10 ms·

Userbenchmark is useless, it can barely compare between Intel microarchs. Best it can do is """help""" compare two processors from the same family. The last bi

Userbenchmark is useless, it can barely compare between Intel microarchs. Best it can do is """help""" compare two processors from the same family. The last big microarch jump for intel was Skylake (6th gen) and then 7, 8, 9 and some of 10th gen (but not all, but they needed "10th gen" for marketing reasons) are just incremental improvemnts over it. So Skylake was about a 15-20% increase in performance (real) but took a few shortcuts (Meltdown and Spectre). From then on, you had a 5-10% performance improvement per generation (let's round it at 7%). That's not bad per se, but it was done out of comfort. The "cursed" 10nm node which never came is also part of the reason why Intel stalled, they insisted on "tick-tock" and were unwilling (or unable?) to deliver a new microarch over 14nm+++, the super refined node (and the extra refinement allowed them to pull the clock higher on each new gen, now alowing those sweet close-to and with this new weird turbo [that seems more marketing than substance] surpass the 5Ghz barrier). And that's how you come to Intel's situation on 2020: no 10nm node, no real new microarch in the last 5 years. Edit: As for why we never saw something like this... I think we never had a situation before where one of the competitors had such an advantage (the microarchs before Zen on AMDs part were bad, and that hit them hard for a long time) that they could pull this duckish moves for so long. For a long time, Intel chose not to improve their offering, because they were so far ahead of competition and they could keep milking the Whatever-Lake microarch family. They'd blame on the difficulty of extracting more IPC, Moores Law coming to an end (or not coming to an end, depending on what marketing was trying to sell). In the end it came to AMD+TSMC offering superb processors for Intel to slash the prices in half, and then still being crushed performance wise, in desktop, server and now mobile too. And all Intel has to show is AVX512, which after having it gatekeeped for years in the server market, they've released on consumer processors - but hey, nobody's using it because there was no support!

0x64 6y ago
> And that's how you come to Intel's situation on 2020: no 10nm node, no real new microarch in the last 5 years. Chips fabbed on 10 nm have been shipping in volume since last summer. The Sunny Cove CPU cores found in Ice Lake chips, fabbed on the 10+ node, average roughly around 16 % higher performance vs. Skylake, clock-for-clock, before accounting for improvements in memory bandwidth. Tiger Lake (fabbed on 10++) will be shipping this year, alongside with Ice Lake Xeons.
- dr_zoidberg 6y ago
  That 10nm node (10nm+ as they call it) is not as mature as 14nm+++ and they can't reach the same clock speeds (or yields) and that's why they have to resort to comparing clock-for-clock against their previous generations. Extract from Anandtech[0] comparing Comet Lake vs Ice Lake: Some of these tests rely heavily on turbo, such as the PCMark tests, and so the Comet Lake i7-10710U can hit 4.7 GHz on the latest variant of Skylake, while the Ice Lake i7-1065G7, despite its higher IPC difference, can only do 3.9 GHz. This means in a lot of bursty workloads (which a lot of business workloads are), the Comet Lake wins and we see that play out. Again, that means that some chips aren't produced in 10nm, because if they were, they'd end up being straight up slower, despite having an 18% IPC improvement. As I said before: I really hope Intel sorts this mess up and comes along with something that can compete against AMDs offering. Otherwise we might see a turn of the tables, with AMD eventually stalling progress, just because they aren't pushed hard enough by the competition. [0] https://www.anandtech.com/show/15385/intels-confusing-messaging-is-comet-lake-better-than-ice-lake https://www.anandtech.com/show/15385/intels-confusing-messag...
  - caycep 6y ago
    Are they just going to skip to 7nm at this point?
    - dr_zoidberg 6y ago
      Intel's 7nm should be comparable to TSMC's 5nm (actually Inte'ls original 10nm, the one that failed spectacularly, would've been better than TSMC's 7nm and 7nmP in terms of transistor density see [0]). And it was started a long ago (like new nodes usually are) so it should be coming along somewhere around the same time maybe? Intel hasn't said much about its future nodes, but there's some information and a few marketing slide around[1]. They seem to have learned from their 10nm mistake, so they're apparently planning to have + and ++ versions of the following nodes, that could handle production of a new microarch if the next node isn't ready, but only time will tell how that goes along. [0] https://en.wikipedia.org/wiki/7_nm_process#7_nm_process_nodes_and_process_offerings https://en.wikipedia.org/wiki/7_nm_process#7_nm_process_node... [1] https://www.anandtech.com/show/15217/intels-manufacturing-roadmap-from-2019-to-2029 https://www.anandtech.com/show/15217/intels-manufacturing-ro...
fomine3 6y ago
Userbenchmark is totally useless because too many users run benchmark in very varying environments. Better to just watch proper review sites like AnandTech, GamersNexus, Phoronix.

