A report citing semiconductor industry sources indicates that TSMC reportedly has difficulty with its 3nm process yields. Taiwan’s DigiTimes says that if the 3nm yield problem continues, many customers might extend their use of the 5nm process node. In addition, TSMC’s struggles could impact the product roadmaps of the PC world’s favorite names like AMD and Nvidia.
It’s essential to take the report with a pinch of salt. The people in the know might be correct, but TSMC hasn’t publicly admitted any N3 delays so far. On the contrary, it has asserted that it is “on track with good progress.”
The source report’s critical rumor is pinned upon TSMC, finding it very difficult to achieve satisfactory yields with its 3nm FinFET processes. It explains that TSMC has “continuously revised” its 3nm offerings, and the foundry is seemingly doing this to find a sweet spot for yields (the percentage of chips that are not faulty). The latest TSMC introduction is N3E, a lower-cost version of TSMC’s 3nm manufacturing process, which surprised industry watchers by arriving a year after N3. TSMC also makes N3B processors for some customers, depending upon design and cost constraints. Despite TSMC’s process wrangling, outlined above, and “constant revision,” the insiders say yields continue to remain lower than expected.
Due to the 3nm family issues, some TSMC’s customers are looking at rejigging plans, which means changing their roadmaps. In addition, customers like Apple and Intel have paid a lot to secure N3 process chips in the coming months. Other partners like AMD must not have felt the urgency or need for such lavish pre-payments so that they will feel the most substantial effect of TSMC’s yield issues.
The latest TSMC N3 schedule I heard was production starts late 2022, resulting in first products available in early 2023. I don’t see TSMC N3 as that interesting as it is an incremental improvement over the currently shipping N4. The TSMC N2 is where they move to GAA (gate all around) which should be a substantial improvement in performance/watt.
Alan
Alan said: The latest TSMC N3 schedule I heard was production starts late 2022, resulting in first products available in early 2023. I don’t see TSMC N3 as that interesting as it is an incremental improvement over the currently shipping N4. The TSMC N2 is where they move to GAA (gate all around) which should be a substantial improvement in performance/watt.
Technically N4 is a derivative of the N5 (5 nm) node. AMD using N4 for Zen 5 and RDNA 4 would be a safe bet. Otherwise, AMD needs to start moving Zen 5 to the N3 design rules. AMD could do both, but while doing both for Zen 5 chiplets is no big deal, there are also APUs and GPUs that are going to require much more design/porting effort.
My guess? AMD will go with N4 for Zen 5 and save N3 for Zen 6. (AMD has not said anything publically about Zen 6, but a third CPU family for AM5 makes a lot of sense.) I guess a Zen 4+ might make sense*–just like Zen 3+, a plan to deal with a delay in N3. Zen 3+, which is actually shipping now in laptops was to build in flexibility in case DDR5 was ready early. DDR5 is here, but at the moment, it is more expensive than DDR4 for equivalent performance.
AMD could offer Zen 4c (Bergamo) cores say with 16 single-threaded cores per chiplet, and 8, 12, 16, 24, and 32 core desktop products. They could, but I have no clue as to whether Zen 4c will ever show up in desktop products.
* AMD could offer Zen 4c (Bergamo) cores say with 16 single-threaded cores per chiplet, and 8, 12, 16, 24, and 32 core desktop products. They could, but I have no clue as to whether Zen 4c will ever show up in desktop products.
Was thinking myself that a 32 core desktop would be a pretty perfect response to Alder lake, send Intel back to the drawing board and force Pat Gelsinger “Intel CEO says AMD is in the rearview mirror and ‘never again will they be in the windshield’” into saying “Never say never again”.
a 32 core desktop would be a pretty perfect response to Alder lake, send Intel back to the drawing board
Last week intel did a demo of the Raptor lake upgrade to Alder lake. It has 16 E-cores and 8 p-cores for a total of 32 threads. Expected performance improvement over alder lake is 15% in single threaded applications, and 40% in multithreaded. Due to be released in the September/October 2022 time frame it will go into the same motherboards as Alder Lake.
