Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.

CrpyticComedian

macrumors newbie
Original poster
Nov 8, 2021
5
0
Amd just teased their new Zen 4 epyc cpu with 128 cores what are your guys thoughts on it
 

mi7chy

macrumors G4
Oct 24, 2014
10,628
11,300
Sure, if you can come up with an ingenious way to dissipate 320W of heat. 280W Threadripper already surpassed the ability to be air cooled and require liquid AIO cooling. And, it'll be plug in only.

Here's an idea.

 
Last edited:

jinnyman

macrumors 6502a
Sep 2, 2011
762
671
Lincolnshire, IL
While they are both a CPU, they are totally different product meant for totally different sectors.
4 x M1 Max (40 cores, 128 gpu cores) theoretically can consume 400W given a single Max uses 100W. Then perhaps.
 

leman

macrumors Core
Oct 14, 2008
19,530
19,709
Hey man you never know AMD’s engineering has come a long way

Just in case you are not jut being funny: how much do you want to pay for a laptop like that? I mean, that CPU alone is likely to be $10K at least, maybe more. Its power draw is estimated to be approximately 400W.

Anyway, why would you want a server chip inside a laptop? Who ever offered a product like that? Why 128-core EPYC? Why not the 32 or 64 core EPYC which are awailable today?
 

leman

macrumors Core
Oct 14, 2008
19,530
19,709
Hot new form factor for server class laptops; the roll-around Zero Halliburton carry-on case...!

bbe3628297e56da1177dc419fe296004--plasma-tv-om.jpg
 
  • Haha
Reactions: Boil

Spindel

macrumors 6502a
Oct 5, 2020
521
655
While they are both a CPU, they are totally different product meant for totally different sectors.
4 x M1 Max (40 cores, 128 gpu cores) theoretically can consume 400W given a single Max uses 100W. Then perhaps.
That’s 400 W for CPU+GPU, that is currently kind of hard to beat for the competition considering that the GPU part is very capable.
 

UBS28

macrumors 68030
Oct 2, 2012
2,893
2,340
The only problem is, is that a laptop with a 128 core Epyc cpu will be loud and drain the battery in a couple of minutes.

Best to put that CPU in a desktop / workstation imo, rather than in a laptop where it belongs.

I have always said it many years already, Apple should have switched to AMD instead of sticking with Intel.

We could have had 128-core Mac Pro’s basically and I am not sure if 4 x M1 Max would beat it.
 

throAU

macrumors G3
Feb 13, 2012
9,273
7,439
Perth, Western Australia
1. it's a ~280 watt server part
2. no - laptop users and apps do not need or can make use of 128 cores at the moment. they need faster rather than more
3. no - totally different use case.

core count on laptops MAY increase in the future but its going to be a very very long time before 128 cores are useful in one. servers on the other hand have many users connected (or more likely, have many virtual machines running on them) and can benefit from such core counts.
 

leman

macrumors Core
Oct 14, 2008
19,530
19,709
The only problem is, is that a laptop with a 128 core Epycc cpu will be loud and drain the battery in a couple of minutes.

That's the only problem you see here? :oops:

I have always said it many years already, Apple should have switched to AMD instead of sticking with Apple.

Why would you want them to switch to inferior tech?
 
  • Like
Reactions: JMacHack

throAU

macrumors G3
Feb 13, 2012
9,273
7,439
Perth, Western Australia
For the markets apple cater to with their pro machines (basically high end video editing is the most intensive stuff they do) an M1 Max will likely outperform a 128 core EPYC unless the EPYC has a hefty GPU accelerator on it AND the app is optimised to use it.

Do not under-estimate the video accelerators in the Apple M1 Pro/Max SOCs. They've basically got a built in afterburner card and afterburner on the Mac Pro is way faster than doing things in CPU or GPU.

These EPYC processors are not aimed at video specifically. They're general purpose VM hosting boxes - designed to run many many copies of Windows server, SQL server, Linux, etc.

Doing video on them is like trying to go race your truck.
 

jinnyman

macrumors 6502a
Sep 2, 2011
762
671
Lincolnshire, IL
That’s 400 W for CPU+GPU, that is currently kind of hard to beat for the competition considering that the GPU part is very capable.
Well of course. The combined CPU-GPU solution is incomparable to anything PC side is offering. I was mentioning that in terms of thermodynamics vs AMD 400w chip.
 

