Discussion i7-11700K preliminary results

[GunsMadeAmericaFree]

early benchmarks

more benchmarks

Just thought I would provide these so folks could draw their own conclusions. Evidently the processors have been getting sold for a few days now in Germany.

 

[inf64]

The IPC results are telling, intel hit the 19% only in FP heavy code. In integer it is 13% in ST and 7% in MT. Zen3 has a massive IPC advantage over this thing and clocks ~5% lower out of the box. Power numbers are insane but expected. 11900K will add maybe 4-5% to these numbers so nothing major, it's still inferior product in almost every single way (except AVX512 which is like 0.01% of all workloads).

It's DOA unless AMD is able to screw up the supply side for the next 8 months.

I hate to say it I told you so, but I told you so.

 

[majord]

Maybe people are being dumb on purpose beating the dead horse, OK it has far worse figures than expected, but saying like "104°C" and "290W!" misses this situation where it's happening:

Hint: that's a little more than your average 19% IPC.

As for gaming definitely big disappointment, the latency hit that hard, maybe some much faster memory even in 1:2 mode will manage something... but at this point it's worth it over Comet only if you get it for a decent price. Or not at all, with Alder coming out in a few months.

I agree.. citing the AVX512 power numbers without context is just hyperbole. The perf/watt is still bad without resorting to that, but unfortunately this is 2021


That said, Your'e dreaming if you think Alder lake is a few months away

 

[SAAA]

That said, Your'e dreaming if you think Alder lake is a few months away

You are right... I forgot to add "intel" months, like nm, with a Q6 retail availability.

Now I wonder what it will happen with Alder if the IPC tests do show big gains but they move to Tiger Lake cache setup. Hopefully intra-ring stays as adding mesh on top of this could kill it for good.

 

[Racan]

I agree.. citing the AVX512 power numbers without context is just hyperbole. The perf/watt is still bad without resorting to that, but unfortunately this is 2021

AVX512 Perf/watt isn't bad in that case then, but still considering that in their test setup they were using a freaking ~2kg copper aircooler and were still reaching 104 degrees C, taming the AVX512 beast will be hard.

 

[scineram]

avx 512 is just a mill weight around intel's neck.

They just need to drop that already. Those engineers likely could do amazing things with the freed up die space.

Ayye. I wish AMD just added a third fully independent AVX2 pipeline for consoomer Zen 3, instead of AVX-512. Just look at brutal M1 FPU perf.

 

[Asterox]

AVX512 Perf/watt isn't bad in that case then, but still considering that in their test setup they were using a freaking ~2kg copper aircooler and were still reaching 104 degrees C, taming the AVX512 beast will be hard.

Great, but in real life or Windows+aplications where can you use that mighty AVX512?Can we use that red ghost to do(accelerates rendering) a epic video like this?Red is just the ghost, you may even happen to see him in action on PC but only if you are very lucky.

 

[Mopetar]

Unless you've got some AVX-512 code to run there isn't too much of a point in upgrading from anything recent.

It'll sell anyway though. There just aren't enough wafers at TSMC for AMD to put a sizable dent in Intel right now.

 

[RTX2080]

Worse than expected......Assume if Intel bring 10nm SunnyCove to desktop instead of backporting it, what will it look like. After seeing TGL-H has 5.0Ghz turbo I have to doubt 10nm yield is that suffer. To me looks like Intel make a strategic mistake......
And IMO I think AVX512 is not a successiful ISA, not being widely used, draws a lot of heat and power that hard to handled by a mainstream platform, no matter how much performance it could produce, it still doesn't make sense......

 

[jpiniero]

Worse than expected......Assume if Intel bring 10nm SunnyCove to desktop instead of backporting it, what will it look like. After seeing TGL-H has 5.0Ghz turbo I have to doubt 10nm yield is that suffer. To me looks like Intel make a strategic mistake......

You still have to deal with the defect yield. Tiger Lake is clearly a lot better than Ice Lake in that respect but you can't say at this point how many wafers Intel will actually devote to H45. It's a lot easier to hide actual supply behind BGA laptops where they sell similar models with processors from different families including AMD.

