Lets start out with the bad before we get to the really bad followed by a glimmer of hope. Intel is getting beaten like a drum in the CPU market. Yields on 10nm are still not economically viable and never will be, their cries of, “it is better than expected” and, “yields are improving fast” seem to always omit the baseline. The numbers SemiAccurate knows are still startling bad which is why Intel is intentionally limiting 10nm wafer starts to a paltry number.
Things like this tend to come out when Intel lets the press ask questions so as is their norm they sent an unannounced ‘briefing’ deck, nothing more than a few chirpy bullet points really, at 7:33pm, 13 hours and 27 minutes before the embargo lift. The intent is obviously to have as many people as possible miss it until the last second so they don’t have time to think about how bad this part is much less ask questions. Any that are asked will be sent in after business hours on the west coast so they might get a response a few minutes before the deadline. This intentional but plausibly deniable stiff-arming of questions is a sure sign the product is a dog, and trust us this one is. Sleazy moves Intel.
Now on to the really bad, this Rocket Lake part and the ‘new’ Cypress Cove core. We will admit we were wrong when we said that Intel would never backport a 10nm architecture to a 14nm process because that is exactly what Rocket/Cypress is, take Ice Lake and backport it. Also backport the Xe GPUs and slap them on for good measure.
Rocket Lake core features
What we should have said is that Intel would never be dumb enough to backport a 10nm architecture that was not designed with process portability in mind to a 14nm process. It would be too big, too inefficient, and simply not competitive with IP blocks designed and tuned for 14nm from the start. A company would need to be desperate and have their backs to the wall so badly that flushing tens of millions of dollars on this port for nothing more than a headline grab and an eventual black eye is worth it. That’s what we should have said but we would have been wrong there too, Intel it seems _IS_ that dumb.
Why is Rocket/Cypress so bad? If you look at the layout of modern CPUs on differing processes, the core area doesn’t go down much with a shrink. Wider datapaths, ISAs like AVX-512 and VNNI, and other bloats like increased caches grow the core area to something close to what it was on the older process. You get more features and higher core IPC but not a smaller one.
Companies don’t backport because a core is designed for a process. It’s efficiency and architecture is based on the transistors that it will use from the design start, caches are sized based on the process too, and a lot of the architectural gains are due to the added transistor counts. If you backport a CPU you lose the efficiency of these new transistors so the energy use goes up and the clocks likely go down as well. Cores take up 2x the size, or at least a lot more than the older core did on the older process, so costs go up too. You can either cut out bits of the core and lose performance or eat it on area and therefor cost. Backporting a design not made for portability isn’t a lose/lose proposition, it is a lose/lose/lose/lose proposition. But Intel is desperate so…
On top of this there are a few more fundamental takeaways. Intel can’t make 10nm devices, yield still isn’t there and Rocket/Cypress is a shining example of the problem. Intel’s 10nm process is unsuitable for high performance, high power devices. We can’t explain the details yet, it will take some digging, but if Intel needed a 10nm halo product, this would be it. Rocket Lake is 14nm.
Lastly it shows SemiAccurate was right about the stunningly low 10nm wafer start capacity, we have been saying for over a year that Intel is intentionally misleading anyone listening about their 10nm capacity, Rocket/Cypress is again a shining example of the numbers we have been talking about. Intel doesn’t have 10nm capacity, period, their total wafer starts are far less than a full modern fab worth but they spread it across three sites to misdirect. No one asks the right questions about this which is what Intel wants.
So the backport is a desperate attempt to spin a headline from a shatteringly bad idea. 10nm yields are not economically viable, 10nm wafer starts are barely extant, 10nm is unsuited for high performance products, and Intel is doing their damnedest to shut down even the tamest questions on the topic. They want happy headlines to distract from the fact that AMD has the product they want and is using them to beat Intel into the ground in the market. Several quarters ago Stacy Rasgon asked Intel management when they would cross over from 14nm to 10nm production and they dodged the question. Now you know why, they never will exceed 14nm production with 10nm, ever.
Intel Rocket Lake platform features
With this disaster now being shouted about by Intel, albeit in code, what are they talking about today? Rocket Lake is the 14nm backport of Ice Lake and uses the Cypress Cove core. The bullet points are up to 8C/16T, DDR4/3200, PCIe Gen 4, Xe GPU, new media encoders, USB 3.2 Gen 2, up to 3 4K monitors, and VNNI. Sounds good right? Sure until you ask some questions.
Lets start out with the core. Intel calls it “new” with IPC improvement but it is just the year plus old Ice core backported. Technically they are right but no sane observer would call that new unless they were technically unaware and not given the chance to ask questions. The 20 PCIe4 lanes which Intel says, “Allows both SSD and Discrete Graphics Direct CPU Attach” shows just how bad their current line is because you can’t do that now.
