2 ms·

Honestly I'm surprised a 62% single-core improvement is all they can muster from TSMC's process improvements. In the same time period, Intel managed a 27.5% im

Honestly I'm surprised a 62% single-core improvement is all they can muster from TSMC's process improvements. In the same time period, Intel managed a 27.5% improvement in single core performance. Intel went from Alder Lake (Nov 2021, Intel 10nm) to Arrow Lake (Oct 2024, TSMC 3nm). [0] This is not true and not how heterogeneous core architectures or single-threaded benchmark scores work. This is true based on Geekbench 6. [0]https://browser.geekbench.com/processor-benchmarks https://browser.geekbench.com/processor-benchmarks 12900KS vs 285k

bigyabai 2y ago
When you mean "improvement" you mean Geekbench score. There are all sorts of metrics that year-over-year improvements bring, and performance uplift is really just one of them. > This is true based on Geekbench 6. But Apple Silicon does not have a homogeneous core architecture. Your claim that "each core" was improved by 62% is simply flat-out wrong. A single-core test does not average out the performance of the efficiency core and performance core, it just schedules it to the fastest available thread. Furthermore, you're again conflating "improvement" and "Geekbench score" without controlling for variables like power consumption and ISA divergence. It's like the cherry-picked PowerISA vs Intel benchmarks all over again. OP's statement still strikes me as overwhelmingly true. The promise of ARM dominating the desktop is entirely unrealized in 2025. It's barely even reaching fruition in the datacenter.
- aurareturn 2y ago
  When you mean "improvement" you mean Geekbench score. There are all sorts of metrics that year-over-year improvements bring, and performance uplift is really just one of them. I'm not sure what this has to do with anything I said. But Apple Silicon does not have a homogeneous core architecture. Your claim that "each core" was improved by 62% is simply flat-out wrong. A single-core test does not average out the performance of the efficiency core and performance core, it just schedules it to the fastest available thread. Furthermore, you're again conflating "improvement" and "Geekbench score" without controlling for variables like power consumption and ISA divergence. When I said "each core", I meant each P core. Geekbench does not test E cores in ST. In MT, E cores are factored in. OP's statement still strikes me as overwhelmingly true. The promise of ARM dominating the desktop is entirely unrealized in 2025. It's barely even reaching fruition in the datacenter. Weird statement devoid of any facts. Apple's laptops are generations better than anything AMD and Intel can muster in 2025. Even Qualcomm's first generation Nuvia core laptops are matching AMD and Intel's very best in CPU benchmarks. Desktop is a small market. In data centers, Graviton has had a meteoric rise.[0] Every major cloud company has its own ARM server CPU including Meta, Google, Microsoft, Tencent, Baidu, Alibaba, Nvidia. [0]https://x.com/patrickmoorhead/status/1863986630366757279?s=46 https://x.com/patrickmoorhead/status/1863986630366757279?s=4...
  - bigyabai 2y ago
    > I'm not sure what this has to do with anything I said. I'm pointing out that your goalposts are arbitrary. We could compare a NEON vs AVX benchmark if we wanted to watch ARM have a meltdown, but we don't. I'd hope you're smart enough to know the composite improvements of x86 are greater than the sum of it's IPC uplift. > When I said "each core", I meant each P core. Ah, thank you. That's much more clear. > Weird statement devoid of any facts. Apple's laptops are generations better than anything AMD and Intel can muster in 2025. Yeah, and their desktops? > Even Qualcomm's first generation Nuvia core laptops are matching AMD and Intel's very best in CPU benchmarks. Once again, I'm sorta dumbfounded that I have to explain this to HN. IPC comparisons are not a direct shootout on CPU performance! They do not reflect real-world workloads! Comparing ARM benchmarks to x86 ones is literally the exact same problem as Apple's infamous Intel vs PowerISA benchmarks.
    - aurareturn 2y ago
      I'm pointing out that your goalposts are arbitrary. We could compare a NEON vs AVX benchmark if we wanted to watch ARM have a meltdown, but we don't. Ah yes. Since Apple Silicon chips perform so much better in single thread than AMD and Intel chips in popular benchmarks like Geekbench 6, Cinebench 2024, and Speedometer, people have started to claim that niche applications accelerated by AVX512 matter a lot more than they really should. I'd hope you're smart enough to know the composite improvements of x86 are greater than the sum of it's IPC uplift. If it's slower in Geekbench 6, and SPEC, then its ST is slower to me. If you want to buy chips based on specific applications such as the need for AVX512, then go for it. Overall ST performance still goes to Apple. Then again, Apple sells SoCs and not standalone CPUs so we should really be factoring all the ways an Apple Silicon chip can accelerate applications. PS. IPC uplift really doesn't mean as much as total ST uplift. You can have a CPU architecture that is great for IPC but can't clock high and vice versa. Only total ST performance matters. Yeah, and their desktops? I like how you said desktops - as if Apple wanted to dominate desktops. Desktops is not Apple's target audience. Yes, they have the Studio and Mac Pro, which they care little about. Quite frankly, an M4 Max is a desktop replacement since its CPU is basically a 9950x and desktop RTX 4070. Comparing ARM benchmarks to x86 ones is literally the exact same problem as Apple's infamous Intel vs PowerISA benchmarks. They run mostly the same applications with a few exceptions. There is nothing wrong with comparing benchmarks between x86 vs ARM.