 

[Hulk]

Worse than expected......Assume if Intel bring 10nm SunnyCove to desktop instead of backporting it, what will it look like. After seeing TGL-H has 5.0Ghz turbo I have to doubt 10nm yield is that suffer. To me looks like Intel make a strategic mistake......
And IMO I think AVX512 is not a successiful ISA, not being widely used, draws a lot of heat and power that hard to handled by a mainstream platform, no matter how much performance it could produce, it still doesn't make sense......

It does seem like the *simple* solution for Intel would have simply to bring Tiger Lake 8 core to the desktop since they R&D and production is already there. But there could be as yet unresolved technical challenges with operating 10SF at or near 5GHz all core for 8 cores, namely thermal dissipation challenges. Evidence for this would be the fact that Zen 3 while quite efficient exhibits high temps most likely due to high transistor density creating hotspots. Intel needs max clocks and production capability now more than efficiency. Hence Rocket Lake 14++++++++++++++ (it's fun to hold down the "+" key).

But more importantly Intel has a very good idea of how many orders it needs to fulfill over the next 6 to 12 months and I don't think they could make those shipments without 14++++++++++++++ taking care of much of it. I don't consider Rocket Lake a fail or a success, it's more of a stopgap stutter step. It will keep production moving along until they fully transition the fabs from 14+++++++++++ to enhanced 10SF.

All that being said (written), if there are further delays with 10nm OR Alder Lake isn't able to decisively beat Zen 3 core for big core then things are going to get a little more serious. Intel has a big "to do" list:

Intel "To-Do" List
1. Transition fabs from 14+++++++++++ to 10SF+
2. Ensure Golden Cove decisively beats Zen 3 and can compete with Zen 4.
3. Ensure Alder Lake 8+8 can beat 5900X and compete with 5950X.
4. Get Alder Lake as described above out the door in quantity in 2021.

 

[jpiniero]

https://twitter.com/x/status/1368238602279399425

Ian thinks the 8C Rocket Lake die is 270 mm2. Talk about yuge.

 

[fleshconsumed]

I'm enjoying all the rubbernecking around the dumpster fire that is Intel right now lol...

 

[dmens]

TSMC 7nm is going to be at least 50%+ more efficient than Intel 14nm by default at peak performance.

This sentence is entirely nonsensical.

I don't know why people act so surprised about this. What's interesting is that Intel could push these chips way above their efficiency zones and still remain in the ballpark. Maybe it says something about how much AMD is also pushing TSMC 7nm.

Saying that shoving amps into a part is interesting is like saying watching paint dry is interesting. From an engineering point of view, it is a total abandonment of engineering trade-off consideration and diligence.

 

[Hail The Brain Slug]

Are we taking bets on how Intels going to market this thing?

My top guesses are taglines like
"the fastest 8 cores you've ever seen" or "the fastest 8 cores we've ever made"
"No one needs more than 8 cores for gaming",
Maybe using timespy extreme CPU score (AVX512) as if it were indicative of game performance

 

[JoeRambo]

After all it could be BIOS / uCode problem. On OC.net people are getting ~proper latencies with fixed core clock / uncore speeds. For example from:

Overclocking 11700k/11900k results, bins and discussion

www.overclock.net

Everything looks about right for clocks involved.

I think what is happening with Anandtech sample + mobo is some seriuos trouble with uncore clocks.

https://images.anandtech.com/doci/16535/Bounce-11700K_575px.png

Probably uncore is not ramping up clocks properly, cause core 2,3 intercore latencies look ~ right for 4Ghz uncore these things are supposed to run by default. Yet somehow it is broken for others. Anandtech could shed some light into situation by simply checking core/uncore clocks during these tests.

 

[AnandThenMan]

Intel Bulldozer'd themselves. I know all the signs were pointing in that direction but I refused to believe it.

 

[tamz_msc]

After all it could be BIOS / uCode problem. On OC.net people are getting ~proper latencies with fixed core clock / uncore speeds. For example from:

Overclocking 11700k/11900k results, bins and discussion

www.overclock.net

View attachment 40712


Everything looks about right for clocks involved.

