In this article, I’m going to mostly focus on the power and performance of the latest Apple M5 (especially Pro/Max) and Qualcomm Snapdragon Elite X2E chips across three benchmarks: Cinebench R24 ST (CPU), Cinebench R24 MT (CPU), and Cyberpunk 2077 1080p Ultra (GPU). The data as always comes from Notebookcheck, while the master plots used to generate the following figures can be found on https://powervsperformance.tiiny.site/.
(All figures are thumbnails, click to enlarge)
CPU
Starting with single threaded Cinebench R24:
Fig. 1 and 2:
(Fig 1) Apple has been increasing the ST power draw of the M-series of chips for every generation we have data though the M5 power draw is still less than the most efficient x86 chips and delivers far more performance as well. At first, I thought the most recent change in the M5’s power/efficiency might've been due to the original test machines having a much larger RAM configuration than before, increasing power consumption, but a more recent test of the M5 with a base configuration of RAM drew even slightly more power. While I can’t discount higher RAM configurations drawing more power for every M5 processor here, it seems unlikely to be the main contributor and M5 ST power draw is likely just going up. However, it should be noted that efficiency across the M5 variants seems to be holding fairly steady, with only a slight dip in the base M5.
The new Qualcomm X2E processors show a similar pattern: a collinear increase in power and performance (relative to the most performant X1E variant above, the X1E-84) leading to a similar efficiency rating as to their previous chips. However, the new Qualcomm chip is significantly larger than its previous iteration in terms of the number of CPU cores and its effective CPU size* (Fig 2) - so some increase in its power draw is likely due to that rather than just pushing ST performance (i.e. Oryon P-core architecture improvements may be delivering better efficiency than what we see at first glance). Relative to Apple, it’s a little faster than an M3, but still much slower than an M4 - most similar in performance in fact to the A18 Pro, which shares the same P-core as the M4, down-clocked. The new chip is on the same node generation as the M5 (X2E-88 is N3P while X2E-94 is N3X). Compared to the most recent x86 processor on the market, the Intel 388H, the new Qualcomm chip is 17% faster in Cinebench R24 and 13% more efficient despite the Qualcomm's larger effective CPU size. The only x86 chips that come close to it in performance are the desktop class chips from Intel and AMD drawing more power.
*as explained on the website, effective CPU size is a heuristic I created to account for things like SMT and different core types when trying to approximate the total expected multi-threaded throughput of a CPU as opposed to naively counting cores/threads.
For the CB R24 multithreaded benchmark, starting with just the recent Apple M4 and M5 and Qualcomm processors:
Fig. 3:
(Fig 3) Qualcomm’s MT performance is around the 14-core M4 Pro though it needs 12 P-cores and 6 M-cores to get there. As it reduces power, it is around the base M4/M5-level in efficiency and performance, but again it requires a much lager processor to achieve that level of performance/efficiency. The two Qualcomm variants are supposedly manufactured on different nodes, N3X and N3P but I don’t see much evidence of that here, they seem to be on pretty much the same power/performance curve. So if so, I’m unsure as to what the point of the different nodes was.
The Apple M5 data points are interesting … the circled processor is the full M5 Max thermally throttled and (potentially power throttled) in the 14” MacBook Pro chassis. It is however, possibly by coincidence, almost exactly where I would expect the binned 15-core M5 Pro CPU to be. The MT tests that Notebookcheck run are run in “High performance” mode. The M5 Max in the 16” laptop seems to allow for higher thermals than its Pro sibling despite them sharing the same core count and chassis. However, in this High Power mode, the increase in performance of the Max doesn’t match its increased power draw over the Pro lowering its relative efficiency (side note: NBC went back and reran the Pro CPU tests and got much more sane results than they did the first time).
