Since Arrow Lake is hopefully months away I think it deserves it's own thread. Especially now that Meteor Lake is available and we have some idea of how Intel 4 is performing.
To get this going I will submit my prediction for Arrow Lake.
I am thinking Arrow Lake will be a 8+24 part with max clock for the P's to be about 5.5GHz and about 4.3GHz for the E's.
Here's my reasoning.
Intel has painted themselves into the same corner they did with Haswell/Devil's Canyon when moving to Broadwell. We have discussed this. Devil's Canyon was on a very mature 22nm process while Broadwell was on a brand new 14nm process. Broadwell didn't have enough frequency and IPC to match much less beat Devil's Canyon so it was in reality a mobile only part. This is much like the situation with Raptor Lake and Meteor Lake today. Except this time it will be even harder for Intel to beat Raptor Cove because Raptor is monolithic vs tiles ARL and Raptor is currently reaching insane frequencies albeit at the expense of insane power.
What is Intel to do?
Based on current Meteor mobile frequency I think 5.5GHz for desktop is realistic for the 2nd go at Intel 4. I am also going to go out on a limb and predict Lion Cove will show about 15% better throughput (IPC) than Raptor Cove. But, this will be offset by the clock speed disadvantage and latency penalty of Arrow Lake. I will estimate this penalty to be 10%. About 6 or 7% do to clock speed reduction and the rest due to the disaggregated architecture.
This means that Arrow Lake would have a 5% performance improvement over Raptor Lake in ST. Many people will complain like mad about this not realize that Intel had to overcome a large clock deficit and tile latency to achieve it. Both the people who complain about the relative lack of ST performance increase and those who claim Intel did well for moving to a new node and tiled architecture will be correct in their appraisal
Okay so they manage to increase ST performance by removing HT structures from Lion Cove, freeing up area for the core to be wider, additional OoO logic, etc... They also offset the increased thermal density in moving to Intel 4 by removing HT from the big cores.
But how do they make up the MT performance? I think they replace the 8 logical P cores + 16 Gracemont E cores (24 total threads) with 24 Crestmont E cores. The E core clusters are quite area efficient and there should be room for them with the move to Intel 4. This not only improves on MT performance because E's are more performant than HT logical cores, but it also makes "life easier" for the Thread Director since now it only has to schedule P's and E's, no more logical threads in the mix. Also the penalty for putting a thread on a logical core when it should be on a P physical core is somewhat mitigated as that mistake would be now placing the thread on what is at least a physical E core. But I have not real evidence for this, it just seems reasonable. If Crestmont has better IPC than Gracemont then ARL will do even better in MT.
Finally ARL should have an efficiency advantage over Raptor Lake. Not only in terms of power but in terms of the very real problem that you can't really extract full performance from Raptor without a custom loop and/or delicate underclocking. If my hypothetical 8+24 5.5/4.3 ARL can run those frequencies on air or perhaps a 240 AIO then that is better than Raptor performance that can be attained by mere mortals without custom loops and God-like knowledge of undervolting and/or winning the silicon lottery and getting a highly binned part.
So in conclusion I see the 15900K being 8+24, no HT, running 5.5/4.3 with 5% better ST performance and 10% better MT performance, but this performance is actually attainable and won't require 300+ watts and commensurate cooling solutions.
Okay, have at me!
To get this going I will submit my prediction for Arrow Lake.
I am thinking Arrow Lake will be a 8+24 part with max clock for the P's to be about 5.5GHz and about 4.3GHz for the E's.
Here's my reasoning.
Intel has painted themselves into the same corner they did with Haswell/Devil's Canyon when moving to Broadwell. We have discussed this. Devil's Canyon was on a very mature 22nm process while Broadwell was on a brand new 14nm process. Broadwell didn't have enough frequency and IPC to match much less beat Devil's Canyon so it was in reality a mobile only part. This is much like the situation with Raptor Lake and Meteor Lake today. Except this time it will be even harder for Intel to beat Raptor Cove because Raptor is monolithic vs tiles ARL and Raptor is currently reaching insane frequencies albeit at the expense of insane power.
What is Intel to do?
Based on current Meteor mobile frequency I think 5.5GHz for desktop is realistic for the 2nd go at Intel 4. I am also going to go out on a limb and predict Lion Cove will show about 15% better throughput (IPC) than Raptor Cove. But, this will be offset by the clock speed disadvantage and latency penalty of Arrow Lake. I will estimate this penalty to be 10%. About 6 or 7% do to clock speed reduction and the rest due to the disaggregated architecture.
This means that Arrow Lake would have a 5% performance improvement over Raptor Lake in ST. Many people will complain like mad about this not realize that Intel had to overcome a large clock deficit and tile latency to achieve it. Both the people who complain about the relative lack of ST performance increase and those who claim Intel did well for moving to a new node and tiled architecture will be correct in their appraisal
Okay so they manage to increase ST performance by removing HT structures from Lion Cove, freeing up area for the core to be wider, additional OoO logic, etc... They also offset the increased thermal density in moving to Intel 4 by removing HT from the big cores.
But how do they make up the MT performance? I think they replace the 8 logical P cores + 16 Gracemont E cores (24 total threads) with 24 Crestmont E cores. The E core clusters are quite area efficient and there should be room for them with the move to Intel 4. This not only improves on MT performance because E's are more performant than HT logical cores, but it also makes "life easier" for the Thread Director since now it only has to schedule P's and E's, no more logical threads in the mix. Also the penalty for putting a thread on a logical core when it should be on a P physical core is somewhat mitigated as that mistake would be now placing the thread on what is at least a physical E core. But I have not real evidence for this, it just seems reasonable. If Crestmont has better IPC than Gracemont then ARL will do even better in MT.
Finally ARL should have an efficiency advantage over Raptor Lake. Not only in terms of power but in terms of the very real problem that you can't really extract full performance from Raptor without a custom loop and/or delicate underclocking. If my hypothetical 8+24 5.5/4.3 ARL can run those frequencies on air or perhaps a 240 AIO then that is better than Raptor performance that can be attained by mere mortals without custom loops and God-like knowledge of undervolting and/or winning the silicon lottery and getting a highly binned part.
So in conclusion I see the 15900K being 8+24, no HT, running 5.5/4.3 with 5% better ST performance and 10% better MT performance, but this performance is actually attainable and won't require 300+ watts and commensurate cooling solutions.
Okay, have at me!