Adarna

Suspended
Jan 1, 2015
685
429
Using TSMC's chiplet tech it is possible to mask stitch 9-16 M1 Max dies together to reach 90-160 CPU cores together.
Form Factor (As of 9 Nov 2021)​
AMD Zen 4 Epyc 128 core cpu Rival​
AMD Zen 4 Epyc 128 core cpu Rival​
Mac Pro
iMac Pro​
Mac mini Pro
Mac Pro
iMac Pro​
MBP 14"
MBP 16"
Mac mini Pro
iMac 24"
iMac Pro​
300mm² Silicon Wafer​
Apple silicon chip​
M1 Max Jade-16C​
M1 Max Jade-9C​
M1 Max Jade-4C​
M1 Max Jade-2C​
M1 Max​
?​
Launch​
?​
?​
Q2 or Q4 2022​
Q2 or Q4 2022​
Q4 2021​
?​
# of dies​
16​
9​
4​
2​
1​
49​
CPU​
160​
90​
40​
20​
10​
490​
performance cores​
128​
72​
32​
16​
8​
392​
efficiency cores​
32​
18​
8​
4​
2​
98​
GPU core​
512​
288​
128​
64​
32​
1568​
Neural Engine core​
256​
144​
64​
32​
16​
784​
memory bandwidth​
6.4TB/s​
3.6TB/s​
1.6TB/s​
800GB/s​
400GB/s​
19.6TB/s​
Max Memory​
1024GB​
576GB​
256GB​
128GB​
64GB​
3,136GB​
Hardware-accelerated H.264, HEVC, ProRes, and ProRes RAW​
16​
9​
4​
2​
1​
49​
Video decode engines​
16​
9​
4​
2​
1​
49​
Video encode engines​
32​
18​
8​
4​
2​
98​
ProRes encode and decode engines​
32​
18​
8​
4​
2​
98​
Peak Quoted Transistor Densities using TMSC's 5nm (2020) at the same die size​
171.3 million transistors per mm²
912 Billion​
513 Billion​
228 Billion​
114 Billion​
57 Billion​
5.19 Trillion​
Estimated Die Size​
17.004cm²​
12.753cm²​
8.502cm²​
6.3765cm²​
4.251cm²​
30cm²​
Peak Quoted Transistor Densities using IBM's 2nm (2025) at the same die size​
333.33 million transistors per mm²
1.774 Trillion​
998.24 Billion​
443.66 Billion​
221.83 Billion​
110.92 Billion​
10.1 Trillion​
Estimated AMD Ryzen 9 5950X Performance​
16x​
9x​
4x​
2x​
1x​
49x​
Estimated RTX 3080 Performance​
16x​
9x​
4x​
2x​
1x​
49x​
 
Last edited:

altaic

Suspended
Jan 26, 2004
712
484
For what purpose would you have a server class cpu in a laptop?
Obviously to provide kubernetes* to your local coffee shop

* Service not guaranteed. Access is $9/cup. Harassing your barista may have criminal penalties depending on your jurisdiction.
 
  • Like
Reactions: Toutou

UBS28

macrumors 68030
Oct 2, 2012
2,893
2,340
Using TSMC's chiplet tech it is possible to mask stitch 9-16 M1 Max dies together to reach 90-160 CPU cores together.
Form Factor (As of 9 Nov 2021)​
AMD Zen 4 Epyc 128 core cpu Rival​
AMD Zen 4 Epyc 128 core cpu Rival​
Mac Pro
iMac Pro​
Mac mini Pro
Mac Pro
iMac Pro​
MBP 14"
MBP 16"
Mac mini Pro
iMac 24"
iMac Pro​
300mm² Silicon Wafer​
Apple silicon chip​
M1 Max Jade-16C​
M1 Max Jade-9C​
M1 Max Jade-4C​
M1 Max Jade-2C​
M1 Max​
?​
Launch​
?​
?​
Q2 or Q4 2022​
Q2 or Q4 2022​
Q4 2021​
?​
# of dies​
16​
9​
4​
2​
1​
49​
CPU​
160​
90​
40​
20​
10​
490​
performance cores​
128​
72​
32​
16​
8​
392​
efficiency cores​
32​
18​
8​
4​
2​
98​
GPU core​
512​
288​
128​
64​
32​
1568​
Neural Engine core​
256​
144​
64​
32​
16​
784​
memory bandwidth​
6.4TB/s​
3.6TB/s​
1.6TB/s​
800GB/s​
400GB/s​
19.6TB/s​
Max Memory​
1024GB​
576GB​
256GB​
128GB​
64GB​
3,136GB​
Hardware-accelerated H.264, HEVC, ProRes, and ProRes RAW​
16​
9​
4​
2​
1​
49​
Video decode engines​
16​
9​
4​
2​
1​
49​
Video encode engines​
32​
18​
8​
4​
2​
98​
ProRes encode and decode engines​
32​
18​
8​
4​
2​
98​
Peak Quoted Transistor Densities using TMSC's 5nm (2020) at the same die size​
171.3 million transistors per mm²
912 Billion​
513 Billion​
228 Billion​
114 Billion​
57 Billion​
5.19 Trillion​
Estimated Die Size​
17.004cm²​
12.753cm²​
8.502cm²​
6.3765cm²​
4.251cm²​
30cm²​
Peak Quoted Transistor Densities using IBM's 2nm (2025) at the same die size​
333.33 million transistors per mm²
1.774 Trillion​
998.24 Billion​
443.66 Billion​
221.83 Billion​
110.92 Billion​
10.1 Trillion​
Estimated AMD Ryzen 9 5950X Performance​
16x​
9x​
4x​
2x​
1x​
49x​
Estimated RTX 3080 Performance​
16x​
9x​
4x​
2x​
1x​
49x​