One of the major exciting changes in the M5 generation is that Apple debuted a new SOC design for the M5 Pro/Max, introducing M-cores for the first time and reducing the number of P-cores. This allows Apple to push both power and performance without losing efficiency (at least in the more appropriately tuned M5 Pro) relative to the full M4 Max. But the real goal of such M-cores is to attain area efficiency - use less silicon to achieve the same or better multi-core performance. Thus to really see if Apple's new M-core has paid off requires die shots of the new CPU and see if/by how much it reduced the silicon footprint of the M5 Max/Pro CPU relative to the M4 Max (or perhaps more accurately what the M5 Pro/Max would’ve been if it had kept the old design). According to my simple heuristic, Apple's multi-threaded design in the new 18-core M5 Pro/Max should deliver roughly the same number of "Effective P-threads", slightly more, as the 16-core M4 Max (Fig. 4).
However, the power increase and new design has meant there is now a sizable gap between the most performant base M5 device and the M5 Max with fewer variants to bridge it. The solid line in Fig. 3 is the gap between the base M5 and the M5 Pro with the same 18 cores as the Max. The dashed line is the M4 gap between base and Max. The previous M4 generation had two binned CPUs to fill that gap at fairly regular power and performance intervals while the M5 generation has only one, the binned M5 Pro represented by the constrained M5 Max, and again a much larger gap. If Apple keeps a similar structure in the next M6 generation, then there’s almost space for a third CPU design beyond the base and Max … perhaps a unique Pro CPU similar to the M3 generation?
Zooming out to view all the processors at once, plotting efficiency on the x-axis and using the effective CPU size as the bubble size, another interesting pattern emerges:
Fig. 4:
(Fig. 4) One can really see for Apple almost a line or target efficiency for any processor that isn’t thermally constrained - roughly around 24 pts/W in Cinebench R24. While AMD and Intel are happy to reduce efficiency to increase performance, Apple uses a greater number of cores to climb in performance, yes increasing power consumption, but staying within that efficiency window. Even the M5 Max, which pushes this limit, doesn’t push it very far (21.9 pts/W). Qualcomm also seems to like this efficiency window for its CPUs. When compared to larger AMD/Intel chips with similar performance, both are far more efficient.
In conclusion, Apple and Qualcomm currently enjoy a healthy lead over Intel and AMD in Cinebench R24. They are able to generate more MT performance at lower power using smaller processors. Similarly, for ST performance, AMD and Intel have to burn a lot more power to get close to Qualcomm's latest X2E chip while not getting close to matching the M5's performance, never mind its efficiency.
GPU
Moving on to the GPU, let's start the Cyberpunk 2077 1080p Ultra based GPU analysis with the smaller GPUs. This is a raster-only benchmark - no ray tracing, no upscaling or frame generation.
Fig. 5 and 6:
(Fig 5) First thing to note is that Qualcomm is running Cyberpunk 2077 under Prism emulation. At 1080p Ultra, this emulation may matter as the CPU performance still affects performance and efficiency. This is especially true for larger GPUs (more on that later), but this can also be true even for smaller GPUs (though maybe the smallest like the A18 Pro are completely GPU bound). However, I decided to include the X2E's Adreno X2-90 GPU data points anyway as, unlike the earlier Qualcomm chips and earlier x86 emulation, the tested devices do pretty well despite running under x86->ARM emulation.
The second thing to note is that this is not the standard CP2077 FHD test Notebookcheck runs and reports results for. For this test, there is a 30 minute warm-up prior to running this benchmark making this version of the benchmark closer to the end of an endurance test rather than a sprint. This is something I talk about on the main webpage. I’m still a little unclear exactly about the particulars of the test, but most processors suffer a 9-10% reduction in performance compared to Notebookcheck’s standard benchmark results for the same games and settings. There are exceptions however: the base M5 (not Pro or Max models) both binned and full doesn’t suffer any noticeable degradation compared to its standard CP2077 FHD Ultra benchmark results and the A18Pro actually improves on its standard FPS in the endurance test. How much of that is noise is unclear.
