Unlocked AM1 APUs... why not?

[VirtualLarry]

Just curious if anyone had any thoughts or suppositions. It's a SoC, and BCLK OC is risky due to PCI-E and SATA crapping out, but what if they shipped them with an unlocked multiplier?

I'm honestly a bit curious how high the Kabini silicon can clock.

Not to mention, they would make nice little "speed racers", and, at the same price point as Intel big-core Celeron CPUs, if AM1 were unlocked, they might offer better value at the low end, and might win sales from Intel.

 

[coercitiv]

These chips are not built for speed. By the time you reach 3Ghz they will likely end up using more juice than Steamroller at higher frequencies. The system power consumption delta between Kabini @ 1.6Ghz and Kabini @ 2.7Ghz is almost 30W. The stock cooling solution was already having difficulties in keeping temps in check with the increased power usage.

Also, if you really wanted to find out how high it can clock, you could have done that by now. People overclocked the heck out of it.

 

[VirtualLarry]

These chips are not built for speed. By the time you reach 3Ghz they will likely end up using more juice than Steamroller at higher frequencies. The system power consumption delta between Kabini @ 1.6Ghz and Kabini @ 2.7Ghz is almost 30W. The stock cooling solution was already having difficulties in keeping temps in check with the increased power usage.

Also, if you really wanted to find out how high it can clock, you could have done that by now. People overclocked the heck out of it.

I got my 3850 Semprons up to 1.65Ghz, which made them almost bearable. But then I discovered that the DVD drive wouldn't work at those higher BCLK speeds.

 

[NTMBK]

Same reason Intel doesn't ship Pentiums, Celerons and i3s with unlocked multipliers (apart from a single one-off exception, and Anniversary Edition). They would cannibalise their more valuable products.

 

[Magic Carpet]

1.5 @ 15W
4.0 @ 95W ?

Kabini even @ 4.0 Ghz wouldn't suck so bad, I would imagine. Provided it would scale properly and the power consumption doesn't go thru the roof. Should be 3 times faster than my A4-5000 per core. Not bad at all.

The people who designed this SoC, should know everything about these things but they won't share this info with us?

 

[Abwx]

dp

 

[Abwx]

The system power consumption delta between Kabini @ 1.6Ghz and Kabini @ 2.7Ghz is almost 30W.

Since from 1.6 to 2.7 TDP increase by 2.85x you are saying that at 1.6 Kabini s TDP is 16W, wich is completely wrong.

Measurements i made are barely 15W with Prime 95 and 12.5W with Cinebench 11.5, at 2.05GHz since it s an Athlon 5350.

So at 2.7 it would be 26W and 22W respectively, that is 11W and 9.5W deltas, we are really very far from your 30W, even if we would take the measurement at the main and starting from 1.6GHz since your delta is more than the APU TDP at 2.7....

 

[superstition]

The motherboards are very unlikely to be built to handle the power draw.

 

[Abwx]

AM1 is designed for 25W continuous power, so that shouldnt be a problem if little coolers are sticked on the VRMs, so far there has been higher frequencies than 2.7 and the power deliveries were enough.

 

[superstition]

if little coolers are sticked on the VRMs

That's a problem, though. AMD is unlikely to sell unlocked chips and tell buyers to purchase VRM sinks. It also makes the value proposition worse. The popular copper ones are $16 on Newegg or something.

 

[myocardia]

There is a reason that only the most expensive CPUs have unlocked multipliers. It's called profit.

 

[Abwx]

That's a problem, though. AMD is unlikely to sell unlocked chips and tell buyers to purchase VRM sinks. It also makes the value proposition worse. The popular copper ones are $16 on Newegg or something.

The Asus, wich are ocking dedicated, have said coolers..

That said there s not much use to overclock, at 2GHz the chip is already quite enough, its main advantage is a very good SATA interface, this more than a few hundreds Mhz is much more important for the actual user experience.

