Intel’s hope for recovery reminds me of Apple’s back in 97. They have one asset which is (greatly diminished but nonzero) momentum. An absolutely enormous number (though slowly diminishing) number of people can’t switch away from their product. So they have time to try and do something new.
Apple’s had a much shorter fuse. They really needed that $150MM from MS (who needed Apple to fend off antitrust). Talking to ppl in Intel that sense of urgency hasn’t made it through the organization (and why should it? There are tons of other jobs around). But they could pull it off.
The momentum is what allows you to avoid the (mythical?) “Osborne effect”: the only thing Intel can brag about is the future; today’s products will still be bought by a lot of people and so hopefully that will stem some of the design losses.
MS was able to do the same thing (though their situation was far from dire). Motorola and HP couldn’t.
I agree that their current products aren't competitive, but I'd disagree with your multiple years out remark. The Alder Lake platform is out Q1 next year. PCIe Gen 5, DDR5, hybrid cores, the new Xe graphics architecture, and the first new node process in over 6 years (Intel 7). From what I've seen from early leaks, it looks like it has the potential to be a flagship CPU for the first time in a long time.
That is partly true and they have been doing this for two years now. Regular readers would have notice a lot more leaks coming out from website and tweets than usual. And I think their sales and marketing department did an exceptional job during Intel's darkest hour.
Not that is matter much because both Intel and AMD was constrained by capacity.
To me, it looks like Intel is playing catch up, and trying every heroic hack like HBM that they can think of in order to be competitive with AMD.
It reminds me a lot of the last days of Netburst, with heroic hacks like huge (for the time) caches, high clock speeds, and then bumping up against TDP limits.
This product is competing against NVidia for the deep learning space, not AMD. NV’s A100 features enormous memory bandwidth (1.4 TB/s) thanks to HBM and peak flops (300 TFlops/s) thanks to tensor cores, so using HBM and Advanced Matrix Extensions here is pretty clearly aimed at that segment.
The P4 high clock speeds were not a hack, they were the root of the problem. High clock speeds were an architectural constraint driven by the _marketing_ department. Intel's architect knew that this wasn't the way to go for actual performance but delivered what was asked of them.
Golden cove core highlights[1] include a 6 wide decoder (up from 4; M1 is 8), reorder buffer of 512 (up from 352 in Ice/Tiger/Rocket lake, or 224 in Skylake; M1 is estimated around 650), and 12 execution units (up from 10 in Ice/Tiger/Rocket lake or 8 in Skylake).
Still not as wide as the M1, but it looks like a huge step in the right direction on paper.
Apple CPUs have a much better energy efficiency than either Intel or AMD, in part due to the improved TSMC 5-nm process and in part because the Apple CPUs are more brainiac, doing 50% more work per clock cycle at a 2/3 times lower clock frequency than the Intel/AMD CPUs.
However, for absolute performance values, the desktop Intel and AMD CPUs remain faster than even M1 Max, both in single-thread and in multi-thread tasks.
M1 Max should be able to match the best laptop CPUs in some tasks and be faster in the tasks that are limited by the memory bandwidth.
On the other hand, when the laptop Alder Lake CPUs will be introduced in Q1 2022, they should be much faster in ST than M1 Max or M1 Pro, but while consuming a much higher power.
Apple only just took the multithreaded performance crown on a laptop with the M1 Pro & M1 Max, neither of which are actually available yet. And neither the M1 Pro nor M1 Max are competitive with AMD's top-end consumer desktop CPUs at multicore performance (ie, Ryzen 5900X & 5950X), to say nothing of the prosumer ones (aka, Threadripper). Which given the differences in power budget isn't really that surprising, but still.
To call that "lapping" seems a bit overly extreme.
Also this article is about server CPUs, so I don't know why the M1 is being mentioned at all. It would get humiliated in a server environment by existing Xeons & Epyc CPUs, to say nothing of what Intel is demoing here.
Until Apple CPUs run on Amazon, Azure and Google clouds, it doesn't really matter how much better they are, other than 10% worldwide desktop users, and some tier 1 markets where iOS sales take place.
It's very important to remember that Apple is always in the bleeding edge of process nodes. In this case, TSMC's 5nm, which is the most advanced process in shipping products right now. AMD is still in 7nm but will be moving later, and Intel is in its own 10nm, or I guess "Intel 7" now.
I am completely not surprised by this move, HBM2 solves a major issue for quite a few applications that live and die on high speed memory and assuming sapphire rapids has enough ports to make use of it optimally will make these an absolute beast. I guess I'm only surprised AMD didn't do this first as they have more experience with HBM2 on their GPUs. Then again they might have and never made it a public SKU.
When Intel salespeople first started showing designs like that I asked them just this, and just like the HN crowd, they didn't seem terribly amused. I did, though.
[+] [-] kfprt|4 years ago|reply
[+] [-] gumby|4 years ago|reply
Apple’s had a much shorter fuse. They really needed that $150MM from MS (who needed Apple to fend off antitrust). Talking to ppl in Intel that sense of urgency hasn’t made it through the organization (and why should it? There are tons of other jobs around). But they could pull it off.
The momentum is what allows you to avoid the (mythical?) “Osborne effect”: the only thing Intel can brag about is the future; today’s products will still be bought by a lot of people and so hopefully that will stem some of the design losses.
MS was able to do the same thing (though their situation was far from dire). Motorola and HP couldn’t.
[+] [-] Aromasin|4 years ago|reply
[+] [-] ksec|4 years ago|reply
Not that is matter much because both Intel and AMD was constrained by capacity.
[+] [-] drewg123|4 years ago|reply
It reminds me a lot of the last days of Netburst, with heroic hacks like huge (for the time) caches, high clock speeds, and then bumping up against TDP limits.
[+] [-] bertr4nd|4 years ago|reply
[+] [-] speed_spread|4 years ago|reply
[+] [-] celrod|4 years ago|reply
Still not as wide as the M1, but it looks like a huge step in the right direction on paper.
[1] https://en.wikipedia.org/wiki/Golden_Cove
[+] [-] kingosticks|4 years ago|reply
[+] [-] thanatos519|4 years ago|reply
[+] [-] adrian_b|4 years ago|reply
However, for absolute performance values, the desktop Intel and AMD CPUs remain faster than even M1 Max, both in single-thread and in multi-thread tasks.
M1 Max should be able to match the best laptop CPUs in some tasks and be faster in the tasks that are limited by the memory bandwidth.
On the other hand, when the laptop Alder Lake CPUs will be introduced in Q1 2022, they should be much faster in ST than M1 Max or M1 Pro, but while consuming a much higher power.
[+] [-] kllrnohj|4 years ago|reply
To call that "lapping" seems a bit overly extreme.
Also this article is about server CPUs, so I don't know why the M1 is being mentioned at all. It would get humiliated in a server environment by existing Xeons & Epyc CPUs, to say nothing of what Intel is demoing here.
[+] [-] pjmlp|4 years ago|reply
[+] [-] nine_k|4 years ago|reply
[+] [-] LtdJorge|4 years ago|reply
[+] [-] leeter|4 years ago|reply
[+] [-] buildbot|4 years ago|reply
[+] [-] schmorptron|4 years ago|reply
[+] [-] ngcc_hk|4 years ago|reply
[+] [-] baybal2|4 years ago|reply
[+] [-] simooooo|4 years ago|reply
[+] [-] formerly_proven|4 years ago|reply
[+] [-] kevingadd|4 years ago|reply