The third thing is that, for Apple, CP2077 tends to be on the lower end of performance relative to other chips. In other words, when comparing Apple to Nvidia or AMD or Intel across different benchmarks like Steel Nomad, CP2077, Assassin’s Creed Shadows, Baldur’s Gate 3, and Total War: Pharaoh - Apple tends to struggle in CP2077 more than the others (though to be fair to AMD, Steel Nomad’s performance on AMD GPUs is quite bad compared to AMD’s actual gaming performance, unsure why - Steel Nomad is not so unrepresentative with any other GPU).
(Fig 6) According to just the FP32 unit count (as well as the memory bandwidth), the new X2-90 GPU should be something in between the M5 Pro and base M5 GPU and roughly double the 8-core M5 GPU in size. However, according to silicon die shots, it is reported to be roughly the same size, in silicon, as what the 10-core M5 is likely to be. As noted in the powervsperformance website, different GPU manufacturers can achieve different levels of compaction in their design based on things like whether or not FP16 and Int have separate functional units, the amount of special math function units, the amount and type of cache, as well as the presence and design of matrix and ray tracing acceleration units. In this benchmark, the Qualcomm Adreno GPU behaves more similarly to the base M5 in performance and efficiency such that you could draw a pretty straight between the three points. However, given all the caveats mentioned earlier, this concordance with silicon size is more likely coincidence than not. The more standard CP2077 FHD Ultra test shows the Adreno X2-90 GPU to be more performant than in this benchmark here (whilst the base M5 stays the same) and of course there is the issue of x86 emulation vs native performance.
Against the 890m and B390, it is roughly on par or better than these GPUs, but it doesn't seem to scale as well as the B390, both of which have fewer FP32 units, but, again, don't labor under emulation.
(Fig 5, 6, and 7) As we move into larger GPUs, the Qualcomm GPU can’t really compete here with the 25% larger 5050m, 8060S, and 20-core M5 Pro GPUs. This despite the fact that the Qualcomm CPU is more a competitor to the Apple Pro-level and AMD Halo-level CPUs. The Qualcomm iGPU is fine when competing with smaller GPUs from Apple, AMD, and Intel, but is a mismatch with its large, powerful CPU. That might make sense if Qualcomm was ever optionally pairing their SOC with a larger dGPU from say Nvidia, but that hasn't happened and doesn't seem likely in the near future.
Fig. 7:
(Fig. 7) At this power level, in this benchmark, the Nvidia chip and the Apple M5 Pro share almost the same performance and efficiency. They both have 2560 FP32 units (Fig 8), though the Nvidia chip has slightly higher bandwidth (384GB/s vs 307GB/s). Again, despite the similarities of the two GPUs, this alignment of the resulting performance/efficiency is probably once again a bit coincidental: 1) the CPUs, which matter for a 1080p Ultra benchmark, are very different (Apple’s should be much more efficient than the Intel one in the PC), 2) this benchmark is one of Nvidia’s best, 3) Nvidia is on an older manufacturing node. The combination of these three factors all seem to cancel each other out. Different benchmarks and a different PC CPU would likely yield a different relationship.
Having said that, most of these Nvidia chips were paired with Intel CPUs and when doing so we see a pretty good concordance for this benchmark even beyond the M5 Pro and 5050m. The M4 Max and M5 Max chips** correspond roughly to the 5070m and 5070M Ti GPUs (when the latter are set to similar power draws) and which makes sense given that these Nvidia GPUs bracket the Apple ones in size (Fig. 9). Nvidia dGPUs can of course be set to much higher power draws with concomitant reductions in efficiency, but as shown here can be just as efficient as their Apple counterparts as well.
**not only is the M5 in the 14” thermally constrained, it also appeared to draw from the battery during the benchmark. I have kept the data point in Figure 7 (but not 8 or 9) to note that is what happened, but it may not be otherwise a very useful data point - unsure how much it was actually drawing from battery as the result is pretty similar to what one would expect just from thermal throttling, but there is a definite asterisk next to its result. For the M5 Max in the 16" MBP, no such battery drain was noted during the test but NBC never published a full review of the 16" model and so I had to estimate the idle power from previous 16" MBPs. Given how consistent Apple's designs are across the generations, I'm probably very close here and the effect on the resulting Power and Efficiency measures is likely minimal.
