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

crazy dave

macrumors 65816
Sep 9, 2010
1,256
930
if they run the same frequency at the same clock, then since the voltage is the same (presumably), it’s going to be the same power efficiency.

Sorry, I’m tired and I’m not quite following. His statement was that firestorm cores and avalanche cores look identical because at the same clock speed they get the same score. My statement was that we don’t know how much power the avalanche cores are drawing at that clock speed. Therefore the avalanche cores may be more power efficient at the M1’s clock speed than the firestorm cores.
 
  • Like
Reactions: wyrdness

Argon_

macrumors 6502
Nov 18, 2020
423
255
Especially if there’s more headroom in this design. In M1, maybe you can’t increase the clock very much because you are bandwidth limited.

The M1 memory bandwidth would bottleneck a hypothetical 32 core ASi GPU. Eight channel LPDDR5 would be able to feed it though. ~205 GB/s
 
  • Like
Reactions: Ruftzooi

cmaier

Suspended
Original poster
Jul 25, 2007
25,405
33,471
California
Sorry, I’m tired and I’m not quite following. His statement was that firestorm cores and avalanche cores look identical because at the same clock speed they get the same score. My statement was that we don’t know how much power the avalanche cores are drawing at that clock speed. Therefore the avalanche cores may be more power efficient at the M1’s clock speed than the firestorm cores.
Power is going to essentially be determined by clock speed times voltage squared, assuming the cores are around the same size as before, so my point is that at the M2’s clock speed, if they are getting the same performance as firestorm would get, they are also burning the same power.
 

cmaier

Suspended
Original poster
Jul 25, 2007
25,405
33,471
California
The M1 memory bandwidth would bottleneck a hypothetical 32 core ASi GPU. Eight channel LPDDR5 would be able to feed it though. ~205 GB/s

Sure, there are two issues. Parallelism and frequency. Increasing the bandwidth allows you to speed up the CPU clock without starving the data or instruction pipe, and also allows you to support more graphics bandwidth.
 

cmaier

Suspended
Original poster
Jul 25, 2007
25,405
33,471
California
Ha. So the geekbench scores are 10% higher single core and 18% higher multicore. If you pretend that averaging them is a thing that makes sense, that’s 14%, which is what i predicted here a month ago or something. I’ll take my cookie now.
 

deconstruct60

macrumors G5
Mar 10, 2009
12,227
3,827
Dylan Patel says A15 has MMU and Nested virtualisation.

The more telling tweets there isn't the first one in the thread. It is 4 and 5. In 5 (chart for individual elements of Metal score) there is a ginormous skew for the depth of field calculation. in 5 (chart for single/multiple) breakdown a high skew on speech recognition. The multiple core zeros for crypto , AES , and machine learning

Apple has multiple matrix multipliers with different emphasis. "Neural" cores, AMX ( hanging off the P cores) , and some customization in the GPU to facilitate. Those plus the larger cache to feed those probably is a big factor here. But also indicative that these "lowest common denominator" benchmarks may not sure work done if most work is off in semi-custom accelerators. ( going forward every CPU isn't necessarily going to have the same set of accelerators. )


As for MMU... If have been running Mach (as iOS has) then there has been an MMU. "new" MMU and nested virtualization probably are not entirely decoupled as two "new" thing. Really two parts of the same new thing. That isn't flowing out of the benchmarks though. ( perhaps some tidbit passed along from someone with early access to a device. )

If Apple wants to put in more layered defenses on the iOS devices also then this is useful even on a "mere phone" ( not really a phone, but a smaller, but full capable, computer. ). For example, Windows 11 using virtualization to disable being to inject code changes into the Windows 11 kernel. Apple has some other mechanisms ( read only kernel RAM) , but cutting off breach vectors from multiple directions can be even safer.

VWmare ( substantive feature for them ) and Microsoft ( long term for Window 11 ) probably thought that was missing "years ago". Apple's virtualization framework couldn't provide what it didn't have. Should get more traction on Mac side , but wouldn't be shocking if it showed up in iPadOS not that far out into the future.
 

crazy dave

macrumors 65816
Sep 9, 2010
1,256
930
Power is going to essentially be determined by clock speed times voltage squared, assuming the cores are around the same size as before, so my point is that at the M2’s clock speed, if they are getting the same performance as firestorm would get, they are also burning the same power.

Do we know they are using the same voltage? I just assumed firestorms in the M1 drew more volts than firestorms in the A14 to increase the clock speed by 10% and thus the avalanche cores might be able to achieve those same clocks with less voltage - ie say what the firestorms pull in for the A14. Or is that not how it works?
 

crazy dave

macrumors 65816
Sep 9, 2010
1,256
930
Ha. So the geekbench scores are 10% higher single core and 18% higher multicore. If you pretend that averaging them is a thing that makes sense, that’s 14%, which is what i predicted here a month ago or something. I’ll take my cookie now.

Well if we take the higher multi core score of the Max Pro, it’s a 21% improvement in MT making the average 15.5%. So you were wrong by a whopping 1.5% for the bigger models ;) you know as you said if averaging was a thing
 

altaic

macrumors 6502a
Jan 26, 2004
636
426
Recent developments in this thread concur with my expectation that good things are afoot regarding the next M iteration. Thank you all for the thoughtful investigation and insight.
 