 

[coercitiv]

So at 2.7 it would be 26W and 22W respectively, that is 11W and 9.5W deltas, we are really very far from your 30W, even if we would take the measurement at the main and starting from 1.6GHz since your delta is more than the APU TDP at 2.7....

Really?! I would love to see your measurements for Kabini @ 2.7Ghz in Prime 95!

 

[The Stilt]

Measurements i made are barely 15W with Prime 95 and 12.5W with Cinebench 11.5, at 2.05GHz since it s an Athlon 5350.

Did you make these "measurements" with a potato?

AMDs own measurements for GX-420CA (GE420CIAJ44HM) which has the same 25W TDP and 50MHz lower core clocks, measured on factory test equipment (Larne DAP w/ wide-band DAQ):

(APU Power Only)

Pov-Ray = 16.42W (avg), 17.08W (max)
AMD SST = 17.19W (avg), 20.34W (max)

AMD SST is roughly as stressfull as Linpack.

 

[ShintaiDK]

Did you make these "measurements" with a potato?

AMDs own measurements for GX-420CA (GE420CIAJ44HM) which has the same 25W TDP and 50MHz lower core clocks, measured on factory test equipment (Larne DAP w/ wide-band DAQ):

(APU Power Only)

Pov-Ray = 16.42W (avg), 17.08W (max)
AMD SST = 17.19W (avg), 20.34W (max)

AMD SST is roughly as stressfull as Linpack.

:thumbsup:

 

[Abwx]

Did you make these "measurements" with a potato?

AMDs own measurements for GX-420CA (GE420CIAJ44HM) which has the same 25W TDP and 50MHz lower core clocks, measured on factory test equipment (Larne DAP w/ wide-band DAQ):

(APU Power Only)

Pov-Ray = 16.42W (avg), 17.08W (max)
AMD SST = 17.19W (avg), 20.34W (max)

AMD SST is roughly as stressfull as Linpack.

I made the measurement with an Agilent measurement set if that s enough for your taste.

You should know that industrial usage dedicated products must have at least 20% voltage margin, that s why server chips use lower frequencies than DTs, so your power figures are higher anyway than what i measured.

Also you should know that AMD figures are at the higher temperatures, with the worst samples and with a 10% overvoltage, that s their max TDP rating conditions, you can find them in the 990/970 chipsets datasheets among others...

 

[The Stilt]

I made the measurement with an Agilent measurement set if that s enough for your taste.

You should know that industrial usage dedicated products must have at least 20% voltage margin, that s why server chips use lower frequencies than DTs, so your power figures are higher anyway than what i measured.

Also you should know that AMD figures are at the higher temperatures, with the worst samples and with a 10% overvoltage, that s their max TDP rating conditions, you can find them in the 990/970 chipsets datasheets among others...

How did you isolate the different voltage planes in order to make the measurements? Which voltage planes did you isolate / measure and which motherboard you were using?

The SoC AMD recorded these numbers with was a normal production part (retail) and the test conditions are specified extremely accurately. In this test they did not measure the power consumption or performance in elevated temperatures but in normal operating conditions.

GX-420CA is rated for 1.100V - 1.3000V in P0 PState (depending on the leakage of the specimen).

Athlon 5350 happens to have exactly the same specifications (1.100 - 1.3000V for P0). GX-420CA operates at 2000MHz in P0, whereas Athlon 5350 operates at 2050MHz in P0. Where are those "at least 20% voltage margins"?

Is there anything else I should know, since I´m obviously this badly informed

 

[FlanK3r]

you can to buys top Kabini, its 5370 Athlon. Maybe I put it to LN2 for some short fun

 

[coercitiv]

Maybe I put it to LN2 for some short fun

Be sure to post results if you do!