Fig. 8:
(Fig 8) Focusing on just the efficiency of the iGPUs, like with the CBR24 MT benchmark, we see a pretty clear target window for efficiency in the last two generations of Apple processors (when not thermally constrained).
In conclusion, Apple and Qualcomm are delivering solid GPUs, but in this 1080p raster-only gaming test, we don't see the advantage here that they enjoyed on the CPU-side. All 5 vendors deliver competitive gaming GPUs using very different architectures. While Nvidia has yet to deliver a consumer iGPU (a couple of SOCs are rumored to be coming soon), they enjoy sole ownership of the large mobile dGPU market with only Apple supplying a competitor, the Max, for its midrange options and Apple's Ultra reserved for the desktop Studio mini-workstations.
A small note on the deficiencies of this test as we get into higher GPU sizes:
Fig. 9:
(Fig 9) Because the test is run at 1080p Ultra, the benchmark will become increasingly CPU-bound as the GPUs gets larger. This can be demonstrated by plotting GPU size vs performance. I’ve drawn two lines: a solid line for a higher performance, lower efficiency setting for the Nvidia GPUs and a dashed line for the lower performance, higher efficiency setting. In both cases, we can see the efficiency of the Nvidia GPU drop as size increases, especially for the lower power/higher efficiency setting. Attempting to maintain efficiency meanwhile results in lower performance gains in the largest GPUs (curved lines). Looking at results on Notebookcheck's website from the standard benchmarks run from low settings to 4K meanwhile, one finds the larger GPUs increasing in their performance advantage over smaller ones, especially at 4K indicating they are finally being taxed enough at the higher resolutions for their size to matter. Further, not pictured here, but some of the 5080 (desktop) results paired with AMD X3D CPUs show a huge improvement in performance and efficiency relative to the Intel CPU tested in the 5080 (desktop) shown above. Finally, a possible contributory factor is that, starting with the 5070M Ti, both memory bus and GDDR7 speeds increase as the Nvidia GPUs gets bigger no doubt resulting in higher energy costs compared to lower end GPUs. If this is a contributory factor, I expect it to be a smaller one.
(All figures are thumbnails, click to enlarge)
CPU
Starting with single threaded Cinebench R24:
Fig. 1 and 2:
(Fig 1) Apple has been increasing the ST power draw of the M-series of chips for every generation we have data though the M5 power draw is still less than the most efficient x86 chips and delivers far more performance as well. At first, I thought the most recent change in the M5’s power/efficiency might've been due to the original test machines having a much larger RAM configuration than before, increasing power consumption, but a more recent test of the M5 with a base configuration of RAM drew even slightly more power. While I can’t discount higher RAM configurations drawing more power for every M5 processor here, it seems unlikely to be the main contributor and M5 ST power draw is likely just going up. However, it should be noted that efficiency across the M5 variants seems to be holding fairly steady, with only a slight dip in the base M5.
The new Qualcomm X2E processors show a similar pattern: a collinear increase in power and performance (relative to the most performant X1E variant above, the X1E-84) leading to a similar efficiency rating as to their previous chips. However, the new Qualcomm chip is significantly larger than its previous iteration in terms of the number of CPU cores and its effective CPU size* (Fig 2) - so some increase in its power draw is likely due to that rather than just pushing ST performance (i.e. Oryon P-core architecture improvements may be delivering better efficiency than what we see at first glance). Relative to Apple, it’s a little faster than an M3, but still much slower than an M4 - most similar in performance in fact to the A18 Pro, which shares the same P-core as the M4, down-clocked. The new chip is on the same node generation as the M5 (X2E-88 is N3P while X2E-94 is N3X). Compared to the most recent x86 processor on the market, the Intel 388H, the new Qualcomm chip is 17% faster in Cinebench R24 and 13% more efficient despite the Qualcomm's larger effective CPU size. The only x86 chips that come close to it in performance are the desktop class chips from Intel and AMD drawing more power.
