Leaked Intel Rocket Lake tests show double digit perf gains

[rave_alan]

The size matters, bigger chip - better .

[badass]

But by they time they have EUV working, both Samsung and TSMC will have had many years with the technology in production and that's why at least TSMCs roadmap for new fabrication shrinkages seem to be so aggressive, it's nuts. If Intel doesn't ramp up and TSMC keeps their rate of acceleration going, Intel 7nm will be competing with TSMC 3nm of which Intels 7nm was meant to be feature similar to TSMCs 5nm.

Whilst Intel's transistor densities have generally been close to the next node down from TSMC, the performance from a user perspective has not - i.e. transistor performance.

From a customer perspective the more transistors crammed in the same space makes for potentially cheaper chips if yields are the same assuming adequate competition. However those intel 10nm +++++++++123ABC transistors do not perform better than TSMC 7nm ones. There are of course lots of caveats to this as various nodes can be tuned for performance under different power characteristics. e.g. Low power tuned for mobile means lower peak frequencies for example. Intel's 10nm Superfin is clearly tuned for performance at "higher" power.

Another way of looking at this: Intel's advantage at the "same" node would be cost to Intel to produce chips. But their yields are so bad there's no saving.


BTW I've used lots of quotes as nodes are just a load of marketing BS these days and have been for a long time now.

[Tabbykatze]

Whilst Intel's transistor densities have generally been close to the next node down from TSMC, the performance from a user perspective has not - i.e. transistor performance.

From a customer perspective the more transistors crammed in the same space makes for potentially cheaper chips if yields are the same assuming adequate competition. However those intel 10nm +++++++++123ABC transistors do not perform better than TSMC 7nm ones. There are of course lots of caveats to this as various nodes can be tuned for performance under different power characteristics. e.g. Low power tuned for mobile means lower peak frequencies for example. Intel's 10nm Superfin is clearly tuned for performance at "higher" power.

Another way of looking at this: Intel's advantage at the "same" node would be cost to Intel to produce chips. But their yields are so bad there's no saving.


BTW I've used lots of quotes as nodes are just a load of marketing BS these days and have been for a long time now.

Oh absolutely on everything you've said especially the node BS marketing.

I wonder if the major issue is not necessarily the transistor quality but the actual architectures themselves.

Take Zen 1 to Zen 3, there is a nominal 89% performance increase in 4 years, 4 years! If we look over to Intel they have had around 15-20% performance increase in the same time period (6700k to 10900k flagship to flagship) and the vast majority of that was realistically just frequency uptick. Some of AMDs increase was frequency bumping but the lions share was IPC.

So is Intels problem more architectural design team problem mixed in with the fact their nodes after 14nm are not very performant?

[six_tymes]

impressive considering the core count. I am still holding off for socket 1700 + DDR5

[kompukare]

impressive considering the core count.

Surely core count has nothing (or everything to do with it).
The article states:

The single threaded CPU benchmarks in the screenshots are "roughly 12 per cent higher than typical CPU-Z Bench single-thread numbers for the current-gen i9-10900 (non-K) and i9-10900K Comet Lake-S processors," notes TPU, the publisher of CPU-Z.

So yes, 6 cores clocked extra high because 14nm won't allow them to squeeze more in without affecting their capacity and the max frequency.

[six_tymes]

I think Intel should figure out a way financially to skip 7nm and go right to 5. 5nm has been reported to provide much better yields than 7. and we all know that is what Intel has needed for years, better yields.