Intel Demonstrates 65W Broadwell-K Socketed Processors at GDC 2015

[mikk]

The article clearly state that it doesnt throttle during benches, only when Prime 95 + Furmark are running both.

This sounds a bit different. GPU clock varies heavily depending on the CPU load.

Der Maximaltakt beträgt bei unserem Testgerät laut HWinfo64 950 MHz, variiert aber je nach Prozessorlast mitunter stark.

 

[Abwx]

This sounds a bit different. GPU clock varies heavily depending on the CPU load.

To be precise the chip throttle in the CPU + GPU stress test but not when only Prime 95 is used.

A genuine 15W CPU can drain 25W during peaks on the order of a millisecond, thoses 25W peaks wont appear at the main level even with good measurement tools as the excess energy over time is low enough that the peak is litteraly leveled by the power supply time constants.

What Intel did was to extend this notion of instantaneous power comsumption to periods as long as 30s or 1mn, to summarize they re assuming, and that s right mathematicaly speaking, that 25W during 30s followed by 0W during the same period amount to 12.5W on average.

I find this kind of logic somewhat dubbious even if founded technicaly speaking, at least it s not a pure scam like the race to idling myth.

In this latter case, and excluding trivial and almost irrelevant solutions, the math says that optimal perf/Watt is when the CPU comsumption is equal to the rest of the system comsumption, if the CPU power drain exceed the rest of system power drain then efficency decrease because starting from this point the CPU performance delta in %age will be less than the total system power delta in %age, hence battery life will be less at equal work.

This is due to the CPU power increasing as a square law in respect of performance, while the race to idling myth is implying that it s a linear law....

 

[Dufus]

This should allow to extrapolate the caracteristics of a 4C/8T variant :

http://www.notebookcheck.com/Test-Intel-NUC-5i5RYH-Mini-PC.138750.0.html

Possibly you might be able to get a ballpark figure but first there's a few spanners in the works that need attention.

• Probably foremost binning is going to make a difference.
• Power limiting should be avoided.
• My own testing of BDW i7-5500U using Cinebench 11.5 ran at 2.9GHz on 2C/4T with a power draw hovering around 15W, just on the edge of power limiting and scored 3.3, posted in crashtech's thread. It did have a -50mV offset for that. Fairly consistent with the score you linked at 2.6GHz.
• HSW mobile quads nominal TDP is 47W. At default bins / voltages and without power limiting and using Intels inbuilt RAPL estimations my own i7-4700MQ will pull over 47W with Cinebench and 80W with Linpack.

I do not have a dual core HSW but can disable 2 cores in the BIOS for my quad core i7-4700MQ. Cinebench at the same clock of 2.9GHz using 2C4T scores 3.17 showing BDW with an IPC of around 5%. Where BDW shines is in power use, the BDW using IA core power of 13.7W while package power is at 15W. Haswell used 19.8W core and a whopping 28.1W of package power, nearly twice that of BDW. And this is why BDW performs better in power limited scenarios, not only more efficient but theres more power for the cores to use.

Using your "extrapolation" on the BDW i7-5500U using my own calculations would put a similar quad core variant of 65W at a base of 3.5GHz and 4 core turbo at 3.9GHz with lower core usage at probably 4.0GHz and 4.1GHz. Perhaps Intel might be able to squeeze a couple of more bins on that.

 

[Dave2150]

Possibly you might be able to get a ballpark figure but first there's a few spanners in the works that need attention.

• Probably foremost binning is going to make a difference.
• Power limiting should be avoided.
• My own testing of BDW i7-5500U using Cinebench 11.5 ran at 2.9GHz on 2C/4T with a power draw hovering around 15W, just on the edge of power limiting and scored 3.3, posted in crashtech's thread. It did have a -50mV offset for that. Fairly consistent with the score you linked at 2.6GHz.
• HSW mobile quads nominal TDP is 47W. At default bins / voltages and without power limiting and using Intels inbuilt RAPL estimations it will pull over 47W with Cinebench and 80W with Linpack.

I do not have a dual core HSW but can disable 2 cores in the BIOS for my quad core i7-4700MQ. Cinebench at the same clock of 2.9GHz using 2C4T scores 3.17 showing BDW with an IPC of around 5%. Where BDW shines is in power use, the BDW using IA core power of 13.7W while package power is at 15W. Haswell used 19.8W core and a whopping 28.1W of package power, nearly twice that of BDW. And this is why BDW performs better in power limited scenarios, not only more efficient but theres more power for the cores to use.

Using your "extrapolation" on the BDW i7-5500U using my own calculations would put a similar quad core variant of 65W at a base of 3.5GHz and 4 core turbo at 3.9GHz with lower core usage at probably 4.0GHz and 4.1GHz. Perhaps Intel might be able to squeeze a couple of more bins on that.

Looking unlikely that Intel can surpass the 4790k clocks of 4Ghz/4.4Ghz turbo then.

Would have been nice to see Broadwell-k being 4.2Ghz stock, 4.6Ghz turbo, coupled with the 5% IPC increase, would have made a adequate upgrade from the 4790k.

