Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Vattila · Oct 6, 2019

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

jamescox · Jun 16, 2022

Schmide said:
3d cache is multi-layer chip stacking . When integrated on package it constitutes one chiplet.

So you are disagreeing with yourself? What are you referring to with “a custom chiplet that contains only cache” not existing? I generally consider “chiplets” to be things used in stacking. Most of the devices that AMD has made just use chips in an MCM, not really chiplets.

Please explain what you mean by this:

“The point that I'm making is there is little room for any custom chiplet that contains only cache. (which I think is what you are proposing).”

We obviously do have custom chiplets with only cache, so what type of chiplets are you referring to? Are you just saying “not stacked chiplets of only SRAM cache” won’t exist? That is kind of obvious. The HBM style interface is quite narrow compared to the level of connectivity used for stacked cache, so they are unlikely to connect a stack of SRAM chiplets via silicon bridges like what would be used for HBM.

Schmide · Jun 16, 2022

jamescox said:
So you are disagreeing with yourself? What are you referring to with “a custom chiplet that contains only cache” not existing? I generally consider “chiplets” to be things used in stacking. Most of the devices that AMD has made just use chips in an MCM, not really chiplets.

Please explain what you mean by this:

“The point that I'm making is there is little room for any custom chiplet that contains only cache. (which I think is what you are proposing).”

We obviously do have custom chiplets with only cache, so what type of chiplets are you referring to? Are you just saying “not stacked chiplets of only SRAM cache” won’t exist? That is kind of obvious. The HBM style interface is quite narrow compared to the level of connectivity used for stacked cache, so they are unlikely to connect a stack of SRAM chiplets via silicon bridges like what would be used for HBM.

Whatever.

The "AMD chiplet design" is an interposer connecting an IO die with 1 or more chiplets. If you want to include any sublayer of the chip and call that a chiplet, I guess. But there are many layers of varying materials that make up this unit many refer to as a chiplet.

Even if my point was strained because of improper usage of the term chiplet. It should be easy to understand. Anything outside what I call a chiplet that has to go over infinity fabric really doesn't fit in the hierarchy of latency and bandwidth.

moinmoin · Jun 16, 2022

Ajay said:
Thanks was a link from Tomshardware.com. Articles is talking about several upcoming processors but the links and pics don't seem to match their descriptions. Or, I'm just dumb today.

Just for the record 7320U was mentioned in the Zen 2 APUs thread where Mendocino belongs.

Page 37 - Zen 2 APUs/"Renoir" discussion thread

Page 37 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

forums.anandtech.com

There I also complain that giving a Zen 2 chip the 7000 series model number is messed up. Thanks for showing exactly why.

maddie · Jun 16, 2022

Schmide said:
Whatever.

The "AMD chiplet design" is an interposer connecting an IO die with 1 or more chiplets. If you want to include any sublayer of the chip and call that a chiplet, I guess. But there are many layers of varying materials that make up this unit many refer to as a chiplet.

Even if my point was strained because of improper usage of the term chiplet. It should be easy to understand. Anything outside what I call a chiplet that has to go over infinity fabric really doesn't fit in the hierarchy of latency and bandwidth.

The "AMD chiplet design" is an interposer connecting an IO die with 1 or more chiplets.

I think "One "AMD chiplet design" is an interposer connecting an IO die with 1 or more chiplets." might be accurate. Anything fabbed as a unique part and then "glued" together with others can be called a chiplet.

MadRat · Jun 16, 2022

Schmide

Nothing stops the chiplet designer from using cache in ways for which HBM is unsuited. Technology has came a long ways since 1995. Designers squeeze out every nanosecond they can to decrease response and transmission times. The thing with cache today is that whatever interface is used, it typically cycles in smart ways through multiple banks to hide the pauses for the cache banks caught in a refresh cycle. This keeps the bus full of data and all those banks refreshes are invisible to the end user. And the buses keep multiplying, leaving more paths for transmissions. They have to balance the number, length, and width of the bus interfaces to meet their latency needs. So calling anything 'just cache' means something different today than in 1995. Back then you felt those refresh cycles. Today we can't feel the difference in cache speeds in most cases, but we do feel the lack of cache in operations that benefit from it.

eek2121 · Jun 17, 2022

Schmide said:
HBM is on package. You could say the same thing for apple's M1 memory. The point that I'm making is there is little room for any custom chiplet that contains only cache. (which I think is what you are proposing).

