AMD 5nm 96-core 'Genoa' Zen 4 CPU arriving in 2022, 128-core 'Bergamo' lands early 2023

Cal Jeffrey

Posts: 3,267   +894
Staff member
In a nutshell: Team Red revealed it has a couple of heavy-duty processors coming out over the next couple of years. One has 96 cores targeted toward high-end general-purpose workstations. The other is a 128-core monster for use in data centers running cloud-based applications. Both are based on TSMC's 5nm process.

During its Accelerated Data Center keynote on Monday, AMD unveiled the roadmap for its upcoming fourth-generation Epyc silicon. AMD chief Lisa Su notes the two HPC processors are aimed at enterprise and data center applications. Reuters notes that Meta (Facebook) just joined long-time AMD partners Microsoft (Azure), Amazon (AWS), and Alphabet (Google Cloud) and will be among the first companies to use the chips.

The new CPUs should offer twice the density and power efficiency while supplying 1.25 times or more performance.

The first chip is codenamed "Genoa." Designed for heavy-duty, general-purpose computing, Genoa features 96 Zen 4 cores thanks to TSMC's 5nm process and comes with DDR5 and PCIe 5.0 support. When announced in 2019, Genoa silicon was expected to land sometime this year but was delayed by the on-going global supply-chain issues. Su said Genoa should start arriving in 2022. She did not have a firmer launch date, but since AMD has already sent out samples, it should probably be sooner in the year than later.

The second chipset, codenamed "Bergamo," is designed with cloud-based data centers in mind. The silicon is still built on the same 5nm process, but AMD has optimized it for cloud-native computing. As such, it has designated the architecture as Zen 4c. At its essence, Zen 4c is compatible and functionally the same as its cousin but optimizes cache hierarchy to allow a higher core/thread density.

Bergamo packs a whopping 128 Zen 4c cores on a single CPU. Like Genoa, it will also support DDR5 and PCIe 5.0, CXL 1.1, RAS, and Infinity Guard security. It will also be socket compatible (SP5) with Genoa. However, it should offer "breakthrough performance per socket." Su said Bergamo silicon should start shipping in the first half of 2023.

Team Red also unveiled its third-generation Milan-X server CPUs, which will sport up to 768MB of L3 cache. We should see those hitting the market in Q1 2022.

Permalink to story.

 

Dimitriid

Posts: 1,530   +2,990
So they're going BIG-BIG or little-little? No BIG-little for the time being?

Well you they just do that anyway with a dual socket SP3 board though
 

Cycloid Torus

Posts: 4,845   +1,664
Lurch, bump, sway back, slump forward..... 96 cores...gulp..... 128 core .... uh.h.h..h...h

Please Mr. AAA dev, don't make a game which can use all this.
 

grvalderrama

Posts: 291   +134
So they're going BIG-BIG or little-little? No BIG-little for the time being?

Well you they just do that anyway with a dual socket SP3 board though

"The new roadmap covers the fourth-gen EYPC processors. The 96-core Genoa will come on the 5nm process in 2022, while the 128-core Bergamo, also on 5nm, will come to market in 2023. In addition, Bergamo comes with a new type of 'Zen 4c' core optimized for specific use cases, meaning that AMD's Zen 4 chips will come with two types of cores, with the 'c' cores obviously being the smaller variants".
Read this in Tom hardware.
 

Tom Yum

Posts: 138   +302
So they're going BIG-BIG or little-little? No BIG-little for the time being?

Well you they just do that anyway with a dual socket SP3 board though

AMD has less need for a big/little design as Zen3 seems to scale pretty well from a power perspective (Cezanne has it go down to around 1W/core, 5800X shows it going up to 15W/core). Tiger Lake struggles to scale down, which is why at 15W it struggles to beat Cezanne in single core but beats it when allowed to go to 28W. Golden Cove is effectively an enlarged Willow Cove, and it seems Intel needs to rely on architecture power tradeoffs (using the little cores) because they can't get Golden Cove to scale down.

Zen4 may start to change that, as it is expected to introduce AVX512 support which will increase the size of the core. Zen 4c will be that little (littler may be more accurate) core, basically a Zen 4 with smaller caches and denser design libraries (trading die size for clock speed). AMD have said that they want to stick with an ISA heterogeneous design, which means they can't remove AVX512 from Zen 4c to reduce the die size further.

The only issue for AMD in creating a big/little design would be it would require yet another die which costs money to design and manage. An all big Zen4, an all little Zen4c, and another die for a hybrid. AMD may assess that the juice isn't worth the squeeze from a cost/benefit perspective if Zen4 is able to scale down enough for the consumer market performance/watt targets. Intel's scale just means they can amortize the cost of a die tapeout much more easily than AMD, which is also why AMD stuck with a single 8 core Matisse design, even when it impacted their ability to target lower price points (as they haven't had sufficient defective dies to support lower bin cheaper parts). Whereas Intel has a smaller and cheaper die for <=6 core Alder Lake's for example.
 

