News It Turns Out, Intel's Fastest Rocket Lake-S Chips Will Boost To 5 GHz After All

PCWarrior

Distinguished
May 20, 2013
110
29
18,610
0
1500 single score at 5Ghz? it is DOA, even skylake make much more.
????? You are probably looking at the top 10 overclocked results. Easy trap. For the record the 10900K normally does between 1400-1500.
https://browser.geekbench.com/v5/cpu/search?utf8=✓&q=10900k

This does 1507. Also let’s not forget that this is an engineering sample, who knows how early one and with what other system specs (RAM frequency etc) it was tested. We don’t know the boost speed at which this single-core test was actually ran at (the information you read under processor information is data logged from the chip’s ID or motherboard, not what the software measures during the benchmark). Also being an engineering sample we don’t know architecturally how final it is in terms of critical path optimisations etc.

Regardless this shows one thing. That Intel is targeting 5GHz with Rocketlake and that there is a working 8-core/16-thread sku. Still no signs for a 10-core Rocketlake so apparently the rumours are true that the Rocketlake flagship is going to top at 8c/16t. Even then I expect the Rocketlake 8-core to beat the 10-core 10900K even in multithreaded workloads but in most such instances it will probably be a less than 5% victory.
 
Feb 27, 2020
40
6
35
0
????? You are probably looking at the top 10 overclocked results. Easy trap. For the record the 10900K normally does between 1400-1500.
https://browser.geekbench.com/v5/cpu/search?utf8=✓&q=10900k

This does 1507. Also let’s not forget that this is an engineering sample, who knows how early one and with what other system specs (RAM frequency etc) it was tested. We don’t know the boost speed at which this single-core test was actually ran at (the information you read under processor information is data logged from the chip’s ID or motherboard, not what the software measures during the benchmark). Also being an engineering sample we don’t know architecturally how final it is in terms of critical path optimisations etc.

Regardless this shows one thing. That Intel is targeting 5GHz with Rocketlake and that there is a working 8-core/16-thread sku. Still no signs for a 10-core Rocketlake so apparently the rumours are true that the Rocketlake flagship is going to top at 8c/16t. Even then I expect the Rocketlake 8-core to beat the 10-core 10900K even in multithreaded workloads but in most such instances it will probably be a less than 5% victory.
Only intel fanboy can be excited for result like that. 8/16 > 10/20 skylake? This is joke, even 3600 make more score than this in multi, don't be fooled.
 

bit_user

Splendid
Ambassador
????? You are probably looking at the top 10 overclocked results.
Yeah, I was worried about how to avoid picking one. It'd be nice if they had like a histogram of the scores for a particular CPU model, so you could pick one at mode.

For the record the 10900K normally does between 1400-1500.
https://browser.geekbench.com/v5/cpu/search?utf8=✓&q=10900k
Here's a comparison of the mystery chip with a i9-9900K:


The mystery chip is about 6.5% faster, single-threaded.

Also being an engineering sample we don’t know architecturally how final it is in terms of critical path optimisations etc.
I'm pretty sure engineering samples have all the silicon-level optimization done. There's not another round of test chips, after engineering samples, which is what you'd need if you were going to make a bunch of optimizations in the datapath.

Even then I expect the Rocketlake 8-core to beat the 10-core 10900K even in multithreaded workloads but in most such instances it will probably be a less than 5% victory.
Well, if the 6.5% single-thread performance boost were carried over to multi-core performance, then the performance on fully-threaded workloads would be 85.2% as fast as Comet Lake i9-10900K. Of course, that 6.5% number could be hampered by slow RAM. Also, it might be that Rocket Lake is even better at hyper-threading, in which case you'd get a per-core boost of more than 6.5%.

Still, for a multi-core improvement of 5%, you'd need a per-thread improvement of over 31%. That's just unrealistic.
 

InvalidError

Titan
Moderator
Still, for a multi-core improvement of 5%, you'd need a per-thread improvement of over 31%. That's just unrealistic.
Since Skylake -> Willow Cove Redux goes from 8 execution units to 10 execution units, the ideal SMT scaling would be about 25% more throughput for a given clock frequency excluding major scheduler improvements beyond increasing the number of execution ports it can manage. Of course, there will be some collateral SMT gains from reduced execution unit contention between threads, deeper re-order queue and other improvements throughout the pipeline improving execution unit packing a handful of points beyond Skylake. A ~30% SMT throughput per core gain should be within the realm of plausible for some workloads.
 

bit_user

Splendid
Ambassador
Since Skylake -> Willow Cove Redux goes from 8 execution units to 10 execution units, the ideal SMT scaling would be about 25% more throughput for a given clock frequency excluding major scheduler improvements beyond increasing the number of execution ports it can manage. Of course, there will be some collateral SMT gains from reduced execution unit contention between threads, deeper re-order queue and other improvements throughout the pipeline improving execution unit packing a handful of points beyond Skylake. A ~30% SMT throughput per core gain should be within the realm of plausible for some workloads.
First, the 25% IPC improvement depends on having the right mix of instructions to keep all of those units occupied.

Second, we don't know if the improvements you cite, beyond the increase in execution units, are needed simply to keep up with the wider architecture, or actually go further to yield additional benefits.

Finally, when people are talking about the per-clock efficiency of an architecture, it's assumed not to apply only to one cherry-picked workload. It needs to be the average IPC improvement, across an entire benchmark suite.

It should suffice to say that I remain skeptical. >= 30% IPC improvements aren't something we've seen from Intel since Core 2. I don't expect they'd suddenly be coming on fast and furious, this late in the game.
 

InvalidError

Titan
Moderator
It should suffice to say that I remain skeptical. >= 30% IPC improvements aren't something we've seen from Intel since Core 2. I don't expect they'd suddenly be coming on fast and furious, this late in the game.
Willow Cove is two generations and 5-6 years newer than Skylake and Ice Lake has already proven a ~18% IPC gain over Skylake, so we're looking at ~20% over Skylake as a starting point.

Another way to look at it is that had Intel kept up with its ~5% annual architecture improvement rate, today's CPUs would have ~25% higher IPC than Skylake. Nothing groundbreaking, merely catching up with where it should have been had its historic slow pace simply kept chugging along.
 

bit_user

Splendid
Ambassador
Ice Lake has already proven a ~18% IPC gain over Skylake,
Was that measured by running at the same clocks, or by normalizing. Because, I'm sure you'll know that performance doesn't scale linearly with frequency. What I'm saying is that by clocking Rocket Lake at 5 GHz, we might find a few % of those IPC gains from Ice Lake seem to melt away.

Also, there's the whole issue of back-porting it to 14 nm. That should tax the IPC gains, as well.
 

spongiemaster

Commendable
Dec 12, 2019
780
351
1,260
0
Reactions: bit_user

ASK THE COMMUNITY

TRENDING THREADS