So on a processor you typically

latency

L1 < 4ns
L2 < 20ns
L3 < 40ns
mem < 80ns

and bandwidth

L1 1-4 TB/s
L2 0.5-2 TB/s
L3 300-600 GB/s
Mem 30-128 GB/s

Zen3 ddr4 ~45GB/s
HBM 8 stack ~128GB/s (give or take a bunch based on generation and width)

infinity fabric is provisioned to roughly match the memory speed.

The general point being, once you go to infinity fabric you're at memory speeds and latency.

Yes, I saw the replies. I wish I had what you are smoking right now, especially with how this week has gone.

Stop trolling these forums. Take a step back, and learn. Your account is 2 decades old!

Cache numbers are way off, as well as your "Zen3 ddr4" speeds. Maybe at stock JEDEC speeds you'll see those numbers, but not anything beyond, and those cache numbers are Intel fan fantasy level stuff.

Apple can have all the bandwidth they want if nothing utilizes it. The 5800X3D is the king of memory bandwidth and the number of workloads that it excels in are limited to a very few select apps. This isn't AMD's fault. I have an M1 Max Macbook Pro and guess what? It isn't any faster in any of my workloads. My workloads are 80% software development, 15% content creation (photoshop, video creation/encoding, a bit of blender, etc.) and 5% hardware design (specifically, circuit board design applications). (Note: my PC that has a 5950x and a 3090 runs workloads that are 40% gaming, 40% software development, and everything else is similar, though blender dominates the last 15%) The memory bandwidth would normally be awesome for gaming, except once again, GUESS WHAT?!?! Most games don't run on a modern Mac, and even if they did, the GPU is 3050-3060 territory at best. I estimate the real world performance impact of my current Mac vs. my previous (last gen) Mac to be < 5%, and as I've stated before, I've had the added joy of encountering Apple M1 specific bugs.

Sure, theoretical performance SHOULD be better on the current Apple stuff, but theoretically, a lot of stuff should be true.

Schmide · Jun 17, 2022

eek2121 said:
Yes, I saw the replies. I wish I had what you are smoking right now, especially with how this week has gone.

Stop trolling these forums. Take a step back, and learn. Your account is 2 decades old!

Cache numbers are way off, as well as your "Zen3 ddr4"

I'm not trolling. Relaxing numbers to encompass a larger sample of processors is not a bad thing.

You had a bad week? Sorry that my imperfections triggered you. That's on you.

eek2121 said:
speeds. Maybe at stock JEDEC speeds you'll see those numbers, but not anything beyond, and those cache numbers are Intel fan fantasy level stuff.

Apple can have all the bandwidth they want if nothing utilizes it. The 5800X3D is the king of memory bandwidth and the number of workloads that it excels in are limited to a very few select apps. This isn't AMD's fault. I have an M1 Max Macbook Pro and guess what? It isn't any faster in any of my workloads. My workloads are 80% software development, 15% content creation (photoshop, video creation/encoding, a bit of blender, etc.) and 5% hardware design (specifically, circuit board design applications). (Note: my PC that has a 5950x and a 3090 runs workloads that are 40% gaming, 40% software development, and everything else is similar, though blender dominates the last 15%) The memory bandwidth would normally be awesome for gaming, except once again, GUESS WHAT?!?! Most games don't run on a modern Mac, and even if they did, the GPU is 3050-3060 territory at best. I estimate the real world performance impact of my current Mac vs. my previous (last gen) Mac to be < 5%, and as I've stated before, I've had the added joy of encountering Apple M1 specific bugs.

Sure, theoretical performance SHOULD be better on the current Apple stuff, but theoretically, a lot of stuff should be true.

What's with this Mac dogma?

I'm not a mac guy but I've read about them and a lot of their power and performance seems to come from very fast L2 and very local memory. I was only using them as an example of large caches and on package memory.

None of this is my argument. I threw out some ballpark figures to basically argue against a cache only chiplet not taking into account that the stacked ram on top of the L3 could be considered a chiplet. Sorry I got the vernacular wrong.