Dimitriid

Posts: 1,530   +2,990
Quit using so much FAB usage on CPU's. There ARE people who would like a video card at retail prices.
While I sympathize with the sentiment, you've gotta think whenever you'd like to sell a 6900xt chip for close to 1000 USD (Rest of the price is probably AIB margins) or one of the new 96 core Genoa chips for probably USD 10,000 maybe even more.

Sucks to be a gamer right now but there's no way AMD would give up 10x more money just to make more consumer GPUs.
 

ypsylon

Posts: 436   +389
AMD believes that BIG is best. :D

Everything I hope for for new gen of Threadrippers 6/7000 whatever is USB4/TB support out of the box without any shenanigans and TR Pro being available at the same time as normal TR; not year later when it basically became irrelevant because everybody who was in the market got either 'normal' TR or EPYC year earlier.

I would love to see TR Pro units like some EPYCs where you have less cores (say 16) but triple or quadruple the cache on chip. Any extra MB of on-chip cache is pure gold pressed latinum.

I'm in for BIG.BIG. No fuss, no stupid hybrid design, no ridiculous power draw. Just works.
 

Yenega

Posts: 302   +206
This is massive. Lisa Su knows where the money is at and the future of computing.

Intel still sits at 95% or more of Enterprise- and Servermarkets but Nvidia is grapping tons of marketshare with it's enterprise GPUs. Alot of workloads are being done with GPUs instead of CPUs these days, they are simply much faster for some tasks.

AMD having PCIe 4.0 before Intel was a big reason why some went for AMD for servers, now Intel will have PCIe 5.0 long before AMD, this matters alot for server/enterprise, since bandwidth is a huge bottleneck.

Intel will do up to 56/112 with Sapphire Rapids and probably doubling these numbers with Next Gen on Intel 4 node, so AMD will need every single of these cores to compete that is for sure.
 
Last edited:

Foobario

Posts: 18   +23
Intel still sits at 95% or more of Enterprise- and Servermarkets but Nvidia is grapping tons of marketshare with it's enterprise GPUs. Alot of workloads are being done with GPUs instead of CPUs these days, they are simply much faster for some tasks.

AMD having PCIe 4.0 before Intel was a big reason why some went for AMD for servers, now Intel will have PCIe 5.0 long before AMD, this matters alot for server/enterprise, since bandwidth is a huge bottleneck.

While PCIE4 was an advantage, it was much less of a sales driver than the lower TCO and higher IPC that EPYC has had over anything Intel has put out up through Ice Lake.

AMD is sitting at 12% data center share. ARM has 3% so your estimate of Intel's current server share is way above reality. About 10 points above reality.
 
Last edited:

Kosmoz

Posts: 539   +968
So they're going BIG-BIG or little-little? No BIG-little for the time being?

Well you they just do that anyway with a dual socket SP3 board though
According to MLiD they are going big.Little with Zen5, but their version is to use the last gen core as the Little and the next gen core as the big one. In a very basic explanation.

 

Irata

Posts: 1,872   +3,166
So they're going BIG-BIG or little-little? No BIG-little for the time being?

Well you they just do that anyway with a dual socket SP3 board though
It seems like Zen 4C are the new ‚little‘ cores. Getting some technical details would be nice.

Either way, right now they are on medium.medium (compared to Golden Cove), so if AMD went big.little that would mean bigger regular Zen cores.

Intel still sits at 95% or more of Enterprise- and Servermarkets but Nvidia is grapping tons of marketshare with it's enterprise GPUs. Alot of workloads are being done with GPUs instead of CPUs these days, they are simply much faster for some tasks.

AMD having PCIe 4.0 before Intel was a big reason why some went for AMD for servers, now Intel will have PCIe 5.0 long before AMD, this matters alot for server/enterprise, since bandwidth is a huge bottleneck.

Intel will do up to 56/112 with Sapphire Rapids and probably doubling these numbers with Next Gen on Intel 4 node, so AMD will need every single of these cores to compete that is for sure.

Zen 4 Epyc does support PCIe 5 and it‘s already sampling. And in terms of SR, not sure if competing against 56/112 with 96/192 will be a problem, considering that we are talking about Zen 4 cores on TSMC 5nm. We‘ll see.
 

Yenega

Posts: 302   +206
It seems like Zen 4C are the new ‚little‘ cores. Getting some technical details would be nice.

Either way, right now they are on medium.medium (compared to Golden Cove), so if AMD went big.little that would mean bigger regular Zen cores.



Zen 4 Epyc does support PCIe 5 and it‘s already sampling. And in terms of SR, not sure if competing against 56/112 with 96/192 will be a problem, considering that we are talking about Zen 4 cores on TSMC 5nm. We‘ll see.