I think what is happening with Anandtech sample + mobo is some seriuos trouble with uncore clocks.

https://images.anandtech.com/doci/16535/Bounce-11700K_575px.png

Probably uncore is not ramping up clocks properly, cause core 2,3 intercore latencies look ~ right for 4Ghz uncore these things are supposed to run by default. Yet somehow it is broken for others. Anandtech could shed some light into situation by simply checking core/uncore clocks during these tests.

Why would uncore clocking lower affect all but two cores in terms of inter-core latency? Should it not result in higher latency across all cores?

 

[NTMBK]

Why would uncore clocking lower affect all but two cores in terms of inter-core latency? Should it not result in higher latency across all cores?

Are the two cores unaffected not hyperthreads on the same core? Hence why there are 16 "cores".

 

[Mopetar]

Are we taking bets on how Intels going to market this thing?

My top guesses are taglines like
"the fastest 8 cores you've ever seen" or "the fastest 8 cores we've ever made"
"No one needs more than 8 cores for gaming",
Maybe using timespy extreme CPU score (AVX512) as if it were indicative of game performance

"Our hottest chips yet!"
"The hottest new chip on the market!"

 

[tamz_msc]

Are the two cores unaffected not hyperthreads on the same core? Hence why there are 16 "cores".

I think they fix core 0 and load up threads one-by-one all the way till thread 15. That's why you have it increase from 29 to 31 ns as you go "outward". The discrepancy is NOT due to the hyperthreads on one core.

 

[JoeRambo]

Why would uncore clocking lower affect all but two cores in terms of inter-core latency? Should it not result in higher latency across all cores?

No idea, but uncore also follows turbo algorithms and maybe they are broken.

Aida L3/memory latency does not show any significant regression versus Skylake, but AIda L3 latency simply loads a single core 100%. Still Anandtech could investigate it more, by looking into uncore clocks.

 

[beginner99]

AMD has shown it is extremely supply constrained,

Same issue they had in the Athlon days. Way better uArch but not nearly enough supply to make any meaningful dent into intel.

 

[inf64]

I did some quick math regarding the spec numbers from Rocket Lake preview on AT and the performance per clock almost perfectly matches the result AT got when they reviewed Zen3 . Zen3 has around ~15% higher ST integer IPC than Cypress Cove and around ~8% higher ST fp IPC in SPEC2017 rate-1 benchmark suite.

This is almost a perfect match to the Zen3 deep dive results, the only difference is that Cypress Cove has a bit lower performance than WC and is in the middle, basically (between IL and WC).




Below is the chart from Rocket Lake preview, if you adjust Zen3 scores for 4% deficit in clock speed you get the numbers I posted above, almost a perfect match:



So to sum it up, inferior IPC(and not a by a small amount!), worse latency due to (most likely) physical implementation compromises in the cache subsystem, somewhat higher clock speeds but nothing to write home about and extremely bad power draw. Gaming performance suffers due to worse latency than intel 10th gen but lets see the final firmware results. I doubt there will be much of a change, probably +-3% at best. They seem SOL in every workload except AVX512 and pricing is really out of touch with reality (even if you count in the scalper tax right now).

I do get the urge to launch SOMETHING new, but it kinda loses the point if it's so inferior. Looking back at their slides, if all they could show for gaming was ~4% higher numbers in their pre-selected few titles under murky test conditions, that was a tell tale sign of desperation .

 

[scineram]

This is what microcode usually does. Not make things faster.

 

[Dave2150]

I agree.. citing the AVX512 power numbers without context is just hyperbole. The perf/watt is still bad without resorting to that, but unfortunately this is 2021


That said, Your'e dreaming if you think Alder lake is a few months away

There's an army of die hard AMD fanatics who'll continue to reveal their intellectual maturity by posting the AVX-512 power draw without context. Typically, this crowd has little else to do than haunt tech forums/news sites comments section for hours per day, to defend their favourite company.

Have to remember they're only doing this as they feel threatened in some way, I'm not sure why...

The i9 with release Microcode/UEFI should take the gaming crown back, something I never thought possible while still on 14nm.