Yes I believe Apple still holds the efficiency crown in mobile. We need to wait for mobile zen 4 to compare tho.
Got a tip for us?
Let us know
Become a MacRumors Supporter for $50/year with no ads, ability to filter front page stories, and private forums.
[CPU only] Apple M1/2(Max/Ultra) (TSMC 5nm) vs AMD Zen 4 (TSMC 5nm) - Technical Analysis
- Thread starter exoticSpice
- Start date
- Sort by reaction score
Zen4 ST is very impressive. More than I thought.
However, Zen4 did raise TDP. It wouldn't surprise me if it takes 5-10x the package power of the M2 to reach 15% faster ST.
Zen3 took 50w package power during ST. Zen4 is probably around 60-70w. M2 Air takes about 6w during Geekbench.
Where do you get these numbers? 5950X consumes up to 20W per core, ”normal” Zen3 cores in a more moderately clocked CPUs are 10-15W. Package power for desktop Zen is misleading anyway since a lot of it is because the I/O die. If I understand it correctly, that part has been substantially improved with Zen4.
If I were to speculate, much of it probably has to do with the I/O die. It was a power hog on the desktop Zen3 and has been made much more efficient with Zen4. So the entry-level configurations have more power budget for CPU cores. At the high end, diminishing results kick in. At those clocks the top configs must be running well outside the optimal perf/watt range.
Laptops do not use the I/O die so if my speculation is correct this particular path to improving efficiency won’t be relevant for them. But clearly AMD has managed to get impressive improvements from 5nm and uarch tweaks anyway.
IIANM the IO Die went from 12nm to 6nm (I guess they finally finished the contract with GloFo for them) so that would allow for a huge amount of power savings.
I've only looked at titles and announcements and have not dug into the details yet but this seems like a pretty good win given that these CPUs have more compute power than most people need. Better efficiency is always welcome - particularly in Europe and Asia where power prices have been soaring.
I think Apple already spends quite a lot of money on these things
At least part of their success can be attributed to them being able to throw more money at the problem — their chips are not the end product and so they don't need to be competitive at the market. Hence the use of more advanced nodes, wide memory interfaces and custom memory packages (both a big no-go for a commercial consumer chip), large caches etc. I have little doubt that M series chips themselves cost more to produce than most x86 chips.The foundational elements that help with higher IPC start to diminish has crank the clocks higher. That lower 65W limit was likely turning the "Turbo" stuff off. As you crank to the highest possible clocks the improvements fade ( diminishing returns ).
For example , better branch predictions gets you fewer missed branches. Once hit a miss though the bigger the clock gap better CPU and Memory means a bigger pipeline bubble. Same thing with improved load/stores. If do them too quickly can cause a "traffic jam". If there is a better balance between requests and available bandwidth get smoother throughput ( DDR5 is helping with better bandwidth there ).
Bigger caches ( of various forms ... pipeline prediction , instruction/data cache ) help and again diminishing returns curve as memory stall relatively consumes more cycles/time.
If look at the articles where compare the 5000 and 7000 series the 'base' clocks of the 7000 series line up with about 85-95% of the turbo from the previous version. For example base 7950 4.5 and Turbo 5950X 4.9 ( 91% ).
A decent chunk of that is just N5 vs N7 (bigger caches and power savings ) . Faster clocked CPU + faster clocked DDR5 is just a better balance at 65W.
As they get 'greedy' on cranking the CPU clocks that balance disappears ( and are tossing power out the window like they were on N7 at turbo also. )
[ Also as pointed out there is a better power allocation balance with I/O on N6 . Also change in fab. ]
Another factor is that they controlled the bloat. Only implementing DDR5. Double pumping AVX-512 . about the same size L2.
The efficiencies here are mainly for AMD's server product. They are going to have some utility in the laptop sector also. But these are not laptop first focused cores.
Even the Zen 4C (cloud) cores that come later probably won't be laptop focused either. What should make Intel nervous is that these 'big" Zen4 cores are just the warm up for the even more dense packed cores at modest clocks.
Apple spends lots of effort to stay far out in front of the laptop chips competition. The M series cores are not substantially different than the A series cores. The entire desktop line up so far is based on laptop chips. So the notion that they don't have a laptop focus is a huge disconnect from what they are actually doing. Laptops/mobile is all they do so far.
What Apple is not spending tons of money on is 5% faster for 50W more power consumption benchmark wins.
With no 'E' cores there is little chance zen 4 is going to win the mobile efficiency crown on low-modest workloads.