Since we're here. Examples of the LLC (last level cache) : the xbox one and broadwell. Unique performance yet often outclassed by wider or newer ram. The ESRAM of the xbox was up against the wider GDDR5 of the PS4 while broadwell eDRAM well it just faded away.

DrMrLordX · Jun 17, 2022

Stacked L3 ala Vermeer-X is not exactly the same thing as a separate chiplet connected via IF; in fact, AMD went to a great deal of trouble to avoid those kinds of headaches.

nicalandia · Jun 17, 2022

https://twitter.com/x/status/1537756160194912256

gdansk · Jun 17, 2022

nicalandia said:
https://twitter.com/x/status/1537756160194912256

I suppose that lines up for AMD quite well. Is anyone else using TSMC chip stacking in volume products other than AMD and Xilinx? I assume yes but wonder who.

nicalandia · Jun 17, 2022

gdansk said:
I suppose that lines up for AMD quite well. Is anyone else using TSMC chip stacking in volume products other than AMD and Xilinx? I assume yes but wonder who.

On that note. Here are a few shots Milan-X

nicalandia · Jun 17, 2022

September 15th Release date?

https://twitter.com/x/status/1537824618613833728

lobz · Jun 17, 2022

eek2121 said:
Yes, I saw the replies. I wish I had what you are smoking right now, especially with how this week has gone.

Stop trolling these forums. Take a step back, and learn. Your account is 2 decades old!

Cache numbers are way off, as well as your "Zen3 ddr4" speeds. Maybe at stock JEDEC speeds you'll see those numbers, but not anything beyond, and those cache numbers are Intel fan fantasy level stuff.

Apple can have all the bandwidth they want if nothing utilizes it. The 5800X3D is the king of memory bandwidth and the number of workloads that it excels in are limited to a very few select apps. This isn't AMD's fault. I have an M1 Max Macbook Pro and guess what? It isn't any faster in any of my workloads. My workloads are 80% software development, 15% content creation (photoshop, video creation/encoding, a bit of blender, etc.) and 5% hardware design (specifically, circuit board design applications). (Note: my PC that has a 5950x and a 3090 runs workloads that are 40% gaming, 40% software development, and everything else is similar, though blender dominates the last 15%) The memory bandwidth would normally be awesome for gaming, except once again, GUESS WHAT?!?! Most games don't run on a modern Mac, and even if they did, the GPU is 3050-3060 territory at best. I estimate the real world performance impact of my current Mac vs. my previous (last gen) Mac to be < 5%, and as I've stated before, I've had the added joy of encountering Apple M1 specific bugs.

Sure, theoretical performance SHOULD be better on the current Apple stuff, but theoretically, a lot of stuff should be true.

Can you chill out a bit though?

Ranulf · Jun 17, 2022

Wccftech article:

AMD Ryzen 7000 Desktop CPUs & AM5 Motherboards To Launch on 15th September, Market Availability Confirmed

AMD seems to have confirmed the launch & general availability of its next-gen Ryzen 7000 Desktop CPUs and AM5 motherboard for 9th September.

wccftech.com

rtxtwt · Jun 17, 2022

7950X 7900X 7800X 7600X, no low end at launch, absolutely the same naming scheme as Zen3. What the F, I'm surprised. Wondering the price.........

jpiniero · Jun 17, 2022

cortexa99 said:
7950X 7900X 7800X 7600X, no low end at launch, absolutely the same naming scheme as Zen3. What the F, I'm surprised. Wondering the price.........

Given DDR5 prices, low end AM5 products don't make sense. Maybe you will see IGP-busted Rembrandt at some point.

Asterox · Jun 17, 2022

cortexa99 said:
7950X 7900X 7800X 7600X, no low end at launch, absolutely the same naming scheme as Zen3. What the F, I'm surprised. Wondering the price.........

Red is not low end, or by price any AM5 cheepest mobo+CPU combo is not even close to low end.

AM4/R5 5600 is not low end.You can get even more cheeper, but good CPU-s as R5 5500 in combo with older AM4 motherboards.Even R5 5500 is not low end, or blah you can say that cheeper old Zen2/R5 4500 is low end.But what about 4/8 R3 4100 what is this CPU, if R5 4500 is low end.

nicalandia · Jun 17, 2022

cortexa99 said:
7950X 7900X 7800X 7600X, no low end at launch, absolutely the same naming scheme as Zen3. What the F, I'm surprised. Wondering the price.........