But rumor has it that Apple will use 4 x M1 Max. They can probably do more, but then the Mac Pro will not be as quite and cool as what they are targetting.

And seeing how a M1 Max 16" MBP has no problem keeping up with the current Mac Pro in stuff like video editing, 4 x M1 Max is a sufficient bump in speed.
 

deconstruct60

macrumors G5
Mar 10, 2009
12,493
4,053
Using TSMC's chiplet tech it is possible to mask stitch 9-16 M1 Max dies together to reach 90-160 CPU cores together.

Pragmatically not. All of TSMC's "chiplet" technology have reticle limits for 2.5d and 3D packaging. (they are getting bigger each year but there is a cap to current tech ). They could just go with just packages on a effectively just a printed circut board then loose lots in bandwidth/power effectiveness. The other issue is soak up all the package space with CPU die then won't have space for the RAM ( and again loose some bandwidth/power upsides that Apple is leveraging. )

A fan-out of that many tile would also likely have significant NUMA issues to work around. Pretty doubtful Apple wants to get into hard core NUMA triage issues. The GPU won't scale well with high likelihood for a screen output workloads.

macOS doesn't even handle more than 64 threads. No rational reason to go where the OS doesn't go. Even more so on a laptop ( which is 70+% of the macs sold and 100% of the iPad Pro's sold). It isn't where Apple's revenue money is.

Even getting to 4 chiplets/tiles would with minimal NUMA overhead would be a significant technical feat for Apple. Double digits? They have not even a seed of tech to do that right now and still fully meet their "highest perf/watt" mantra.
 
  • Like
Reactions: JMacHack

deconstruct60

macrumors G5
Mar 10, 2009
12,493
4,053
Amd just teased their new Zen 4 epyc cpu with 128 cores what are your guys thoughts on it

There is rumor that AMD my use this size optimized core scaled down to TSMC N3 coupled to a Zen 5 ( also N3) in a APU. From the description it sounds like these Zen4c cores are either N4P ( or an even more highly forked tweak to N5... it isn't standard N5. It probably would be a better fit for N3 but AMD doesn't have time... need to get it out not to address the Neoverse N1 and N2 server SoCs shipping now and in 2022. )

For an AMD APU that also needed to compete with Apple on GPU throughput there could be some upside to trading a larger Zen 5 core for a handful of 'N3 shrunk' 4c cores. ( If the instruction set of Zen 5 basically overlaps 4c then won't have to turn off some features ). AMD then could take that 'saved' space and allocate more GPU cores.

Apple has basically built a GPU with some application (CPU) cores wrapped around it. Most of Intel and AMD APU/laptop processors are skewed closer to the opposite way or closer to 50/50 .

I doubt though that AMD goes "smaller core crazy" like Intel and tries to throw lots of these in an APU ( 6-8 P and 16 c cores ). Probably be closer to Apple's 2 or maybe 4 4c cores. And not see until Zen 5 which is a couple years out. Running two 4c cores where the VCN ( video decode unit ) pumps out some video for the semi-standard battery run time benchmark will probably produce a longer time than a full Zen 5 or Zen 4 core will do. ( just need a core to minimally keep the background OS task handling running. )


The APU cannot be all 4c cores though. These cores are not going to be good for all/general purpose workloads. These Bergamo server packages will be more tuned for a common subset of cloud nodes. ( fixed clocks, relatively narrowly bounded "turbo" , lower focus on single thread maximuization. probably approximated AVX-512 (run 128 twice. ). In short, a smaller core because take some "extra" stuff out that mainly focus on other workloads. They'll be good for Windows 11 assigning background , non high priority tasks to and goosing up MT tech porn benchmark scores. But probably would not want that to be the primary core driving a GUI for single user interactions. The core will be tuned for multiple user workloads; not single users.