*as explained on the website, effective CPU size is a heuristic I created to account for things like SMT and different core types when trying to approximate the total expected multi-threaded throughput of a CPU as opposed to naively counting cores/threads.
For the CB R24 multithreaded benchmark, starting with just the recent Apple M4 and M5 and Qualcomm processors:
Fig. 3:
(Fig 3) Qualcomm’s MT performance is around the 14-core M4 Pro though it needs 12 P-cores and 6 M-cores to get there. As it reduces power, it is around the base M4/M5-level in efficiency and performance, but again it requires a much lager processor to achieve that level of performance/efficiency. The two Qualcomm variants are supposedly manufactured on different nodes, N3X and N3P but I don’t see much evidence of that here, they seem to be on pretty much the same power/performance curve. So if so, I’m unsure as to what the point of the different nodes was.
The Apple M5 data points are interesting … the circled processor is the full M5 Max thermally throttled and (potentially power throttled) in the 14” MacBook Pro chassis. It is however, possibly by coincidence, almost exactly where I would expect the binned 15-core M5 Pro CPU to be. The MT tests that Notebookcheck run are run in “High performance” mode. The M5 Max in the 16” laptop seems to allow for higher thermals than its Pro sibling despite them sharing the same core count and chassis. However, in this High Power mode, the increase in performance of the Max doesn’t match its increased power draw over the Pro lowering its relative efficiency (side note: NBC went back and reran the Pro CPU tests and got much more sane results than they did the first time).
One of the major exciting changes in the M5 generation is that Apple debuted a new SOC design for the M5 Pro/Max, introducing M-cores for the first time and reducing the number of P-cores. This allows Apple to push both power and performance without losing efficiency (at least in the more appropriately tuned M5 Pro) relative to the full M4 Max. But the real goal of such M-cores is to attain area efficiency - use less silicon to achieve the same or better multi-core performance. Thus to really see if Apple's new M-core has paid off requires die shots of the new CPU and see if/by how much it reduced the silicon footprint of the M5 Max/Pro CPU relative to the M4 Max (or perhaps more accurately what the M5 Pro/Max would’ve been if it had kept the old design). According to my simple heuristic, Apple's multi-threaded design in the new 18-core M5 Pro/Max should deliver roughly the same number of "Effective P-threads", slightly more, as the 16-core M4 Max (Fig. 4).
However, the power increase and new design has meant there is now a sizable gap between the most performant base M5 device and the M5 Max with fewer variants to bridge it. The solid line in Fig. 3 is the gap between the base M5 and the M5 Pro with the same 18 cores as the Max. The dashed line is the M4 gap between base and Max. The previous M4 generation had two binned CPUs to fill that gap at fairly regular power and performance intervals while the M5 generation has only one, the binned M5 Pro represented by the constrained M5 Max, and again a much larger gap. If Apple keeps a similar structure in the next M6 generation, then there’s almost space for a third CPU design beyond the base and Max … perhaps a unique Pro CPU similar to the M3 generation?
Zooming out to view all the processors at once, plotting efficiency on the x-axis and using the effective CPU size as the bubble size, another interesting pattern emerges:
Fig. 4:
(Fig. 4) One can really see for Apple almost a line or target efficiency for any processor that isn’t thermally constrained - roughly around 24 pts/W in Cinebench R24. While AMD and Intel are happy to reduce efficiency to increase performance, Apple uses a greater number of cores to climb in performance, yes increasing power consumption, but staying within that efficiency window. Even the M5 Max, which pushes this limit, doesn’t push it very far (21.9 pts/W). Qualcomm also seems to like this efficiency window for its CPUs. When compared to larger AMD/Intel chips with similar performance, both are far more efficient.
In conclusion, Apple and Qualcomm currently enjoy a healthy lead over Intel and AMD in Cinebench R24. They are able to generate more MT performance at lower power using smaller processors. Similarly, for ST performance, AMD and Intel have to burn a lot more power to get close to Qualcomm's latest X2E chip while not getting close to matching the M5's performance, never mind its efficiency.