Maybe with the clock speeds you suggested, plus the 128MB L4 cache for CPU use (when coupled with a dedicated GPU) would let it completely outperform the 4790k regardless?

Hoping for some news soon!

 

[Dufus]

Will only no for sure when they come out.

 

[StrangerGuy]

My conjecture is that the heavy emphasis on Iris Pro, TDP and SFFs, but the lack of any actual frequency suggest that Intel is trying to hide that absolute CPU performance has declined relative to Haswell due to significantly lower clocking headroom, for whatever reason.

 

[crashtech]

Headroom does seem to be decreasing slightly with each iteration ever since Sandy Bridge.

 

[mikk]

To be precise the chip throttle in the CPU + GPU stress test but not when only Prime 95 is used.

Once again, you don't have frequency logs from all these games. And this is where both CPU and GPU are under load. How can you say this chip is running full speed all the time? You simply can't.

 

[jhu]

Back in my day, Pentiums were running at 5V and 16W. And that was considered too hot!

 

[Arachnotronic]

My conjecture is that the heavy emphasis on Iris Pro, TDP and SFFs, but the lack of any actual frequency suggest that Intel is trying to hide that absolute CPU performance has declined relative to Haswell due to significantly lower clocking headroom, for whatever reason.

Yup. This looks like a successor to the 4770R in LGA form, not to the 4790K.

 

[coercitiv]

My own testing of BDW i7-5500U using Cinebench 11.5 ran at 2.9GHz on 2C/4T with a power draw hovering around 15W, just on the edge of power limiting and scored 3.3, posted in crashtech's thread. It did have a -50mV offset for that. Fairly consistent with the score you linked at 2.6GHz.

I do not have a dual core HSW but can disable 2 cores in the BIOS for my quad core i7-4700MQ. Cinebench at the same clock of 2.9GHz using 2C4T scores 3.17 showing BDW with an IPC of around 5%. Where BDW shines is in power use, the BDW using IA core power of 13.7W while package power is at 15W.

My own testing of i7-4510 using same Cinebench 11.5 ran at 2.8Ghz on 2C/4T with a max power draw of 16.7W package power. (same undervolt as you, -50mV)

My 4700HQ, limited to 2.9Ghz and running 2C/4T Cinebench 11.5 scores 3.18 while using no more than 18W package power, with the average bellow 17W. (standard voltage)

Haswell used 19.8W core and a whopping 28.1W of package power, nearly twice that of BDW. And this is why BDW performs better in power limited scenarios, not only more efficient but theres more power for the cores to use.

I don't know how you got that HW power usage data, but something went definitely wrong. I can only get ~28W package power if I run 4C/8T undervolted by -50mV.

By corroborating your data and mine, in Cinebench BDW U uses ~10% less power while running 0.1Ghz closer to 3Ghz than HW U. Maybe this can help when extrapolating towards desktop use.

 

[Dufus]

Thanks for the feedback coercitiv, nice to see some other figures.

My 4700HQ, limited to 2.9Ghz and running 2C/4T Cinebench 11.5 scores 3.18 while using no more than 18W package power, with the average bellow 17W. (standard voltage)

I don't know how you got that HW power usage data, but something went definitely wrong. I can only get ~28W package power if I run 4C/8T undervolted by -50mV.

That's interesting. What core power did you get? My package power when on load always seems to sit from about 7W to 9W above core power.

I have checked with battery power by using the EC battery current and voltage values to show power drawn while operating on battery. I would hope those values are at least reasonable in the interests of the battery. It appears to tie in quite well with CPU package power delta.

I don't know about the HQ but my MQ does seem to tie in with with other MQ data I have seen on the net. Of course some small variance is expected from chip to chip.

How does your HQ fair power wise with the below results using one of the higher multi's?
http://forum.techinferno.com/thrott...-bridge-i-have-some-shocking-tdp-results.html

 

[mikk]

http://chinese.vr-zone.com/146637/i...top-only-have-i7-5775c-and-i5-5675c-03232015/


Base clock slighty higher and turbo slightly lower than HSW GT3e by the looks of it. GPU clock still unknown.

 

[coercitiv]

That's interesting. What core power did you get? My package power when on load always seems to sit from about 7W to 9W above core power.

Here's a HWInfo log file for a 2c4t run @ 2.9Ghz (no undervolt). You have all the data there, and I ran the test on battery so you also have battery drain info to compare.

The log starts and ends with ~20 seconds of low/idle power, score in Cinebench was 3.18 pts

 

[Enigmoid]

http://chinese.vr-zone.com/146637/i...top-only-have-i7-5775c-and-i5-5675c-03232015/


Base clock slighty higher and turbo slightly lower than HSW GT3e by the looks of it. GPU clock still unknown.

4 MB cache?

Weird. With an inclusive cache system that's really only 3 MB unique cache.

 

[mikk]

4 MB cache?

Weird. With an inclusive cache system that's really only 3 MB unique cache.

4MB and 6MB. Cache size doesn't differ to HSW GT3e, not sure why it is weird.

