With the volume of Intel 7nm* product we see in the market, I suspect they filled the AZ fab 42 with 7nm equipment rather than reserving it for 4nm per the “previous” plan.
By early 2023 Intel should have about $10B of capital deployed for 4nm production. Given chiplet technology I predict this should be enough to supply almost 50% of the x86 market.
Intel has designs in process on TSMC 6nm, 5nm, 4nm, and 3nm so it seems they should have a pretty good idea which TSMC processes they align with.
Instead of using the ITRS process names, this is using the TSMC node names.
Intel 4nm* * Instead of using the ITRS process names, this is using the TSMC node names.
I’m utterly confused.
Proposal: Let’s stop using “nm”. It is nonsense. Let’s use the official marketing names. From TSMC: N7, N7P, N6, N5, N5P, N4, N4P, N3 (preferably preceded by “TSMC”, i.e. “TSMC N5” or “TSMC’s N5 process”). From Intel: Intel 7, Intel 4, Intel 3, Intel 20A, Intel 18A (for clarity, use quotes or perhaps an underscore, e.g. “N7 vs ‘Intel 7’ density” or “their Intel_7 process”).
The ITRS laid out a very nice framework and benchmarks for process naming. TSMC chose to name their process that met the 10nm ITRS specifications as a 7nm process, which has created a huge amount of confusion; probably more so than your current confusion:-) It seems to me we should have a standardized name for a given set of process benchmarks, making it easier to compare TSMC, Samsung, and Intel. Given TSMC is the leader, it currently makes sense to me to use their names as the standard.
Our best data right now is the Intel 4 process is about equivalent to the TSMC 4nm process. As we get more data it might be prudent to reevaluate that equivalency.
Alan
I suspect they filled the AZ fab 42 with […] equipment [for their Intel_7 process] rather than reserving it for [Intel_4] per the “previous” plan.
Since Intel_4 is their first use of EUV, is this due to lack of EUV machines, perhaps?
By the way, as an interesting side note, it seems Intel is only planning on staying with this generation of EUV technology for just Intel_4 and Intel_20A. Then, with the Intel_18A process they are betting on stepping up to the next innovation, ie. high numerical aperture EUV machines (TWINSCAN EXE:5200). Although I guess they probably will use both type of machines in the same process for different processing steps. In any case, it looks like an aggressive plan in which much must go right to stay on schedule.
TSMC’s N4 or N4P?
Since there is no difference in transistor density and only a 6% difference in performance I am going to count that as in the noise…
Regarding your next comment: Since Intel_4 is their first use of EUV, is this due to lack of EUV machines, perhaps?
More likely due to the slip in the Intel 4 process schedule. This would have resulted in the need for more 7 to bridge the gap. It also meant the Ireland fab 34 would be ready for Intel 4.
There was an update from ASML last year that “a customer” had asked for a delay in receiving EUV machines, but they were able to ship those machines to somebody else.
Alan
Since there is no difference [from N4 to N4P] in transistor density and only a 6% difference in performance I am going to count that as in the noise.
Well, don’t dismiss the power improvements. Although not explicitly stated for N5 to N4, nor for N4 to N4P (not in the WikiChip article, at least), the improvement from N5 to N4P totals -22%. If we assume a similar ratio of improvement from N4 to N4P as for N5 to N5P, for which +7% performance and -15% power is stated, then we get 6/7*(-15) ~= -13% improvement, and an improvement around -10% from N5 to N4 (sanity check: (1-0.10)*(1-0.13)-1 ~= -22%).
PS. Thanks for adopting my proposal and not using “nm”.
Here is a December update which includes N4X. They often quote either a performance improvement OR a power reduction, but not both at the same time. I think the N4 family is still a little fuzzy on that front. https://www.anandtech.com/show/17123/tsmc-unveils-n4x-node-h…
Alan
Thanks for the reminder; N4X slipped my mind. It certainly can make a difference in terms of HPC performance achieved at TSMC.
