Remix.run Logo
▲ bhouston 3 hours ago

> Moreover, single-core CPU performance has largely stalled since the early 2000s, pushing people towards large multicore CPUs such as Intel Xeon, AMD Zen, and NVIDIA Vera.

Huh? Have you tried the Apple's A + M-series of processors? It has dramatically improved performance and tends to increase single core around 15% per year and it compounds.

The drive towards multi-core architectures is real but it isn't because single core performance has stalled, it is that we want to scale faster than single core performance increases.

These are basic facts that this article is getting wrong. How many more basic errors are in this?

▲AmazingTurtle an hour ago | parent | next [-]

> It has dramatically improved performance and tends to increase single core around 15% per year and it compounds.

Mostly because of SoC designs and dedicated/specialized compute units and versatile instructionsets yadda yadda

▲bhouston 12 minutes ago | parent [-]

> Mostly because of SoC designs and dedicated/specialized compute units and versatile instructionsets yadda yadda

Very true.

▲nottorp 2 hours ago | parent | prev [-]

> It has dramatically improved performance and tends to increase single core around 15% per year and it compounds.

Does it, or is that because of the large caches and memory bandwidth? Honest question.

▲bhouston 2 hours ago | parent | next [-]

> Does it, or is that because of the large caches and memory bandwidth? Honest question.

The burden to prove your "honest question" claim is on you, not me.

CPUs are incredibly complex and many things contribute to sustained improved performance year over year. To simplify it to just one or two things is showing a lack of understanding of the topic.

▲postalrat 33 minutes ago | parent [-]

You made the claim and refuse to back it up.

▲bhouston 13 minutes ago | parent [-]

> You made the claim and refuse to back it up.

I didn't make this claim that he wants me to prove: "Does it, or is that because of the large caches and memory bandwidth?"

He made that claim, not me.

▲hajile an hour ago | parent | prev [-]

Intel's Lion Cove core is 3.43mm2 without L2 cache and 4.53mm2 with private L2.

M5 S-core is 3.4mm2 including private L2.

But Intel's Lion cove is on N3B while M5 is on N3E which is about 10% less dense. do the math to scale up Lion Cove to N3E and it becomes around 5mm2.

That's 25-35% bigger for worse performance on the same node.

AMD could have gone with 3nm, but went with 5nm (N4P is 5nm++) instead. 3.72mm2 for the core and 4.77mm2 with private L2 cache.

N3E has the same size SRAM as N5/N4. Theoretical logic scaling is around 1.5x (216 MTr/mm² and 143.7 MTr/mm), but high-performance layouts don't get anywhere near that number due to the extra spacing needed for wires and transistors.

If we assume there's zero SRAM anywhere in the core (not true with large register files, L1, uop cache, etc), we get about 2.51mm2 for the core plus another 1.05mm2 (no cache shrink) for L2 giving 3.56mm2 (bigger than M5 S-core). With a more realistic shrink, we could be talking a total area of 3.9-4.2mm2.

TL;DR -- even in the absolute best-case scenario, both Zen5 and Lion Cove are bigger than Apple's M5 S-core.

▲nottorp an hour ago | parent [-]

Is it because the cores are more complex? They have to run x86-64 not arm instructions. I know they all run microcode internally, but maybe the translation layer is simpler on arm.

Honest question again <waves to whoever was ticked off by that>.

I could ask a LLM but HN could use some actual discussion.