Likely not either on GPU side either (as adapting down the desktop GPU into mobile use . The AMD - Samsung partnership still has significant work to do. ).
Zen 4C is likely not an exceptionally good mobile 'E' either.
P.S. benchmark scores that don't factor in battery life though Zen4 should turn in scores. Higher end laptops they should make more market penetration on in the overall market.
Zen 4C cores will be smaller, not more efficient. However, Phoenix cores will be more efficient.
What does Samsung have to do with AMD GPUs? TSMC produces AMD's CPUs and GPUs.
Phoenix appears to be a shift to N4 variant (N4P ?) from N5. I wouldn't expect a more core implementation shift. ( they did some prior to Zen 4 with the N7-> N6 6000 series work. )
Dumping the multiple chip package is likely saving more power and boosting Pref/Watt that any CPU core specific changes. The CPU cores , GPU cores , and RAM controllers all on one die is better for Pref/Watt even if made zero changes to those three subsystem implementations.
Samsung makes SoCs . Their latest top end SoC has AMD RDNA tech in it.
Samsung To Keep Using AMD's RDNA GPUs For Exynos SoCs
Samsung's Xclipse is not set to eclipse.
Yes, that SoC and RDNA implementation is targeting Samsung fab process. If they don't work out their fab problems then that is likely a dead end. But not just for the GPU part of the SoC. If they do get the gate-all-around N3 process bugs worked out then , no AMD isn't going to walk away from that. Samsung is doing most of the port work, but AMD is not completely uninvolved.
But is the I/O die doing anything? If load up the I/O die with an active 2 CU iGPU and pull lots of workload through the PCI-e v5 lanes is it going to be a lower power consuming cohort) than the older I/O dies? If send large chunks of the I/O die to sleep then sure the power goes down.
AMD's benchmarks here in these slide are probably somewhat cherry picking a bit to offload power on the i/O die. ( for such an relatively expensive CPU is likely teamed up with a dGPU though. Video en/decode could be a workload though even with a dGPU present. )
You don't know that unless we have benchmark scores, it could well be more efficient overall.
That's because they don't compete in the traditional PC space, also the reason why you can't buy 128GB RAM & just use it on your Mac or MacBook. Pretty sure I'm fine with slightly less efficient cores if they give me the ability to use a lot more memory or storage!
I/O die has the memory controller, PCIe 5.0 & RDNA2 IGP. Not 100% sure about the last one but going from 12nm to 6nm will do wonders for their power consumption figures overall.
Last edited:
Yeah, the I/O die contributes significantly to the power consumption of Zen3. It hosts the L3 cache and the memory controllers (as pointed out by R!TTER). You can clearly see it in power consumption tests here: https://www.anandtech.com/show/1621...e-review-5950x-5900x-5800x-and-5700x-tested/8 - note the huge difference between core power and package power, almost 20-30 watts for lower loads. If AMD has managed to reduce this to some meaningful figure like 5-10watts, the entry level CPUs will gain a lot of breathing space within their TDP bracket.
Because Apple chips share the RAM with video too. Its shared memory. Also has the LPDDR5 RAM has much higher bandwidth needed for the GPU.
Shared memory isn't the reason why you can't upgrade RAM on Apple chips. LPDDR5 is but that's only a part of the reason, also AMD is probably going to implement this in a gen or two for PC. They're already doing this on consoles & there's something similar for EPYC as well IIRC.
It'll be interesting to see how the M2 Ultra (M2 Ultra X2?) compares to the 7950x and 13900k
It does but its also very expensive to get. For an M1 Ultra you need a $4000 Mac and it comes with a measly 64GB RAM and 1TB SSD and a binned 48 core GPU.
[Going on a tangent]
The 7950X is $699 and say a RX 7800 is $899 and the rest of the system is $1500, so total = $3,100.
That would be $900 less than a Mac Studio. Apple's value just got dropped very hard and the pref is not even top of the range as its lacking in ST.
The AMD system also supports user upgrades like SSD and RAM and even the Motherboard will support Zen 5 for future CPU upgrades and also GPU upgrades that are much cheaper than Apple's. AMD supports 32 bit apps and still has all the baggage of x86 and is still showing excellent numbers.
So this begs the question, why is Apple soldering everything in their desktop Macs and what's the point of removing 32bit support when AMD showed its possible to still get great pref while maintaining backwards compatibility.
Apple needs to update the Mac Studio with faster CPU and GPU. It's already outdated as AMD's CPU has a very high ST and perf per watt is great too.
Last edited: