics engine with a new memory subsystem, and a lot more local cache than anything else Intel has built so far. There is also an enhanced media engine that is supposed to double encode and decode throughput, so that’s good news for creators I guess.
With the 10nm SuperFin technology working in the background, XE can hit higher sustained and burst frequencies than Gen11 at a fraction of the power cost. According to Intel, a 15W XE LP GPU integrated into Tiger Lake will have the same performance as a 25W Gen11 in Ice Lake. And did they prove it? Well no. In fact, during the presentation they did a few side-by-side on-screen demos of Battlefield and Grid without any frame rates or any performance metrics. Honestly, I was expecting a lot at this point, but one way or another – until AMD incorporates their higher-end Navi graphics into their notebook processors – XE coming into the market is certainly something that they have to be afraid of, because as it stands Gen11 graphics is super competitive against the GCN cores in the Ryzen 4000 U-series processors. Graphics is supposed to be AMD’s home turf and here we have Intel coming in and saying “Hey, we are going to double performance!” so that might be a bit of a problem for AMD given how a lot of these key applications are moving towards GPU acceleration these days.
Speaking of graphics, Intel did confirm that they are launching their first discrete GPU this year, but it won’t be coming to a desktop near you. Sorry about that. It’s only being added to higher-end Tiger Lake notebooks and targeting content creators. The underlying architecture will be the same XE LP that is in Tiger Lake, which means it’s based on the chip inside the DG1 development prototype that was rolled out to developers last year. It will have a high frequency and more execution units than the integrated graphics card as well. Intel really wants to avoid calling this a gaming graphics card, because the primary purpose of the DG1 is to accelerate media transcoding and other compute tasks. In many ways I get it, Intel GPU drivers are still a work in progress and their slow rollout doesn’t allow them to quickly push updates for the latest games like gamers expect. Incorporating DG1 into a notebook allows them to test a bit more and play with some interesting concepts as well, like allowing the unified driver architecture to better handle power sharing and switching between the GPU and IGP, or even introducing new technologies like variable rate shading. There is even the possibility of using both graphics processors together and boosting performance in multi-GPU mode.
All of this leads to an update in Intel graphics roadmap. There used to be a derivative of XE HP for enthusiast and mid-range GPUs, but now there is a new XE HPG or High Performance Gaming architecture. HPG incorporates Intel’s learnings from the compute-focused HBC models and the efficiency of the low power notebook GPUs, along with the scalability of XCHB. That’s probably confusing, right? But that last part is super important since it’s their intent for the gaming GPU to scale from an entry-level all the way up to high-end enthusiast parts. The reason why they are focusing on integrated first is simply due to market size. There is a billion plus users without a discrete graphics card in their systems, so starting small and then reaching upwards kind of makes sense. Unfortunately, there weren’t many more new details, but what Intel did reveal was still interesting. Their gaming GPUs will have embedded hardware accelerated Ray Tracing, and will be based off of a GDDR6 memory architecture. That means they are avoiding AMD’s mistakes by ignoring super expensive and hard to produce memory like HBM.
Right now it looks like the first GPUs will be rolled out in 2021, but based on this presentation we are only going to start to see entry-level graphics cards. They are not targeting high-performance gaming until sometime in the future. Intel also confirmed a lot of the rumors by saying that the XC HPG cores will not be produced in their own foundries, but rather using an external process technology. That pretty much wraps this up, I really wish I was able to show you actual benchmark numbers and specs, but this whole presentation was more about talk and short on results. While Tiger Lake is right around the corner, I’m sure there is going to be a lot more to cover until that point, but it looks like the end of this year moving on to 2021 is just going to be crazy in the CPU market. I can’t wait to see what Team Blue and Team Red have to offer. I’m sure you are excited as well.