The 4c cores are probably cache limited relative to the full Zen 4 ( and later, Zen5 ) cores. The "higher performance" that AMD is claiming is likely the overall aggregate performance of the whole package. ( getting more "bang" because have lots more cores. Not that it does single threaded drag racing better in most workloads).
This is same way the Ampere 80 ( and now 128 ) [ a ARM Neoverse N1 implementation ] is beating current AMD and Intel server processors on several multiple threaded workloads ( e.g., java application server , web , etc. ) Don't need a "barn burner" fast, single threaded core to win that race. If have 40 different folks asking the server to do 30 different things then it isn't a top fuel funny car drag race workload.
 

Adarna

Suspended
Jan 1, 2015
685
429
Pragmatically not. All of TSMC's "chiplet" technology have reticle limits for 2.5d and 3D packaging. (they are getting bigger each year but there is a cap to current tech ). They could just go with just packages on a effectively just a printed circut board then loose lots in bandwidth/power effectiveness. The other issue is soak up all the package space with CPU die then won't have space for the RAM ( and again loose some bandwidth/power upsides that Apple is leveraging. )

A fan-out of that many tile would also likely have significant NUMA issues to work around. Pretty doubtful Apple wants to get into hard core NUMA triage issues. The GPU won't scale well with high likelihood for a screen output workloads.

macOS doesn't even handle more than 64 threads. No rational reason to go where the OS doesn't go. Even more so on a laptop ( which is 70+% of the macs sold and 100% of the iPad Pro's sold). It isn't where Apple's revenue money is.

Even getting to 4 chiplets/tiles would with minimal NUMA overhead would be a significant technical feat for Apple. Double digits? They have not even a seed of tech to do that right now and still fully meet their "highest perf/watt" mantra.
Let’s talk about the 2-4 chiplet design within the next 12 months.

I remember reading people strongly disagreeing about RTX 3080 performance as an iGPU within the last 52 weeks.

Then the M1 Max came out. ??‍♂️??‍♂️??‍♂️

We are in an industry where R&D money creates unknown concepts into actual daily products. So when PCMasterRace types start saying "not possible" then what they know is limited to what Intel/AMD/Nvidia marketing tells them what is possible.

Macs has an annual global shipments of ~22.5 million vs ~252 million of Intel/AMD.

To save on R&D cost why not design the M1 Max to be chiplet-compatible so that the ~0.6 million of the 3 million destined for the 14" & 16" can also be used as Jade-2C & Jade-4C to increase volume & utility to as few a chip designs as possible that can be placed into these form factors that can physically fit them in, cool them adequately and have the PSU overhead
  • Mac mini (at most a Jade-2C on a 150W PSU)
  • iMac 24"/27"/Pro (at most my fictional Jade-9C on a 500W PSU)
  • Mac Pro (at most my fictional Jade-16C on a 1.4kW PSU)
Who would have ever imaging that the M1 chip used by ~80% of all Macs would ever be used on an iPad?

Everyone was expecting an iPhone 12 A14-derived chip, A14Z to be used in the iPad Pro. Putting an M1 chip in an iPad Pro justifies the price gap and the "Pro" moniker between it and all the other iPads.

In my table you will find my fictional chips labeled Jade-9C, Jade-16C & Jade-49C. The naming convention is based on the number of "C" (or dies) per chip SKU would have.

I created these fictional chips because this thread asked what's Apple's answer to AMD server chips and because of this 30cm² die size 2.6 trillion transistor chip is there. Assuming the M1 Max was designed for that purpose then great. Odds are only ~1% of MacRumors users would ever buy one.

ARM supercomputers have been the norm as early as Jun 2020 prior to Nov 2020's debut of the the M1. AWS offers ARM instances. Apple has a lot of cloud computers, supercomputers, data centers and other aging infrastructures that would benefit from the performance per watt advantages of Apple Silicon. Businesses are always looking for ways to lower down overhead like power & water utilities. Homeowners would benefit to have this point of view as well.

Countries like the US have very low $/kWh & $/US gal. as compared to the world but any lowering of expenses is always a plus.
 
Last edited:
  • Like
Reactions: JMacHack
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.