A 32 core zen 4 built on TSMC N5 will almost certainly outperform Raptor lake. However, I suspect the timing is such that the AMD N5 client products are going to be competing with the Intel 4 meteor lake products.
There is also a leak of a Sapphire Rapids -X HEDT part that will release in Q3 2022. It will be a competitor to the 32 and 64 core threadripper parts.
FYI,
Alan
A 32 core zen 4 built on TSMC N5 will almost certainly outperform Raptor lake. However, I suspect the timing is such that the AMD N5 client products are going to be competing with the Intel 4 meteor lake products.
There is also a leak of a Sapphire Rapids -X HEDT part that will release in Q3 2022. It will be a competitor to the 32 and 64 core threadripper parts.
So what you’re telling me is that this is becoming a knife fight in an elevator.
A 32 core zen 4 built on TSMC N5 will almost certainly outperform Raptor lake. However, I suspect the timing is such that the AMD N5 client products are going to be competing with the Intel 4 meteor lake products.
Zen 4 will apparently be shipping early in 2H2022. There are indications that it may be announced in May or June. Zen 4c will be a 2023 part. So Zen 4c against Meteor Lake on the desktop makes sense–if Zen 4c does show up on the desktop. There has been some chatter about AMD using Zen 4c cores with Zen 4 to compete with Intel’s P and E core designs. However, I personally don’t see any reason to do that. AMD could manufacture a CPU with one Zen 4 chiplet and one Zen 4c chiplet, but why? Zen 4 will certainly follow Zen 3 in having one L3 cache per chiplet. Zen 4c will likely have more cores per chiplet, but still one L3 cache per chiplet. Sharing those two L3 caches would effectively blend the L3 cache size between the two chiplets. Offering a Zen 4 chiplet with 3d cache seems more the direction that AMD is going.
A meta-comment. I used to say that “I hate to wait,” when it came to computers. Since I got started in computing when machines were a hundred million times slower, I think I can assert that, while stupid programming can make things slow, the cases where I end up waiting today involve NP-complete or NP-hard problems. In theory, we will need quantum computers to speed solving those problems up significantly. More realistically, finding better algorithms is the way forward on those problems. For example, the best algorithm for factoring a huge integer is the Special Number Field Sieve, a subset of the more general Number Field Sieve. (To factor a large number with the SNFS, work is put into finding a field that will work well with that number.)
I’m sort of getting off track. My point above is that finding factors of huge numbers has been improved a lot during my lifetime. I don’t know that any quantum computer will ever have enough qubits to catch up with improvements in algorithms. And any problems that are not NP-complete or NP-hard, are not something that anyone sits around waiting on a computer for.
Same with games, and GPUs. You can still buy low-end GPUs, or CPUs with an integrated GPU that will not run some modern games at an acceptable rate–if you turn up the eye candy. High-end GPUs are almost to the point where frame rates with all the eye candy turned up, are well over 100 frames per second. (Right now, I wouldn’t recommend a Radeon 6500XT to anyone. Why? Because the 6600 is so much better. Does the 6600 cost a lot right now? If you are building a computer for gaming, the difference in cost of the GPUs is small compared to the whole system.
I used to say that “I hate to wait,” when it came to computers. Since I got started in computing when machines were a hundred million times slower, I think I can assert that, while stupid programming can make things slow, the cases where I end up waiting today involve NP-complete or NP-hard problems. In theory, we will need quantum computers to speed solving those problems up significantly. More realistically, finding better algorithms is the way forward on those problems. For example, the best algorithm for factoring a huge integer is the Special Number Field Sieve, a subset of the more general Number Field Sieve. (To factor a large number with the SNFS, work is put into finding a field that will work well with that number.)