GPU
Moving on to the GPU, let's start the Cyberpunk 2077 1080p Ultra based GPU analysis with the smaller GPUs. This is a raster-only benchmark - no ray tracing, no upscaling or frame generation.
Fig. 5 and 6:
(Fig 5) First thing to note is that Qualcomm is running Cyberpunk 2077 under Prism emulation. At 1080p Ultra, this emulation may matter as the CPU performance still affects performance and efficiency. This is especially true for larger GPUs (more on that later), but this can also be true even for smaller GPUs (though maybe the smallest like the A18 Pro are completely GPU bound). However, I decided to include the X2E's Adreno X2-90 GPU data points anyway as, unlike the earlier Qualcomm chips and earlier x86 emulation, the tested devices do pretty well despite running under x86->ARM emulation.
The second thing to note is that this is not the standard CP2077 FHD test Notebookcheck runs and reports results for. For this test, there is a 30 minute warm-up prior to running this benchmark making this version of the benchmark closer to the end of an endurance test rather than a sprint. This is something I talk about on the main webpage. I’m still a little unclear exactly about the particulars of the test, but most processors suffer a 9-10% reduction in performance compared to Notebookcheck’s standard benchmark results for the same games and settings. There are exceptions however: the base M5 (not Pro or Max models) both binned and full doesn’t suffer any noticeable degradation compared to its standard CP2077 FHD Ultra benchmark results and the A18Pro actually improves on its standard FPS in the endurance test. How much of that is noise is unclear.
The third thing is that, for Apple, CP2077 tends to be on the lower end of performance relative to other chips. In other words, when comparing Apple to Nvidia or AMD or Intel across different benchmarks like Steel Nomad, CP2077, Assassin’s Creed Shadows, Baldur’s Gate 3, and Total War: Pharaoh - Apple tends to struggle in CP2077 more than the others (though to be fair to AMD, Steel Nomad’s performance on AMD GPUs is quite bad compared to AMD’s actual gaming performance, unsure why - Steel Nomad is not so unrepresentative with any other GPU).
(Fig 6) According to just the FP32 unit count (as well as the memory bandwidth), the new X2-90 GPU should be something in between the M5 Pro and base M5 GPU and roughly double the 8-core M5 GPU in size. However, according to silicon die shots, it is reported to be roughly the same size, in silicon, as what the 10-core M5 is likely to be. As noted in the powervsperformance website, different GPU manufacturers can achieve different levels of compaction in their design based on things like whether or not FP16 and Int have separate functional units, the amount of special math function units, the amount and type of cache, as well as the presence and design of matrix and ray tracing acceleration units. In this benchmark, the Qualcomm Adreno GPU behaves more similarly to the base M5 in performance and efficiency such that you could draw a pretty straight between the three points. However, given all the caveats mentioned earlier, this concordance with silicon size is more likely coincidence than not. The more standard CP2077 FHD Ultra test shows the Adreno X2-90 GPU to be more performant than in this benchmark here (whilst the base M5 stays the same) and of course there is the issue of x86 emulation vs native performance.
Against the 890m and B390, it is roughly on par or better than these GPUs, but it doesn't seem to scale as well as the B390, both of which have fewer FP32 units, but, again, don't labor under emulation.
(Fig 5, 6, and 7) As we move into larger GPUs, the Qualcomm GPU can’t really compete here with the 25% larger 5050m, 8060S, and 20-core M5 Pro GPUs. This despite the fact that the Qualcomm CPU is more a competitor to the Apple Pro-level and AMD Halo-level CPUs. The Qualcomm iGPU is fine when competing with smaller GPUs from Apple, AMD, and Intel, but is a mismatch with its large, powerful CPU. That might make sense if Qualcomm was ever optionally pairing their SOC with a larger dGPU from say Nvidia, but that hasn't happened and doesn't seem likely in the near future.
