Intel Shows Off Multi-Chiplet Sapphire Rapids CPU with HBM

I would presume because it turns out, in many cases, you have some tasks continuously running that don't require that much CPU, but a nonzero amount, and this way you can shove them on the LITTLE cores while the big core(s) go to town on single big tasks.

According to Intel's claims, the Gracemont cores have more than a half of the throughput for multi-threaded tasks of a big core in about a quarter of the area and a quarter of the power consumption.

That means that a Gracemont thread is supposed to have about the same throughput as a SMT thread on a big core.

Therefore an 8+8 cores Alder Lake with 24 hybrid threads should have about the same MT performance with a CPU having 12 big cores with 24 SMT threads, while having the area and the power consumption of only 10 big cores.

So Alder Lake is expected to have a similar MT performance with the 12-core Ryzen 9 5900X, while having a much better ST performance.

Of course that remains to be seen.

Early next year AMD should counter with Zen 3+ CPUs having a much larger cache memory and probably a little higher clock frequency. Those will probably have much better MT performance than Alder Lake, but Alder Lake should retain the best ST performance until Zen 4, late next year.

It's not a matter of absolute numbers, but a question of die area, to some degree. You can either spend X amount of area on a "big core" that can clock all the way down to idle (where it will spend most time) and all the way up to its max perf, or you can spend a fraction of X on a smaller core with a much, much smaller relative surface area, and get the same performance at idle or near-idle tasks. It just won't be able to scale "up" to max perf. But the net gain is you've saved a significant amount of area, and that translates to better TDP and thermals. The goal is to reap this benefit while also having "big cores" for the throttle-up phase, so you need both.

There's a lot of talk about "background tasks" that do light, but necessary processing in asymmetric multicore setups like this; for example on macOS Spotlight performs background indexing continuously, this is a great fit for a "small core" like in the M1. There's no amount of idle clocking you can do to match the perf-per-watt here. Similarly, there is also talk of "dark silicon" and "increasing heterogenous systems", like hybrid designs. But these are all the same problem, really: modern designs are not about having the most powerful components, but the correct blend of components to maximize the work-per-watt you can achieve. You can't just throw "more transistors" at every single design anymore and get free wins on all fronts 1 year later when transistors shrink.

Yes Thunderbolt is an Intel IP. You don't pay for a license fee, although I think they still need to buy the controllers from Intel if I recall correctly.

Intel seems more like Sun, they once produced competitive products but are now entering a period of stagnation and decline due to confusing product offerings and executive mismanagement. Most of their sales come from legacy entrenchment.

Disclaimer: I f***ing love AMD.

But... between mindshare, supply constraints, and OEMs being slow to make compelling products with AMD, it takes a long time to switch large numbers from Intel to AMD.

With AMD laptops, you'll struggle to find the ideal combination of display (resolution, quality, brightness), ports (Thunderbolt) and other attributes. They only got high-end discrete GPUs and 4K screens this year. And often not in the same machine. Thunderbolt is technically available for AMD platforms now, but still exceedingly rare.

For desktop, yes... I switched to AMD fifteen years ago and I haven't looked back. That's as someone who builds their own.

In the consumer space (outside DIY), desktops are like laptops. You need to find OEMs putting things together in the right configuration. Not to mention selling them at your local Best Buy for a mainstream consumer to come across it and pick it up. And not turn away because it's missing the "Intel Inside" sticker seared into their brain from decades of marketing.

Well the people designing the boards care. Many huge enterprise companies care (for reasons I don’t completely understand but I don’t work in that space any more). And AMD’s volume is nothing like Intel’s — one of Intel’s lucky breaks.

I think AMD is mostly a consumer brand and consumers aren’t buying a lot of PCs these days (except for a small transient due to Covid).

> Instructions per Clock matters too, but the market would have a hard time buying an x86 chip that stopped at 3 GHz, even if it was double the IPC.

AMD's response to the later-life Pentium 4's was the Athlon 64, which was clocked substantially lower but matched or beat the P4 in most tasks. The market responded positively.

Similarly, the original Pentium was clocked substantially lower than high end 486 models when it was first introduced. It was also received positively.

In modern times, the market has responded positively to the Apple M1.

Has there actually been a CPU anywhere that has outperformed its market competitor (in price and performance), but failed to gain traction because its clock speed was too low? "The market wants high clock speeds" has been a common assumption, but I've never actually seen it work out that way in practice.

The point is that whatever M1 capabilities are, they are largely irrelevant to any company besides Apple themselves.

Ferrari and Kia both make cars, but customer bases are separate.

Amazon ARM and Apple ARM are a class apart, and are not substitutable. Same ISA but wildly different architectural capabilities and limitations, a bit like Via x86 versus Intel x86 back in the day.

They never will, otherwise they will just be yet another OEM, instead the Ferrari/Bugati/Lamborghini of computers.

They don't want to be the computer people take for groceries, rather the one that drives slowly in the main avenue making everyone aware of their presence.

It depends how you look at it.

Around 50% of Windows PC are Cooperate Business PC. Those are not Apple's target customer.

But Apple is still gaining ground on the so called 90% market. And ARM M1 Chip already allow them to compete at a much more favourable price point. You see higher discount rate on M1 based Mac, and better Education Discount on M1 MacBook Air, along with SKUs hitting $799.

They have plenty of margin to play with and get an official MacBook down to $799 within the next 3 years.

They may never out sell PC in unit volume, but they still have plenty of room to grow. In an optimistic projection I would not be surprised if Apple double their Mac active user by 2031.

The user of Macs, certainly, because there is only one playing field, where Apple places it.

Apple is entirely not represented on the AMD and Intel key playing field, datacenter-grade chips. OTOH it's very much present (and more with M1) in the Qualcomm and Samsung playing field, ARM-based mobile chips, from which AMD and Intel are both absent.