Posted by shioyama 10 hours ago
This ISA seems to be designed to: 1. make the automatic parallelization easy (with workload split between compiler and hardware) 2. make communication costs between tiles very small, so that even very simple workloads can be automatically parallelized.
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?
Mostly because of SoC designs and dedicated/specialized compute units and versatile instructionsets yadda yadda
Very true.
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.
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.
The fact that single core performance has been improving consistently in a compounding fashion in the Apple series of M and A chips is a historical fact.
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.
Honest question again <waves to whoever was ticked off by that>.
I could ask a LLM but HN could use some actual discussion.
Apple doesn’t need uop cache while x86 designs do which means more power and area.
ARM has looser memory ordering which improves performance. x86 can do this too, but they have to add an entire memory speculator and all the circuitry to handle when the predictions are wrong. Again, that’s extra area and power.
More controversially, x86 decode is still a significant part of the core area. ARM A715 cut decoder area by 75% just by dropping ARMv7 support (and it’s nowhere near as complex as x86). If the decoder is just 5% of the core area (and basically always on), cutting that by 75% means reducing core size and power by a significant amount. I’d also note that a company doesn’t choose complex 2x4 or 3x3 decoder implementations with massive complexity and inferior worst-case performance unless something is forcing them.
The situation isn’t quite as bad is it might otherwise seem though because 512-bit SIMD as it adds to core size (though arguably doesn’t add very much to most workloads over 128-bit aside from some great new instructions). We can calculate this with Strix Point which uses 256-bit “double pumped” SIMD. The core plus L2 area drops to about 4.2mm2 (or about 3.1mm2 for just the core). Using perfect 1.5x scaling for the core and no scaling for the L2 gives around 3.2mm2. Assuming more reasonable 1.3x scaling (which might still be too generous) gives 3.5mm2 which is still bigger for worse performance on hypothetical N3E vs M5 S-cores.
Even in the best scenario, x86 is way behind in total performance, perf/watt, and area. If we assume the engineers at AMD and Intel are good at their job (a good assumption given how long they’ve been beating everyone else on the planet), it only leaves the ISA holding them back complicating things with legacy bloat.
You’ll see a lot of pushback to this idea, but never much of substance to the assertion.
How many times has a bold and maybe even exaggerated claim come out and people respond “this is false hype”, or “this is misleading”, or “ehh, only kind of in some technical sense, but don’t worry”. Only for continual iteration and breakthrough to have all of us agreeing one week later that it is indeed doing whatever thing and even that this is old news now.
If the frequency increases, that’s the intelligence explosion…
Truly the last invention (for humans)