Fig. 7:
(Fig. 7) At this power level, in this benchmark, the Nvidia chip and the Apple M5 Pro share almost the same performance and efficiency. They both have 2560 FP32 units (Fig 8), though the Nvidia chip has slightly higher bandwidth (384GB/s vs 307GB/s). Again, despite the similarities of the two GPUs, this alignment of the resulting performance/efficiency is probably once again a bit coincidental: 1) the CPUs, which matter for a 1080p Ultra benchmark, are very different (Apple’s should be much more efficient than the Intel one in the PC), 2) this benchmark is one of Nvidia’s best, 3) Nvidia is on an older manufacturing node. The combination of these three factors all seem to cancel each other out. Different benchmarks and a different PC CPU would likely yield a different relationship.
Having said that, most of these Nvidia chips were paired with Intel CPUs and when doing so we see a pretty good concordance for this benchmark even beyond the M5 Pro and 5050m. The M4 Max and M5 Max chips** correspond roughly to the 5070m and 5070M Ti GPUs (when the latter are set to similar power draws) and which makes sense given that these Nvidia GPUs bracket the Apple ones in size (Fig. 9). Nvidia dGPUs can of course be set to much higher power draws with concomitant reductions in efficiency, but as shown here can be just as efficient as their Apple counterparts as well.
**not only is the M5 in the 14” thermally constrained, it also appeared to draw from the battery during the benchmark. I have kept the data point in Figure 7 (but not 8 or 9) to note that is what happened, but it may not be otherwise a very useful data point - unsure how much it was actually drawing from battery as the result is pretty similar to what one would expect just from thermal throttling, but there is a definite asterisk next to its result. For the M5 Max in the 16" MBP, no such battery drain was noted during the test but NBC never published a full review of the 16" model and so I had to estimate the idle power from previous 16" MBPs. Given how consistent Apple's designs are across the generations, I'm probably very close here and the effect on the resulting Power and Efficiency measures is likely minimal.
Fig. 8:
(Fig 8) Focusing on just the efficiency of the iGPUs, like with the CBR24 MT benchmark, we see a pretty clear target window for efficiency in the last two generations of Apple processors (when not thermally constrained).
In conclusion, Apple and Qualcomm are delivering solid GPUs, but in this 1080p raster-only gaming test, we don't see the advantage here that they enjoyed on the CPU-side. All 5 vendors deliver competitive gaming GPUs using very different architectures. While Nvidia has yet to deliver a consumer iGPU (a couple of SOCs are rumored to be coming soon), they enjoy sole ownership of the large mobile dGPU market with only Apple supplying a competitor, the Max, for its midrange options and Apple's Ultra reserved for the desktop Studio mini-workstations.
A small note on the deficiencies of this test as we get into higher GPU sizes:
Fig. 9:
(Fig 9) Because the test is run at 1080p Ultra, the benchmark will become increasingly CPU-bound as the GPUs gets larger. This can be demonstrated by plotting GPU size vs performance. I’ve drawn two lines: a solid line for a higher performance, lower efficiency setting for the Nvidia GPUs and a dashed line for the lower performance, higher efficiency setting. In both cases, we can see the efficiency of the Nvidia GPU drop as size increases, especially for the lower power/higher efficiency setting. Attempting to maintain efficiency meanwhile results in lower performance gains in the largest GPUs (curved lines). Looking at results on Notebookcheck's website from the standard benchmarks run from low settings to 4K meanwhile, one finds the larger GPUs increasing in their performance advantage over smaller ones, especially at 4K indicating they are finally being taxed enough at the higher resolutions for their size to matter. Further, not pictured here, but some of the 5080 (desktop) results paired with AMD X3D CPUs show a huge improvement in performance and efficiency relative to the Intel CPU tested in the 5080 (desktop) shown above. Finally, a possible contributory factor is that, starting with the 5070M Ti, both memory bus and GDDR7 speeds increase as the Nvidia GPUs gets bigger no doubt resulting in higher energy costs compared to lower end GPUs. If this is a contributory factor, I expect it to be a smaller one.