AMD Achieves First TSMC N2 Product Silicon Milestone

[btarunr]

AMD today announced its next-generation AMD EPYC processor, codenamed "Venice," is the first HPC product in the industry to be taped out and brought up on the TSMC advanced 2 nm (N2) process technology. This highlights the strength of AMD and TSMC semiconductor manufacturing partnership to co-optimize new design architectures with leading-edge process technology. It also marks a major step forward in the execution of the AMD data center CPU roadmap, with "Venice" on track to launch next year. AMD also announced the successful bring up and validation of its 5th Gen AMD EPYC CPU products at TSMC's new fabrication facility in Arizona, underscoring its commitment to U.S. manufacturing.

"TSMC has been a key partner for many years and our deep collaboration with their R&D and manufacturing teams has enabled AMD to consistently deliver leadership products that push the limits of high-performance computing," said Dr. Lisa Su, chair and CEO, AMD. "Being a lead HPC customer for TSMC's N2 process and for TSMC Arizona Fab 21 are great examples of how we are working closely together to drive innovation and deliver the advanced technologies that will power the future of computing."



"We are proud to have AMD be a lead HPC customer for our advanced 2 nm (N2) process technology and TSMC Arizona fab," said TSMC Chairman and CEO Dr. C.C. Wei. "By working together, we are driving significant technology scaling resulting in better performance, power efficiency and yields for high-performance silicon. We look forward to continuing to work closely with AMD to enable the next era of computing."

View at TechPowerUp Main Site

 

[john_]

Happy to see them advancing fast to 2nm. It shows that they don't take chances and I mean, if they where "limiting" themselves at 3nm for cost reasons, they could fell behind if Intel was coming out with a good 18A manufacturing node.

 

[alwayssts]

Happy to see them advancing fast to 2nm. It shows that they don't take chances and I mean, if they where "limiting" themselves at 3nm for cost reasons, they could fell behind if Intel was coming out with a good 18A manufacturing node.

On one hand, this isn't incredibly dissimilar to the fact Turin was on 3nm, but I get what you're saying. If consumer products actually follow this trend, it would be a big deal signifying exactly that.
On server it makes much more sense regardless as the margins are huge and clients want as many cores as possible on a socket. Moving from 12->16 core CCDs makes sense, even if at a premium.

If they use 2nm/2026 all-around, and A16/2028, I surely hope TSMC get their next node in production fairly quickly (if not the outside chance of Samsung's '1nm') bc they'll need something new come 2030.
Which, DGMW, appears to be very much potentially possible given what we've heard recently. It just wasn't always known that would likely be the case; TSMC was (until now) pretty cagey about what comes post-A16.

This could potentially shorten their runway and lead to higher costs, but if competition (from Intel designs and/or processes) demands it, that's not necessarily an overall bad thing for consumers.
It could lead to lower prices on better tech faster long-term, especially if TSMC ramps production (especially in more fabs) more quickly on newer nodes.

It's like it was only yesterday (and truly was only a few months ago iiirc) TSMC very-much did not want to produce 2nm in the US (until much later)...yet here we are (apparently).

 

[kondamin]

zen6 on epyc before it reaches ryzen?
that’s risky business

 

[billeman]

zen6 on epyc before it reaches ryzen?
that’s risky business

I think it's more a matter of money ; lower initial yields are no problem on a high-income product like server cpu's.

 

[HD64G]

I think it's more a matter of money ; lower initial yields are no problem on a high-income product like server cpu's.

Exactly! More risk for more profit was, is and always will be the financial name of the game in every market.

 

[kondamin]

I think it's more a matter of money ; lower initial yields are no problem on a high-income product like server cpu's.

the risk is pretty high though.
Just imagine there being serious issues with them and losing trust in that market space.

 

[Endymio]

the risk is pretty high though.
Just imagine there being serious issues with them and losing trust in that market space.

There's no direct correlation between yields -- whether low or high -- and the reliability of the finished chip.