Claims have been made that AMD’s upcoming “Zen 4” products are not using vanilla N5. Assuming that’s correct, and it is a custom process adapted to AMD’s needs, I wonder which of N5, N5P, N4, N4P and N4X it is closest to, and how it differs from these off-the-shelf offerings.
There has been talk of an N5HPC for release in Q2 2022. I suspect both AMD Genoa, and INTC Ponte Vecchio use this process.
I’m pretty sure that N5HPC exists and is getting beyond risk lots. Also, remember that AMD CPUs have one or more CPU chiplets and a single I/O chip. AFAIK, since AMD went this route, those chips have been fabricated on different technologies. I expect that to continue into the indefinite future. (Hmm. Say 4Q2025, that makes it the definite future.) My point, I guess, is that the Zen 4 CPU chiplets may use N5HPC, but I don’t think the I/O chiplets will. I’m being a bit pedantic here, but there are multiple new chiplets coming, and I am interested in the processes used for all of them.
I expect Genoa, Raphael, and Bergamo to use a new process for the I/O chips, probably N7 or later. Rembrandt will be a single die containing both CPU and GPU and maybe another process. (That’s not all the Zen 4 chip families. Gets complicated quickly doesn’t it?)
Rembrandt will be a single die containing both CPU and GPU and maybe another process.
Rembrandt, and the Intel ARC alchemist GPU are both on TSMC N6.
First Rembrandt benchmarks are leaking now: https://wccftech.com/amd-ryzen-9-6900hx-zen-3-apu-benchmark-…
–Alan
Yikes - I knew I shouldn’t have clicked that seemingly good news AMD link. The full title is:
AMD Ryzen 9 6900HX ‘Zen 3+’ APU Benchmark Leaks Out: 33% Faster Than Ryzen 9 5900HX But 30% Slower Than Intel Core i9-12900H
I’m getting the sickening feeling I should have sold all my AMD two months ago. Oh well, I probably won’t live enough to need the money anyway. (Sorry - feeling down and very mortal lately.)
I should have sold all my AMD two months ago
The server market is where the money is. While Intel continues to do “server last” while AMD does “server first” AMD will continue to gain market share, and hence $$$.
12 month performance:
AMD +21%
INTC flat
TSM flat
NVDA +72%
As long as you got in early, you should be way ahead of market performance.
Thanks for pointing out the potential AMD upside again. But I still wonder if it will be good enough to overcome the market sentiment which seems to be looking at INTEL eventually regaining total dominance again, even if that might be more than a year out. Yes, the market is always pretty nearsighted, except when it’s not
In any event, I’m still making a healthy profit: the last 3000 shares I bought were at about $31.
Of course, if AMD meets expectations, their 2022 server sales (in dollars) should about double their 2021 numbers.
As for the top-end mobile product comparisons, the portable gaming system market. Systems that can burn your lap if you game with them actually in your lap instead of on a table (or an airplane fold-out tray.) I expect the AMD side to be mostly systems with separate graphics and socketed CPUs. (The same will probably be true for Intel systems.) My son has a desktop system where we bought a socketed 5700G APU. The plan was to wait for GPU prices to come down. I just spent over $500 for an AMD 6600 GPU so he didn’t buy a 6500XT just to replace it in a few months. So he has a lot of experience trying to play AAA games on an APU. He has 32 Gig of DDR4 3600 memory, which was still the bottleneck.
Will these new chips be a decent upgrade for AAA games? Sure–but my son has a 4k monitor, so 1080P doesn’t really cut it. I have about the equivalent to his new card (a Vega 64) and some games are best at 2160P, but 1080P on a 4K monitor is ugly.
Do I regret spending $399 for the Vega 64? Of course not. It is about 1/2 the speed of today’s top-end cards, and I have gotten several years out of it so far. I tend to buy new top-end, or next-to-top-end cards when they give me a 4X improvement. AMD may be about to get there with their dual-chip “chiplet” GPUs. But I’ll wait until prices drop a bit, thank you.