  • Like
Reactions: jabbr

Juraj22

macrumors regular
Jun 29, 2020
169
196
Considering that single M1 core can read at 58GB/s, write at ~33GB/s (source: Anandtech) Apple need to add more memory bandwidth. Considering more GPU clusters,more CPU cores, they need to add a lot of bandwith. Seems to me like they are going to make similar CPU like in Xbox/PS5 with much less heat.
 

cmaier

Suspended
Original poster
Jul 25, 2007
25,405
33,471
California
Do we know they are using the same voltage? I just assumed firestorms in the M1 drew more volts than firestorms in the A14 to increase the clock speed by 10% and thus the avalanche cores might be able to achieve those same clocks with less voltage - ie say what the firestorms pull in for the A14. Or is that not how it works?
The voltage is determined by the physics of the materials being used and the size of the transistors. It likely hasn’t changed. If you lower it too much you can’t turn off the transistors (the equation I gave is for dynamic power - the power due to work being done to process. There is also static power, caused by leaky transistors that don’t shut all the way off, that is independent of work being done). Lowering the voltage also can make the chip unreliable, by causing signal transitions to slow down - that makes them susceptible to cross-coupling to neighboring signals, and can cause wires to have the wrong value at the wrong time and break the chip.


So, the voltage doesn’t change very much except over the course of decades.
 

cmaier

Suspended
Original poster
Jul 25, 2007
25,405
33,471
California
Well if we take the higher multi core score of the Max Pro, it’s a 21% improvement in MT making the average 15.5%. So you were wrong by a whopping 1.5% for the bigger models ;) you know as you said if averaging was a thing

I suck :)

In a world where anyone I know who might know about the performance told me anything (very hypothetical), such a person would probably not be relying on geekbench, of course. CPU designers have their own array of benchmarks. I dont actually know what they use nowadays. Back when, when I was a chip lad, we used to concentrate on SPEC plus a lot of other benchmarks in a suite.
 

MrGunny94

macrumors 65816
Dec 3, 2016
1,031
588
Malaga, Spain
Considering that single M1 core can read at 58GB/s, write at ~33GB/s (source: Anandtech) Apple need to add more memory bandwidth. Considering more GPU clusters,more CPU cores, they need to add a lot of bandwith. Seems to me like they are going to make similar CPU like in Xbox/PS5 with much less heat.
Yep this would fix the current implementation no doubt
 

Ruftzooi

macrumors newbie
Sep 9, 2021
5
45
The more telling tweets there isn't the first one in the thread. It is 4 and 5. In 5 (chart for individual elements of Metal score)
So this tweet shows the 48% increase between 9608 Metal-score for the A14 and a 14216 score for the A15, which is very impressive. Now if the A15 had 4 GPU-cores instead of 5, that would be a score of 14216/5*4 = 11373 which is about 18.3% more than a 4-core A14. If the Metal score scales linearly with GPU cores that would mean a A15-GPU-core is about 18.3% faster than a A14-GPU-core.

M1 Mac mini's with 8 A14-Based GPU-cores show Metal scores between 20700 and 23500. Let's say 22000 for easy math. If the coming generation M-chip (I think it'll be the M2 and not the "M1X") has A15 Based GPU's and the performance increase is similar to the phones, that would mean an M2 with 8 GPU cores would have a Metal score of around 26000. And the hypothetical 16-GPU-core M2 could score up to 52000, which easily beats the Radeon Pro 5600M in the current 16" Macbook Pro, which hovers around a 43000 Metal-Score and should be a very reasonable upgrade. It would even come close to the RX5700 in last years 27" iMacs! This is ofc assuming clock-speeds can stay the same with such a core increase, memory bandwidth allows it and everything actually scales linearly.

Also, during the event they boasted about A15's ability to hardware encode/decode ProRes video. A feature that could put the Mac Pro's afterburner-card to shame and I definitely expect it in the next M-Soc for macs. So that's another reason I think they'll have A15-based chips.
 
  • Like
Reactions: wyrdness and Homy

Homy

macrumors 68020
Jan 14, 2006
2,073
1,924
Sweden
So this tweet shows the 48% increase between 9608 Metal-score for the A14 and a 14216 score for the A15, which is very impressive. Now if the A15 had 4 GPU-cores instead of 5, that would be a score of 14216/5*4 = 11373 which is about 18.3% more than a 4-core A14. If the Metal score scales linearly with GPU cores that would mean a A15-GPU-core is about 18.3% faster than a A14-GPU-core.

M1 Mac mini's with 8 A14-Based GPU-cores show Metal scores between 20700 and 23500. Let's say 22000 for easy math. If the coming generation M-chip (I think it'll be the M2 and not the "M1X") has A15 Based GPU's and the performance increase is similar to the phones, that would mean an M2 with 8 GPU cores would have a Metal score of around 26000. And the hypothetical 16-GPU-core M2 could score up to 52000, which easily beats the Radeon Pro 5600M in the current 16" Macbook Pro, which hovers around a 43000 Metal-Score and should be a very reasonable upgrade. It would even come close to the RX5700 in last years 27" iMacs! This is ofc assuming clock-speeds can stay the same with such a core increase, memory bandwidth allows it and everything actually scales linearly.

Also, during the event they boasted about A15's ability to hardware encode/decode ProRes video. A feature that could put the Mac Pro's afterburner-card to shame and I definitely expect it in the next M-Soc for macs. So that's another reason I think they'll have A15-based chips.

I like that. 32 cores would mean 104 000, on par with Radeon Pro Vega II. :D
 
  • Like
Reactions: Ruftzooi
Register on MacRumors! This sidebar will go away, and you'll see fewer ads.