 

[Bomby569]

Happy to see them advancing fast to 2nm. It shows that they don't take chances and I mean, if they where "limiting" themselves at 3nm for cost reasons, they could fell behind if Intel was coming out with a good 18A manufacturing node.

it will also mean they will be hard to come buy, like gpus

 

[N3utro]

Athlon64 3000+ : really?

 

[Squared]

Athlon64 3000+ : really?

View attachment 395145

That's funny. TPU has a listing on it too.

AMD Athlon 64 3000+ Specs

Venice, 1 Cores, 1 Threads, 1800 MHz, 67 W

www.techpowerup.com

 

[N3utro]

That's funny. TPU has a listing on it too.

AMD Athlon 64 3000+ Specs

Venice, 1 Cores, 1 Threads, 1800 MHz, 67 W

www.techpowerup.com

Of course they do! The A64 3000+ was one of the best gaming processor... 20 years ago

 

[Endymio]

That's funny. TPU has a listing on it too.

AMD is recycling code names from 20 years ago:

AMD Athlon 64 3000+ (Venice) - Hardware Museum

Athlon 64 3000+ (Venice) - code name Venice - is 1C/1T processor, runs at 1800 MHz. Using 90 nm technology, compatible with Socket 939. Introduced in 2005.

hw-museum.cz

 

[john_]

the risk is pretty high though.
Just imagine there being serious issues with them and losing trust in that market space.

You are right here, a design flaw creating instability for example in certain scenarios, would be a small disaster. And I say small because as with Nvidia in GPUs, AMD enjoys a somewhat huge advantage today in server CPUs, meaning if they f somehow Zen6, they will probably get away with it if they manage to come up with a fixed revision quickly. Especially if Intel keeps failing in competing. Look at Nvidia. Blackwell full of problems and that stops no one from buying, either someone building a gaming PC, or an AI super computer, everyone rushes to buy, no questions asked.

 

[ThomasK]

Intel says its 18A node is comparable. Lets wait and see.

 

[Tomorrow]

Happy to see them advancing fast to 2nm. It shows that they don't take chances and I mean, if they where "limiting" themselves at 3nm for cost reasons, they could fell behind if Intel was coming out with a good 18A manufacturing node.

That's a pretty big "IF". TSMC, unlike Intel has a proven track record of multiple successful nodes. Even their less successful ones like initial 3nm still outperformed Intel. So much so that Intel itself used 3nm on Arrow Lake.

zen6 on epyc before it reaches ryzen?
that’s risky business

the risk is pretty high though.
Just imagine there being serious issues with them and losing trust in that market space.

What risk? Server chips are always given the newest node these days. What serious issues? If anyone should be worried about "serious issues" and"losing trust" it's Intel.

Intel says its 18A node is comparable. Lets wait and see.

Intel says a lot of thing in their PowerPoint slides. Trouble is that most of the said things either do not materialize or come out with much limited scope and underwhelming performance. For the sake of competition i hope they get this right.

 

[john_]

That's a pretty big "IF". TSMC, unlike Intel has a proven track record of multiple successful nodes. Even their less successful ones like initial 3nm still outperformed Intel. So much so that Intel itself used 3nm on Arrow Lake.

It seems that the US government is using it's power to make TSMC to.... willingly help Intel to make it's 18A manufacturing viable. I am expecting Intel to be producing ALMOST everything on 18A next year, except those chips already designed for TSMC's 3nm and by 2027 EVERYTHING in it's fabs.

 

[jpvalverde85]

Athlon64 3000+ : really?

View attachment 395145

That CPU was a king of the budget gaming PC of that era, and a pretty badass overclocker too.

 

[Baba]

This story mixes up two pieces of news. 5th gen is being made in Arizona and 6th gen in Taiwan. Which 5th gen though? Turin or Turin dense? Arizona was supposed to be 5nm line. Turin is on N4X which is optimized N5. Or is this N3E Turin dense which is a new node?

 

[kondamin]

What risk? Server chips are always given the newest node these days. What serious issues? If anyone should be worried about "serious issues" and"losing trust" it's Intel.

Intel lost it's position already and they have little left to lose after a 8 horrible years starting with specter.

 

[alwayssts]

Of course they do! The A64 3000+ was one of the best gaming processor... 20 years ago

I didn't bite on that one (I knew it existed, though). I went from Socket A to 939; 754 never appealed to me as I was still impressed with how well you could overclock the mobile Socket A chips (like Barton).

I think I was waiting for something like that again (mobile T-birds/Breds/Barton), and we did get that with Denmark imo; the 939 Opteron X2's. I also think dual-channel was actually a worth-while upgrade.
I had a 165 and 170. Not in that order; the former because I released the magic smoke on the latter by cracking it with too much pressure on a WB mount.
Only time I ever did that, including many-a vapochill mounts. Slightly embarrassing, but it was 20 years ago.

Good times, those Opterons. They should do that again, even if a small run. If the new Opterons are 16 ('c') and the general consumer parts 12 ('non-c'), it would be fun to have that option to play around with.
Even if they're clocked low because of more cores (and general server usage) it could bring some fun back to overclocking to see what they could really do with more power and a decent cooling setup.

It seems that the US government is using it's power to make TSMC to.... willingly help Intel to make it's 18A manufacturing viable. I am expecting Intel to be producing ALMOST everything on 18A next year, except those chips already designed for TSMC's 3nm and by 2027 EVERYTHING in it's fabs.

Except by some (all?) accounts the high-end Nova is also TSMC 2nm, like AL 3nm (instead of 20A).
The lower-end/mobile parts (Panther and perhaps later?) will supposedly be 18A. At least that's what some are saying. There is the idea of 14A looming, but that appears anything but a certainty by then.
That should tell you something.

 

[john_]

Except by some (all?) accounts the high-end Nova is also TSMC 2nm, like AL 3nm (instead of 20A).
The lower-end/mobile parts (Panther and perhaps later?) will supposedly be 18A. At least that's what some are saying. There is the idea of 14A looming, but that appears anything but a certainty by then.
That should tell you something.

If they keep relying on TSMC by 2027, it will be saying to me that I am way too optimistic about Intel's future.

 

[alwayssts]

If they keep relying on TSMC by 2027, it will be saying to me that I am way too optimistic about Intel's future.

IMO you are too optimistic wrt Intel's future; I'm curious about the rollout of 14A more than 18A, which imo (unlike others) doesn't inspire a ton of confidence vs N2X/A16 when looking at product decisions.
That said, that doesn't mean Nova will be a bad arch. It might be very good, especially with DDR6. But it also might be very expensive.

I don't know who will win, but by all accounts AMD appears to be making the right moves to stay competitive on a less-expensive platform. To me, that's what they do best and will continue to do it, which is good.
Even if there are some advantages to using Intel (which may or may not end up being the case), it would appear AMD, at the very least, will not give up the position they've held for decades (of better value).

TBH, I'm more curious of seeing AMD's 12/16-core versus nVIDIA's ARM design (10+10?). That uncertainty, but conceivable excitement, of WoA using nVIDIA, looms large (and if nVIDIA can develop that ecosystem).
From there we may see Sound Wave evolve into an actual competitor (in later gens). That may actually be the future for many people, although I'm not pretending it will be tomorrow.

It's one of those things that in theory both Intel and nVIDIA might fight with slightly better PPC through more cores/parallelism, and AMD through higher clocks...and I'm curious how people respond to that.
Intel may also fight with high clocks, but nVIDIA likely efficient ones (as they do with GPUs), and it will be interesting to see how that all shakes out, especially wrt to how nVIDIA performs on WoA.

I don't think Intel is even thinking about the ARM market seriously yet, and they should. AMD appears to be playing it safe, and may develop it further as nVIDIA establishes it, which at least shows initiative.