SR is releasing soon, based on Intel 7 node and will have 1 year head start on PCIe 5.0 - Next Gen is based on Intel 4 node and should double core count. So Intel will probably have 112/224 when AMD goes 96/192 or soon after.

AMD is not able to use 5nm before late 2022. Apple owns the proces and AMD does not have to money to get in.

First when Apple leaves a node or has enough stock (for the primary series, iPhone 13 = 5nm TSMC), others can get in. TSMC loves Apple and will always prioritize Apple. TSMC's best customer. Without Apple, TSMC would not be in the position they are today.
 
Last edited:

HardReset

Posts: 1,401   +1,034
SR is releasing soon, based on Intel 7 node and will have 1 year head start on PCIe 5.0 - Next Gen is based on Intel 4 node and should double core count. So Intel will probably have 112/224 when AMD goes 96/192 or soon after.
SR will have at most 4-6 months head start. And considering power consumption, SR will have huge disadvantage.

It's not only core count, it's also power consumption. And there AMD will have huge advantage, like it already has. No, Sapphire Rapids will not solve Intel's power consumption problem on servers.
AMD is not able to use 5nm before late 2022. Apple owns the proces and AMD does not have to money to get in.

First when Apple leaves a node or has enough stock (for the primary series, iPhone 13 = 5nm TSMC), others can get in. TSMC loves Apple and will always prioritize Apple. TSMC's best customer. Without Apple, TSMC would not be in the position they are today.
It's early 2022, not late. Currently Apple takes 72K wafers per month, but total output is already 120K wafers per month. And it will get bigger whole time.

To remind:

TSMC began 5nm risk production March 2019. That is, 2 years, 8 months ago). TSMC began 5nm mass production April 2020. That is, 1 year 7 months ago.
 

Yenega

Posts: 302   +206
SR will have at most 4-6 months head start. And considering power consumption, SR will have huge disadvantage.

It's not only core count, it's also power consumption. And there AMD will have huge advantage, like it already has. No, Sapphire Rapids will not solve Intel's power consumption problem on servers.

It's early 2022, not late. Currently Apple takes 72K wafers per month, but total output is already 120K wafers per month. And it will get bigger whole time.

To remind:

TSMC began 5nm risk production March 2019. That is, 2 years, 8 months ago). TSMC began 5nm mass production April 2020. That is, 1 year 7 months ago.
It's late 2022, you will see. No AMD chips will be made on 5nm before late 2022; Not enterprise, not consumer.

Nvidia will use it before AMD, for RTX 4000 series.

Stop your amd fanboy nonsense. Reality calls. AMD don't have the funds to use 5nm before late 2022 when Apple is done with it.

TSMC 5nm is used for Apple SoCs. Has nothing to do with AMD, so why bring risk production up haha. Apple has used 5nm since iPhone 12 and first M1 and still uses the proces for iPhone 13 and M1 Pro + Max. Common knowledge.

5nm is nothing new, but AMD can't afford to use it.

Intel even secured majority of TSMC 3nm proces too, together with Apple. Both can easily afford to use the best proces.

Nvidia can probably can too, but they don't need the best node to beat AMD, as we saw with Ampere.
 
Last edited:

dangh

Posts: 355   +537
It's late 2022, you will see. Nvidia will use it before AMD, for RTX 4000 series.

Stop your amd fanboy nonsense. Reality calls.

TSMC 5nm is used for Apple SoCs. Has nothing to do with AMD, so why bring risk production up haha. Apple has used 5nm since iPhone 12 and first M1.
Lisa saying it is q1 2022, MS already did the testing of those CPU's (https://www.tomshardware.com/news/microsoft-posts-in-depth-amd-epyc-milan-x-benchmarks) but you are saying 'believe me, I'm noone and I know better than CEO when cpu will be released.

Please do the reality check. You're out of sync with it.
 

Yenega

Posts: 302   +206
Lisa saying it is q1 2022, MS already did the testing of those CPU's (https://www.tomshardware.com/news/microsoft-posts-in-depth-amd-epyc-milan-x-benchmarks) but you are saying 'believe me, I'm noone and I know better than CEO when cpu will be released.

Please do the reality check. You're out of sync with it.

Where did she say Q1? Just says 2022 and Late 2022 is still 2022.

Yeah obviously CPUs are tested long before actual release, this is nothing new.

Nvidia is already testing Lovelace and Hopper.

AMD is already testing Zen 4 on AM5 with DDR5. What else is new. None of those products are close to release.
 

SixTymes

Posts: 173   +110
They sure must be loving that massive cash infusion they got from the Saudi's back in 2008, and again in 2009, and again in 2011. I sure wish I had play money then and bought the stock also when it was 2 dollars per share.