bigyabai 2y ago

When you mean "improvement" you mean Geekbench score. There are all sorts of metrics that year-over-year improvements bring, and performance uplift is really just one of them. > This is true based on Geekbench 6. But Apple Silicon does not have a homogeneous core architecture. Your claim that "each core" was improved by 62% is simply flat-out wrong. A single-core test does not average out the performance of the efficiency core and performance core, it just schedules it to the fastest available thread. Furthermore, you're again conflating "improvement" and "Geekbench score" without controlling for variables like power consumption and ISA divergence. It's like the cherry-picked PowerISA vs Intel benchmarks all over again. OP's statement still strikes me as overwhelmingly true. The promise of ARM dominating the desktop is entirely unrealized in 2025. It's barely even reaching fruition in the datacenter.

aurareturn 2y ago
When you mean "improvement" you mean Geekbench score. There are all sorts of metrics that year-over-year improvements bring, and performance uplift is really just one of them. I'm not sure what this has to do with anything I said. But Apple Silicon does not have a homogeneous core architecture. Your claim that "each core" was improved by 62% is simply flat-out wrong. A single-core test does not average out the performance of the efficiency core and performance core, it just schedules it to the fastest available thread. Furthermore, you're again conflating "improvement" and "Geekbench score" without controlling for variables like power consumption and ISA divergence. When I said "each core", I meant each P core. Geekbench does not test E cores in ST. In MT, E cores are factored in. OP's statement still strikes me as overwhelmingly true. The promise of ARM dominating the desktop is entirely unrealized in 2025. It's barely even reaching fruition in the datacenter. Weird statement devoid of any facts. Apple's laptops are generations better than anything AMD and Intel can muster in 2025. Even Qualcomm's first generation Nuvia core laptops are matching AMD and Intel's very best in CPU benchmarks. Desktop is a small market. In data centers, Graviton has had a meteoric rise.[0] Every major cloud company has its own ARM server CPU including Meta, Google, Microsoft, Tencent, Baidu, Alibaba, Nvidia. [0]https://x.com/patrickmoorhead/status/1863986630366757279?s=46 https://x.com/patrickmoorhead/status/1863986630366757279?s=4...
- bigyabai 2y ago
  > I'm not sure what this has to do with anything I said. I'm pointing out that your goalposts are arbitrary. We could compare a NEON vs AVX benchmark if we wanted to watch ARM have a meltdown, but we don't. I'd hope you're smart enough to know the composite improvements of x86 are greater than the sum of it's IPC uplift. > When I said "each core", I meant each P core. Ah, thank you. That's much more clear. > Weird statement devoid of any facts. Apple's laptops are generations better than anything AMD and Intel can muster in 2025. Yeah, and their desktops? > Even Qualcomm's first generation Nuvia core laptops are matching AMD and Intel's very best in CPU benchmarks. Once again, I'm sorta dumbfounded that I have to explain this to HN. IPC comparisons are not a direct shootout on CPU performance! They do not reflect real-world workloads! Comparing ARM benchmarks to x86 ones is literally the exact same problem as Apple's infamous Intel vs PowerISA benchmarks.
  - aurareturn 2y ago
    I'm pointing out that your goalposts are arbitrary. We could compare a NEON vs AVX benchmark if we wanted to watch ARM have a meltdown, but we don't. Ah yes. Since Apple Silicon chips perform so much better in single thread than AMD and Intel chips in popular benchmarks like Geekbench 6, Cinebench 2024, and Speedometer, people have started to claim that niche applications accelerated by AVX512 matter a lot more than they really should. I'd hope you're smart enough to know the composite improvements of x86 are greater than the sum of it's IPC uplift. If it's slower in Geekbench 6, and SPEC, then its ST is slower to me. If you want to buy chips based on specific applications such as the need for AVX512, then go for it. Overall ST performance still goes to Apple. Then again, Apple sells SoCs and not standalone CPUs so we should really be factoring all the ways an Apple Silicon chip can accelerate applications. PS. IPC uplift really doesn't mean as much as total ST uplift. You can have a CPU architecture that is great for IPC but can't clock high and vice versa. Only total ST performance matters. Yeah, and their desktops? I like how you said desktops - as if Apple wanted to dominate desktops. Desktops is not Apple's target audience. Yes, they have the Studio and Mac Pro, which they care little about. Quite frankly, an M4 Max is a desktop replacement since its CPU is basically a 9950x and desktop RTX 4070. Comparing ARM benchmarks to x86 ones is literally the exact same problem as Apple's infamous Intel vs PowerISA benchmarks. They run mostly the same applications with a few exceptions. There is nothing wrong with comparing benchmarks between x86 vs ARM.