That said AMD has been offering 24 as standard for years now so best case Intel is still way behind. That said they don’t break down how many are on the CPU and how many are on the chipset, plus how many are needed for the chipset connection, just vague wording that implies they are all directly on the CPU. Given Intel’s honesty in messaging of late, don’t believe it until you see the real specs. Allowing real questions would clear this up but for some reason Intel is desperate to avoid that eventuality.
The new memory controller will catch Intel up to AMD, a year late but that is better than before, as will the new USB 3.2 Gen 2 controllers. The media encoders and added display resolutions and outputs are nice but barely catch up to AMD’s last gen part. Don’t forget Rocket Lake is not a 2020 part, it is a Q1 2021 device that should be out a month or two before AMD’s updated APUs hit. Good luck beating those Zen 3 cored devices with a 14nm backport.
Then we come to the last bit, the vaunted Xe graphics designed for 10nm transistors. It is backported to 14nm so onward to victory, right? Remember what we said above about backports being power hungry, bloated, and slow? Well Intel is claiming that the GPU in Rocket has 50% more performance than Gen 9 graphics! Wow! Right? Remember that bit about not wanting questions asked?
The first one SemiAccurate would ask would be along the lines of, “Xe on 10nm in Tiger Lake guise is 2x faster than that same Gen 9 part. Why do you spin this as a win rather than a flaming disaster wrapped in a distraction attempt?” We probably wouldn’t get an answer mind you but we would ask but probably in a less friendly way. As you can see the area, efficiency, and raw performance losses of backporting lost double digit performance versus the same architecture on 10nm. Without die area figures you can’t tell how much area Intel burnt to get this loss but we will go out on a limb and say it is substantially larger than the 10nm equivalent.
Then we come to the cores, all eight of them. The current top 14nm part, the 10900K, has 10 cores and runs at the same 125W TDP. If you recall that Ice core has a 18% IPC increase over the 14nm cores which is a great number until you realize that is against a 2015 core and the Cannon Lake generation was omitted from the history lesson. And you don’t know that AMD’s Zen 3 core is a 19% uplift from last year’s Zen 2. But if you don’t have time to think or ask questions before you write, that performance increase of 18% seems pretty darn impressive.
Then we come to the math which may not be SemiAccurate’s strong point but we do think that the 14nm 10900K’s 10 cores are 25% more than the 8 cores in the Rocket Lake part. Where things get tricky is that 25% does indeed seem like a bigger number than 18%. Actually Rocket may have the lead in CPU performance if, and it is a big if, it clocks as high as the 10900K. Given how badly a backport hobbles performance, increases energy use, and all of that, this is unlikely but it could happen. Then again Intel talks about architectural enhancements and IPC but curiously omits any talk of actual CPU performance. They do talk about GPU performance so infer what you will. We will again go out on a limb and say that Rocket will lose to the current parts on most benchmarks for the reasons we discussed above. The details would come out if questions were allowed though, once again any guesses why the disclosure format precluded them?
Last but not least is another kick in the teeth to enthusiasts. If you want a Rocket Lake CPU you will need a new motherboard, it isn’t compatible with anything Intel has done and given the hacky nature of Rocket, it is unlikely to be compatible with anything that come out later. How many generations is AMD on for the same boards now, is it three or four, we forget. Either way if you invest in a new setup you get a slower part. Wait, that’s not good.
If you think there is nothing good about Rocket Lake and Cypress Cove you would be wrong, there is one thing that Intel is doing really right, disclosure. Even for the most desperate bullet point regurgitation they put out their system specs, documented things in a way that is recreatable, and did the right thing for the right reasons. Well done Intel, AMD could learn from you but they won’t. Don’t let it stop you from doing the right thing in the future though.
So in the end what do we have? Desperation on display and millions squandered to make an inferior product for grabbing headlines. Even then Intel had to pull the sleaziest of stuns with the release timing because even their best efforts won’t stand up to minimal scrutiny. The backport is just a bad idea borne of desperation because the Intel 10nm process simply doesn’t work and never will. There is no capacity, no way out, and absolutely nothing to counter AMD with this year. Or next. Or 2022, it really is that dire of a situation. Rocket Lake will do nothing to move the needle forward but could move it back, another in the long line of Intel’s self-inflicted wounds. Please stop digging, this is getting too painful.S|A
Latest posts by Charlie Demerjian (see all)
- Intel looks at laptop power usage - Nov 23, 2020
- AMD slide confirms our delay story - Nov 19, 2020
- AMD stays trendy and delays Milan a bit - Nov 13, 2020
- Intel delays a mainstream server platform once again - Nov 12, 2020
- Intel releases Server GPU and significant software updates - Nov 11, 2020