These are what I think of as “official” leaks. Usually starts about two months before the official announcement. Note that even if Zen 4 and 13th Gen Core parts are announced in late September, shipping could be announced for early October, and in fact, I expect that for both Raptor Lake, b>Raphael and probably Genoa.
I am a bit amazed that Raptor Lake will only have a 5% IPC growth over Alder Lake. The word on the street has been that Raptor Lake improvements over Alder Lake would mostly be clock speed increases and more E cores in some models. Shrug. As for Zen 4 the big IPC improvement will be due to going to 1 Meg of L2 per core from 512k in Zen 3. Again shrug. Much more important will be how long it takes AMD and Intel to get the new CPUs shipping all down the line. I won’t be surprised if AMD only releases three Raphael SKUs to start with, and the same for Intel with Raptor Lake. Upgrading to Raphael will be expensive to start with. Eventually, DDR5 prices will come down, and low-end B650 motherboards will show up. People who are sensitive to price will stay with current products to avoid the need for new motherboards and memory.*
I do expect AMD to release Ryzen 7 7700 and 7800 before Christmas. AMD has been leaving Ryzen 3 SKUs to APUs, and I see this starting to be the case with Ryzen 5. There are currently 3 Ryzen 5 5xxx SKUs that are six cores but without graphics. The 7000 series is expected to have GPUs in all models, including R7 and R9 SKUs. We will see.
I sort of feel silly saying this. I currently have a Ryzen 9 5900. My son in the next room over has a 5700 and his son is visiting, so we fixed him up with a Ryzen 7 5800H and my old graphics card. I needed multiple chips to test the effects of cache size on some software cores. Might as well get more use out of them. I suspect that I’ll need to do some testing with DDR5, so my son may get a new motherboard, CPU, and memory for Christmas.
I am a bit amazed that Raptor Lake will only have a 5% IPC growth over Alder Lake. The word on the street has been that Raptor Lake improvements over Alder Lake would mostly be clock speed increases and more E cores in some models. (emphasis and boldface in original)
Hmmm…
Could this be indicative of an Intel R&D unit that’s still underperforming, coupled with a marketing/PR decision that it’s better to announce something called Raptor Lake, even if the improvements over Alder Lake are underwhelming, to check a box on a product road map?
This is the last? product on the Intel N7 process. The improvement is comparable to previous annual steps. We won’t see more performance gains from INTC until they release their N4 products.
Raptor lake primarily moves the high end desktop product to 32 threads to match what AMD is doing.
Alan
Could this be indicative of an Intel R&D unit that’s still underperforming, coupled with a marketing/PR decision that it’s better to announce something called Raptor Lake, even if the improvements over Alder Lake are underwhelming, to check a box on a product road map?
Not really. From the beginning, this has been the Raptor Lake story. Same process as Alder Lake, some IPC gains and some clock gains. Just pick up the low-hanging fruit. While planning for significant architecture changes–and a new process node (Intel 4)–in 2023 (Meteor Lake), then Arrow Lake and finally Lunar Lake both in 2024. I don’t think Intel will manage to ship Meteor Lake in 2023. Even if we ignore Intel’s use of TSMC’s N3, a new process every year is something that Intel has never really accomplished. AMD has come a lot closer, but usually by using new layout macros with the same physical process.
So it sounds like Raptor Lake will be around a ten percent improvement over Alder Lake, ignoring cases where Intel plans to add CPU cores to otherwise identical models. That’s not such a big deal, but Raptor Lake releases stretching into 2023 will be. Alder Lake may have gotten Intel back to parity with AMD. Trying to keep up may lead to another train wreck. AMD’s slipping of Zen 4 has been due to external market factors.
Translation, since Zen 4 is DDR5 only, AMD couldn’t release Zen 4 without new motherboard designs and DDR5 dropping in price. Zen 5 has kept with its schedule, but AMD will have a different sort of problem. Bergamo is planned for servers in 2023. AMD can, and probably will release low-end Zen 4 desktop and mobile chips based on the same core as Bergamo. That may push Zen 5 into 2024 just to keep a year between new server products. (Bergamo and Zen 5 may use the same TSMC process. As I understand it, TSMC’s N5 and N4 are variations of the same process, rather than different processes. N4 and N3 are different processes, and TSMC has been having issues with getting N3 working. As of right now, AMD is planning to use N4 for initial Zen 5 production and migrate to N3 later if it makes sense. (My guess is that what will make sense is to use N3 for Zen 6, and not at all for Zen 5.)
I don’t think Intel will manage to ship Meteor Lake in 2023.
You think Intel will have that brand new factory in Ireland sit idle for over a year? Intel continues to say production wafer starts this year for Meteor lake, so missing 2023 entirely is a HUGE slip on something that is fairly close in time. I suspect Intel will update us on this at earnings next week.
With the smallest meteor lake die at only 40mm they can ramp the process even if defect densities are high.
Alan
You think Intel will have that brand new factory in Ireland sit idle for over a year? Intel continues to say production wafer starts this year for Meteor lake, so missing 2023 entirely is a HUGE slip on something that is fairly close in time. I suspect Intel will update us on this at earnings next week.