 

[Dave2150]

http://chinese.vr-zone.com/146637/i...top-only-have-i7-5775c-and-i5-5675c-03232015/


Base clock slighty higher and turbo slightly lower than HSW GT3e by the looks of it. GPU clock still unknown.

Disappointing. Intel's 14nm process seems many steps backwards for the high performance mainstream market.

I wonder how shitty Broadwell-E will be, when it launches next year. Wouldn't surprise me if they skip it, since I doubt it's going to match/exceed Haswell-E's clock speeds.

 

[crashtech]

Here's a HWInfo log file for a 2c4t run @ 2.9Ghz (no undervolt). You have all the data there, and I ran the test on battery so you also have battery drain info to compare.

The log starts and ends with ~20 seconds of low/idle power, score in Cinebench was 3.18 pts

Huh, I tested an i5-4300M at 2.9GHz that scored 3.47 in Cinebench 11.5.

http://forums.anandtech.com/showpost.php?p=36940174&postcount=233

 

[DrMrLordX]

Back in my day, Pentiums were running at 5V and 16W. And that was considered too hot!

Do we need to get off your lawn now? 'course I go back to those days too but whatever.

Base clock slighty higher and turbo slightly lower than HSW GT3e by the looks of it. GPU clock still unknown.

Hmm, not bad. Looks like the speculations was true, and that we are seeing the successor to the 4770R. The turbo clock is a bit disappointing but whatever.

Disappointing. Intel's 14nm process seems many steps backwards for the high performance mainstream market.

I wonder how shitty Broadwell-E will be, when it launches next year. Wouldn't surprise me if they skip it, since I doubt it's going to match/exceed Haswell-E's clock speeds.

Hey now, they actually brought the base clock up by 100 mhz. That ain't bad. The real story will be: where's the clock ceiling? The C chips are unlocked, after all, so we will find out just how high they can go . . . something we couldn't do with the 4770R.

Given the base clock, I doubt it will reach 4.5-4.7 ghz easily.

 

[coercitiv]

Huh, I tested an i5-4300M at 2.9GHz that scored 3.47 in Cinebench 11.5.

http://forums.anandtech.com/showpost.php?p=36940174&postcount=233

Funny how Broadwell scoring 3.3pts at 2.9Ghz did not prompt any reaction from you, but a Haswell @ 2.9Ghz scoring 3.18pts is a trigger

 

[crashtech]

Funny how Broadwell scoring 3.3pts at 2.9Ghz did not prompt any reaction from you, but a Haswell @ 2.9Ghz scoring 3.18pts is a trigger

I'm probably just misinterpreting the meaning of the data.

 

[coercitiv]

I'm probably just misinterpreting the meaning of the data.

Might be that too.

9. crashtech | i3-4130T | 2.9GHz | 3.18

 

[Dufus]

Huh, I tested an i5-4300M at 2.9GHz that scored 3.47 in Cinebench 11.5.

http://forums.anandtech.com/showpost.php?p=36940174&postcount=233

While the CPU-z screen shot shows 2.9GHz was it actually running at 3.2GHz during the bench? 3.2/2.9*3.18=3.51

@coercitiv, thanks for the data. Couple of things, you ran with all cores enabled and infinity set whereas my testing was with 2 cores disabled from the BIOS. I re-ran with affinity and 4 cores and strangely the package power was a little lower. Your default VID is much lower than mine, 0.918V vs 0.983V but still does not fully count for the big difference. I tried maximizing power saving features and best I could come up with at 2.9GHz, same score and negative core and cache offset to give 0.901V was

20J package and 13J core with package and battery differential of 19J and 21W
vs yours
16.5J package, 11J core with package and battery differential of 15.5J and 19W.

That's a nice 4700HQ you have. I guess it also shows more data is required for a decent median. I have already seen my predicted calculations for the 65W BDW were wrong. Would be interesting to see some power / performance numbers for them when available. Also wonder if or how much higher vcore will be on BDW compared to HSW.

 

[AtenRa]

http://chinese.vr-zone.com/146637/i...top-only-have-i7-5775c-and-i5-5675c-03232015/


Base clock slighty higher and turbo slightly lower than HSW GT3e by the looks of it. GPU clock still unknown.

Seriously, if Intel will only release those two SKUs then nobody considering upgrading to Broadwell will choose any of those two over a Haswell Devils Canyon Core i5 and especially Core i7 4970K, not even in the same price.

edit: Even if you have a Celeron/Pentium/Core i3 Haswell those two Broadwell SKUs are not worth considering over Haswell Devil Canyon SKUs. And if you are going for a new Build, then better go for the new Skylake.

 

[Dave2150]

Seriously, if Intel will only release those two SKUs then nobody considering upgrading to Broadwell will choose any of those two over a Haswell Devils Canyon Core i5 and especially Core i7 4970K, not even in the same price.

edit: Even if you have a Celeron/Pentium/Core i3 Haswell those two Broadwell SKUs are not worth considering over Haswell Devil Canyon SKUs. And if you are going for a new Build, then better go for the new Skylake.

Unless Skylake's architecture has a huge IPC increase, it will most likely also be gimped by Intel's 14nm process that cannot reach the clockspeeds that 22nm afforded us.