0x64 6y ago

> And that's how you come to Intel's situation on 2020: no 10nm node, no real new microarch in the last 5 years. Chips fabbed on 10 nm have been shipping in volume since last summer. The Sunny Cove CPU cores found in Ice Lake chips, fabbed on the 10+ node, average roughly around 16 % higher performance vs. Skylake, clock-for-clock, before accounting for improvements in memory bandwidth. Tiger Lake (fabbed on 10++) will be shipping this year, alongside with Ice Lake Xeons.

dr_zoidberg 6y ago
That 10nm node (10nm+ as they call it) is not as mature as 14nm+++ and they can't reach the same clock speeds (or yields) and that's why they have to resort to comparing clock-for-clock against their previous generations. Extract from Anandtech[0] comparing Comet Lake vs Ice Lake: Some of these tests rely heavily on turbo, such as the PCMark tests, and so the Comet Lake i7-10710U can hit 4.7 GHz on the latest variant of Skylake, while the Ice Lake i7-1065G7, despite its higher IPC difference, can only do 3.9 GHz. This means in a lot of bursty workloads (which a lot of business workloads are), the Comet Lake wins and we see that play out. Again, that means that some chips aren't produced in 10nm, because if they were, they'd end up being straight up slower, despite having an 18% IPC improvement. As I said before: I really hope Intel sorts this mess up and comes along with something that can compete against AMDs offering. Otherwise we might see a turn of the tables, with AMD eventually stalling progress, just because they aren't pushed hard enough by the competition. [0] https://www.anandtech.com/show/15385/intels-confusing-messaging-is-comet-lake-better-than-ice-lake https://www.anandtech.com/show/15385/intels-confusing-messag...
- caycep 6y ago
  Are they just going to skip to 7nm at this point?
  - dr_zoidberg 6y ago
    Intel's 7nm should be comparable to TSMC's 5nm (actually Inte'ls original 10nm, the one that failed spectacularly, would've been better than TSMC's 7nm and 7nmP in terms of transistor density see [0]). And it was started a long ago (like new nodes usually are) so it should be coming along somewhere around the same time maybe? Intel hasn't said much about its future nodes, but there's some information and a few marketing slide around[1]. They seem to have learned from their 10nm mistake, so they're apparently planning to have + and ++ versions of the following nodes, that could handle production of a new microarch if the next node isn't ready, but only time will tell how that goes along. [0] https://en.wikipedia.org/wiki/7_nm_process#7_nm_process_nodes_and_process_offerings https://en.wikipedia.org/wiki/7_nm_process#7_nm_process_node... [1] https://www.anandtech.com/show/15217/intels-manufacturing-roadmap-from-2019-to-2029 https://www.anandtech.com/show/15217/intels-manufacturing-ro...

fomine3 6y ago

Userbenchmark is totally useless because too many users run benchmark in very varying environments. Better to just watch proper review sites like AnandTech, GamersNexus, Phoronix.