3 ms·
It is estimated that the M1 is 16-billion transistors on 5nm for a 4-big core + 4-little core + iGPU + neural engine SOC. The most comparable laptop processor
by dragontamer 6y ago
It is estimated that the M1 is 16-billion transistors on 5nm for a 4-big core + 4-little core + iGPU + neural engine SOC.
The most comparable laptop processor is AMD Renoir: Zen2 8-core + iGPU for 10-billion transistors.
As such, I'd expect a "hypothetical high end" 32-big core Mac to compete against a 64-big core EPYC in terms of #transistors. Would the high end market be interested in a lower core count computer but higher single-thread performance?
I somehow doubt that. Server / workstation loads scale well to many cores: indeed, any computer that actually scales to 32-cores would probably also scale well to 64-cores (or SIMD-compute), and the M1 is only 128-bit wide SIMD.
------------
There's also the issue of the lack of hyperthreading: the Apple M1 does NOT implement SMT in any way what so ever. Those 32-cores would likely only support 32-threads, while the 64-core EPYC supports 128-threads.
I dunno. "Just scaling up" the M1 doesn't seem like it'd be a winner in the high-end market to me. Maybe if the core gained a wider SIMD-width and SMT (2x threads/core), that'd help a lot.
- toast0 6y agoThere's value in having the same compute ability in fewer cores. Communication/synchronization overhead is usually growing faster than linear in the number of threads. SMT isn't an unalloyed good, it helps for many workloads and hurts for some. That said, the very wide cores in M1 are asking for SMT to increass utilization, I'd guess it'll happen eventually. OTOH, the weaker memory consistency rules on ARM seem to allow for a lot more out of order execution, which seems to keep things busy. I don't think we'll see Intel and AMD increase the width of their cores so much; they need their cores to scale to 5Ghz (even if it takes 1KW of cooling) for marketting and competition reasons, and wider cores are harder to clock so high.
- dragontamer 6y ago> There's value in having the same compute ability in fewer cores. Communication/synchronization overhead is usually growing faster than linear in the number of threads. I agree for sure. But the tradeoffs are pretty steep. Scaling within a core is traditionally sqrt(size), while scaling number of cores on a problem is also sqrt(cores). Balancing between size-of-core vs number-of-cores is a big question. And that also depends very strongly on the task at hand. Web-servers scale better than sqrt(#cores), because each worker-thread is basically fully independent. > I don't think we'll see Intel and AMD increase the width of their cores so much; they need their cores to scale to 5Ghz (even if it takes 1KW of cooling) for marketting and competition reasons, and wider cores are harder to clock so high. I think there's a good argument that AMD / Intel can't scale their decoder much larger than 4-way / 6-way with uOp cache, due to the instruction set issue. But I've seen some interesting innovations: the most recent Intel Atoms have double-decoders (2+2 decoder) on SMT-based Atoms. Such a thing would make SMT more efficient without much increase to power. Its strange to see Apple take the single-thread performance crown. But... I don't think they've proven themselves in the high-end server / workstation market yet. Even if they do scale up, their decisions don't seem to be an obvious winner at the high end.
- vbezhenar 6y agoThey'll sell neural engine, video coding acceleration and other things like that.
- dragontamer 6y ago> They'll sell neural engine Is there even an API to take advantage of that yet? It looks like an impressive DSP but without an API, then that stuff is locked to Apple-only software / hardware. > video coding acceleration Is that really better than another 96-threads that you'd get with a 64-core Threadripper Pro? The thread-deficit of 32-core Apple vs 64-core/128-thread Threadripper is going to be pretty apparent. When you have another +96 threads available, I'm not sure if "video coding acceleration" is even needed. Just run that in software.
- EricE 6y ago"Is there even an API to take advantage of that yet?" Has been for at least three years now (maybe more): https://developer.apple.com/machine-learning/core-ml/ https://developer.apple.com/machine-learning/core-ml/
- dragontamer 6y agoAh right... I guess whenever I look at that neural-engine's specs, I'm actually personally more interested in the VLIW DSP they got in there. VLIW-DSPs can calculate a lot of things, more than just neural-net training. I guess that API is good enough for the deep-learning hype train however.
- EricE 6y agoMetal also leverages the neural engine. It's transparent, of course. I can't find the reference but it may have been one of the original Metal Apple Developer Conference (ADC) sessions where they let that slip. Of course other frameworks could leverage it too. As for direct access, I'm not sure. I think I did mix up API and framework in my response :)
- 6y ago