3 ms·
Intel does not appear to have any serious problem in their capabilities of logic design (even if e.g. the launch of Sapphire Rapids had multiple delays due to d
by adrian_b 4y ago
Intel does not appear to have any serious problem in their capabilities of logic design (even if e.g. the launch of Sapphire Rapids had multiple delays due to design bugs and insufficient validation before the tape-outs), but until now there has been no evidence of Intel recovering their great CMOS manufacturing process disadvantage vs. TSMC.
The Intel Sapphire Rapids CPUs would have been superior to the AMD Genoa, except that the CMOS process used by Intel to make them provides such functional dependencies between the clock frequency and the power-supply voltage and between the power consumption and the power-supply voltage, that whenever the CPU in in a state where the clock frequency is limited by the power consumption, i.e. whenever it runs a multi-threaded application, the clock frequency of a Sapphire Rapids CPU is ridiculously low in comparison with that of an AMD CPU made by TSMC.
This leads to low performance in multi-threaded applications. Moreover, the Intel CMOS process has low yields, which forces them to make only Sapphire Rapids CPUs with fewer cores, smaller cache memories and higher prices than AMD Genoa.
As long as Intel remains handicapped by their inferior manufacturing processes, they will not be able to make competitive products.
At the end of this year, Intel is expected to launch Meteor Lake the first product made (partially, the integrated GPU will be made by TSMC) using the next Intel manufacturing process.
This is their chance to regain some of the lost ground. It remains to be seen if they will succeed.
Until now, the next Intel manufacturing process (Intel 4) does not seem to be in very good shape, because Intel had to cancel the desktop variant of Meteor Lake and replace in with a "Raptor Lake Refresh", presumably because "Intel 4" does not provide yet high enough clock frequencies, exactly like Intel had to replace for desktops Ice Lake with Comet Lake and Tiger Lake with Rocket Lake.
- locustous 4y agoMy understanding is that the process is moving forward quite well, it's the design teams that are struggling to deliver on time. Meteor lake is a big step up in efficiency but will be a regression in performance. Another reason to cancel the desktop variant.
- analognoise 4y agoIn a vacuum that would be a win. AMD is executing on performance and efficiency simultaneously; a perf regression on Intel parts would be a disaster.
- adrian_b 4y agoBecause Meteor Lake is expected to have a greater IPC than Raptor Lake, reducing the gap towards the Apple CPUs, any regression in performance can be caused only by a clock frequency during a power-limited scenario that is lower than estimated when the Meteor Lake design has started, which is the same what has happened with the Intel "10 nm" process during the Cannon Lake/Ice Lake/Tiger Lake years, until Intel eventually succeeded to tweak their manufacturing process and design rules enough to be able to obtain a better performance than with their old process. This means that Intel 4 is still not moving forward well enough, at least not as well as they had predicted at the start of the Meteor Lake project. Hopefully for a healthy competition, this time they will succeed much faster to tune their manufacturing process.
- locustous 4y agoIntel has had process problems. I don't think those continue to be their primary issues. It wasn't process problems that kept sapphire rapids from being released till now. 10nm/Intel 7 has been ramped up and fabbing products for over a year. They are very late on critical products that are clearly not the fault of the foundry side. This is squarely on the design side. Running good engineering orgs is hard, Intel's problems run deep and on many levels.
- ksec 4y agoFew extra notes. The original Sapphire had to be "redesigned" ( or reworked ) for Intel 7. You cant just take an high yield Intel 7 Golden Cove Core and "Edit" it with some extra features for server and then expect it yield. Second being Intel 7 isn't exactly high yield either. At least not 14nm level. It is good enough for small die size chips, but it is something different for large die size ( Aka Sapphire Rapid ). Which is why the comparison between Intel and AMD on server market doesn't make sense. They have a very different economy of scale. Third being Intel doesn't put resources in Intel 7 any more. It is no longer a priority. They are losing one way or another that is why they put everything into their next node Intel 4. Intel is expecting Intel 3 in 2023 and products out in 2024, this isn't too far behind TSMC schedule.