DDR5 and Low end? You must be joking.....

Amd 5700X and 5700G are very capable products with 8 Cores and very good gaming performance. and 5600 non-x for the lower end of the budget

rtxtwt · Jun 17, 2022

sorry, I'm not clear enough, what shocking me is if the naming scheme and core count is the same, means AMD don't care about the competition of big+little. That's why I have to wonder the price..... IF 7600x(6C12T) cheaper than 12600k(6P4E) could it be called low end??? if it's same price or more expensive it won't make sense unless the 6C12T performance could be stand well against 6P4E.........maybe AMD is too confident?? OK, I have to take a rest, my brain is overload

nicalandia · Jun 17, 2022

cortexa99 said:
sorry, I'm not clear enough, what shocking me is if the naming scheme and core count is the same, means AMD don't care about the competition of big+little. That's why I have to wonder the price..... IF 7600x(6C12T) cheaper than 12600k(6P4E) could it be called low end??? if it's same price or more expensive it won't make sense unless the 6C12T performance could be stand well against 6P4E.........maybe AMD is too confident?? OK, I have to take a rest, my brain is overload

No doubt in my mind that the 7600X will beat the 12600K at gaming and in MT performance.

Mopetar · Jun 17, 2022

$300 for the 7600X is the most likely price. If we assume a 10% clock speed bump over over the 5600X and a 5% improvement in IPC, it would be about even with the $300 12600K in terms of performance and both of those are pessimistic estimates.

If we use TPUs figures and a 13% clock speed bump and an 8% improvement in IPC, that hypothetical 7600X is 5% ahead of the 12600K in application performance and 11% ahead in gaming when comparing at 720p.

I'm not sure if there's a good reason why AMD should charge less, particularly when it seems likely that inflation is going to erode a lot of any price over the next year or so. They might even price over $300 just to compensate.

nicalandia · Jun 17, 2022

cortexa99 said:
Classic CPU-monkey start creating data of 7950X into their lists

They actually did a good extrapolation of the performance with the data provided by AMD

AMD Ryzen 9 7950X - Cinebench R23 Benchmark projection

AMD Ryzen 9 7950X - Cinebench R23 Benchmark projection

We give you a preview of AMD's new Zen 4 desktop processors and calculate the performance of the AMD Ryzen 9 7950X top model for the new AM5 socket. The performance data and other technical details come directly from the press kit that AMD released as part of its Financial Analyst Day on June 9...

www.cpu-monkey.com

Ranulf · Jun 17, 2022

Mopetar said:
$300 for the 7600X is the most likely price. If we assume a 10% clock speed bump over over the 5600X and a 5% improvement in IPC, it would be about even with the $300 12600K in terms of performance and both of those are pessimistic estimates.

If we use TPUs figures and a 13% clock speed bump and an 8% improvement in IPC, that hypothetical 7600X is 5% ahead of the 12600K in application performance and 11% ahead in gaming when comparing at 720p.

I'm not sure if there's a good reason why AMD should charge less, particularly when it seems likely that inflation is going to erode a lot of any price over the next year or so. They might even price over $300 just to compensate.

I think for either AMD or Intel the bigger problem is how does a customer justify paying $300 for a 6core these days compared to the 12400 or 5600x's current prices at or below $200.

nicalandia · Jun 17, 2022

Ranulf said:
I think for either AMD or Intel the bigger problem is how does a customer justify paying $300 for a 6core these days compared to the 12400 or 5600x's current prices at or below $200.

Budget minded customers don't purchase bleeding edge tech...

moinmoin · Jun 17, 2022

Mopetar said:
$300 for the 7600X is the most likely price.

Agreed. If AMD keeps the model range and numbering like it currently looks like I expect the MSRPs to be kept as well.

Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

Senior member

Senior member

Diamond Member

Diamond Member

Diamond Member

Lifer

Platinum Member

Diamond Member

Lifer

Diamond Member

Platinum Member

Diamond Member

Diamond Member

Platinum Member

Platinum Member

Senior member

Lifer

Golden Member

Diamond Member

Senior member

Diamond Member

Diamond Member

Diamond Member

Platinum Member

Diamond Member

Diamond Member