If Intel starts (unintentionally) building up unsold Alder Lake or Raptor Lake chips it could be the 10 nm train wreck all over again. Intel had several fabs sitting idle during that mess. I don’t see a similar issue here. What I see is that Intel has several fabs that can manufacture Raptor Lake but not yet Meteor Lake. If Intel introduces Meteor Lake too early, they will end up with Raptor Lake capacity sitting idle or producing chips to be sold at a loss. On the other hand, if Intel pushes out the Meteor Lake release, they can build inventory in Ireland and either have it sit there for a few months or sell the chips into the HPC market. A few high-end HPC systems could use up the capacity in Ireland. We could discuss when Intel will have enough Sapphire Rapids server chip capacity. But I have little or no visibility into that market.
I am a little confused by your post.
Intel only has 2 N7 factories (shared with any other 10nm output), when traditionally they have needed 3 factories to service the market. AMD has taken up some of that slack.
Intel server market is currently a mixture of old 14nm and 10nm icelake. As this shifts to N7 Sapphire rapids Intel needs more N7 capacity.
Intel desktop is currently a mix of old 14nm rocket lake, and new Intel Alder Lake. As this shifts to the larger die Raptor lake Intel will need more N7 capacity.
Intel Notebook market is mostly tiger lake on 10nmSF, and is shifting to the larger die alder lake on N7. Some of this will shift to the even larger die Raptor lake on N7, and some to Meteor Lake on N4.
It seems Intel needs to accelerate the shift of notebooks to N4 in order to have enough N7 capacity to meet desktop and server.
Increased AMD market share along with a contracting market may allow Intel to accelerate the conversion of N7 to the larger die Raptor lake product.
I reread your response, and perhaps should add a bit more.
The 2 old Intel 7 factories (built ten years ago) will never be upgraded as they are not EUV capable. Technically fab 42 is, but by time they are ready to convert they would want high NA machines, which won’t work with it.
Intel capacity does not catch up with Intel demand until 2025… they will need additional wafers from TSMC until that time.
Perhaps I have mentioned it, but the largest die on the meteor lake SOC is built on the old Intel 14nm process. The GPU tile is on TSMC (N6 or N3), and the smallest of these is the CPU cores built on Intel 4.
Now that Intel is more or less trying the foundry model, old factories will be less likely upgraded and more likely remain in production on older nodes for the foundry. They are doing this already by building automotive products in the old 14nm factories.
Alan
Not much. Regarding details of fabs and processes, you are the expert here on Intel. I tend to pick up the details of TSMCs processes. I still do some tuning of libraries, but only the AMD versions. (Other people do the tuning for Intel CPU cores.) This requires knowing the gory details of the various Zen CPU cores. It is not enough to figure out how to do as much work as possible in one clock cycle, I need to tell the CPU when to ask for the next cache line. I find that describing the details quickly puts almost anyone to sleep, so I’ll spare you. What matters here is that I get access to CPU design details and even sometimes access to some ES silicon–by asking nicely someone who has ES silicon to run some code for me.
I have no such access or interest in the details of Intel fabs. So depend on Alan for that. I do need to verify that code highly optimized for Zen CPUs doesn’t mess up when run on an Intel CPU. I usually ask someone with a recent or even engineering sample, Intel CPU to provide me with timings. If something messes up, I have to start digging into Intel CPU internals. As long as the code runs at least half as fast as on the AMD CPU, they should have gotten the compiler flags right. Again, the actual library as distributed has more than one code version–but I assume people here have no want to know how to build the libraries so the right code gets linked in.
eachus: A few high-end HPC systems could use up the capacity in Ireland.
HPC systems take a very small percentage of CPUs made, like 0.1%. There is no way a few could use up a fab’s capacity.
The #1 system in the world, Frontier, has 8,730,112 AMD cores, or 1.1 million chiplets. Each chiplet is 72 mm². That’s about 837 candidates per wafer, let’s say yielding 400 good dies (probably low). To get all the chiplets needed for Frontier, there needs to be 2728 wafers. Most fabs are 5000-7000 WSPW. So a typical fab could turn out ~2 Frontiers per week.
Meanwhile, in April, “TSMC Boosts 5nm Production To 150,000 Wafers/Month Amidst Strong Demand”. That’s the equivalent of 7 Frontiers per day.
All great information! As an aside, there is no HPC product designed on the Intel 4 process. There was a start-up GPU originally, but INTC has said nothing about it other than it is currently on TSMC N5.
Intel 4 is going to be almost exclusively meteor lake clients. The broader product portfolio is designed on the slightly better Intel 3 due out a year later.
Alan
There was a start-up GPU originally, but INTC has said nothing about it other than it is currently on TSMC N5.
I was thinking about GPUs. As I said, I don’t follow Intel’s technology closely, just what I see when looking at AMD’s plans. I gather that the A-series of Intel GPUs is showing much less performance than expected. I don’t know what this means for Ponte Veccio. The problem with Archmage seems to be in the drivers, but HPC drivers are pretty simple. (Mostly do matrix multiplication fast, and use a BLAS that does most of the heavy crunching as matrix multiplication.)
The consumer ARC A-series GPU’s are on TSMC N6, like some of the recent AMD CPU upgrades.
Ponte Vecchio HPC GPU is on TSMC N5HPC, just like AMD Genoa.
I think the problem with ARC is the software requires custom optimization for EVERY software title. AMD and Nvidia have been writing optimizations every year as titles come out, while Intel is playing a game of catch-up. Some titles run great and some not so good.
Alan