 

[FlanK3r]

I will be happy for 3 GHz

There are WR with 5350: http://hwbot.org/benchmark/superpi_...Id=processor_3594&cores=4#start=0#interval=20

 

[Abwx]

How did you isolate the different voltage planes in order to make the measurements? Which voltage planes did you isolate / measure and which motherboard you were using?

In FPU heavy tasks all the delta power is absorbed by the APU, you could as well just use good quality multimeters, in my case the MB has a DC input and there s only two DC conversion to get through (within the MB)..

The SoC AMD recorded these numbers with was a normal production part (retail) and the test conditions are specified extremely accurately. In this test they did not measure the power consumption or performance in elevated temperatures but in normal operating conditions.

GX-420CA is rated for 1.100V - 1.3000V in P0 PState (depending on the leakage of the specimen).

Athlon 5350 happens to have exactly the same specifications (1.100 - 1.3000V for P0). GX-420CA operates at 2000MHz in P0, whereas Athlon 5350 operates at 2050MHz in P0. Where are those "at least 20% voltage margins"?

Is there anything else I should know, since I´m obviously this badly informed

As i told you AMD numbers account for a 10% overvoltage, wich means that if the Athlon votage is specified for 1.3000V then the TDP is measured at 1.3000V + 10% = 1.4300V.

To be qualified a chip must be capable to work at its rated voltage + 10% to account for power supplies tolerances, when measured in such condition TDP will be 10% higher given that frequency is kept the same, this is even more important for the GX wich is used in industrial environments.

Likewise the chip must be capable to work at its rated frequency at 90% of the specified voltage, all this is standard enginering knowledge, i wont go further in this direction as it can be googled for whom is really interested in such issues...

 

[The Stilt]

I didn´t understand most of that gibberish, however you measured the power prior VRMIn? If so even more "impressive", since the AMD numbers have no conversion / conduction losses included.

This test was not made to determine the TDP, but to test the performance and the power consumption of the SKU in real world condition and in real world workloads. The very same data exists for all SKUs and is intended for their partners.

Please post a link to the source which supports your claim about the +10% voltage margin in TDP. I´ve never seen any indication of it in any of the official documents, for any AMD design, regardless of the level.

There is a margin, but that is negative and not positive (load-line, Rll). For 16h 00-0Fh parts Rll is 4.0mOhms, which means that at the TDC of 15A (for 25W parts) the "offset" is -60mV (0.004 * 15).

 

[Abwx]

I didn´t understand most of that gibberish, however you measured the power prior VRMIn? If so even more "impressive", since the AMD numbers have no conversion / conduction losses included.

This test was not made to determine the TDP, but to test the performance and the power consumption of the SKU in real world condition and in real world workloads. The very same data exists for all SKUs and is intended for their partners.

Please post a link to the source which supports your claim about the +10% voltage margin in TDP. I´ve never seen any indication of it in any of the official documents, for any AMD design, regardless of the level.

There is a margin, but that is negative and not positive (load-line, Rll). For 16h 00-0Fh parts Rll is 4.0mOhms, which means that at the TDC of 15A (for 25W parts) the "offset" is -60mV (0.004 * 15).

For voltage margins you can for instance check GF s presentation of the perfs of a 2C A57 using 14nm LPP.

Specified voltage is 0.800V and to be qualified the chip perf/watt is measured at 0.880V while voltage margin is measured a 0.720V.

That s typical for a consumer dedicated CPU, industrial and military requirements are one and two steps above respectively.

As for AMD s testbeds lots of things can be found, i have not the AMD datasheets on this PC but you can find some indications in rapidly googled links like the A77 chipset :

Thermal Design Power (TDP) is defined as the highest power dissipated while running currently available worst case applications at nominal voltages.

The core voltage was raised to 5% above its nominal value for measuring the ASIC power. Since the core power of modern ASICs using 65nm and smaller process technology can var significantly, parts specifically screened for higher core power were used for TDP measurement.

To get back on the Athlon 5350 it looks like my SKU was still quite average, there s only one review so far that (reliaby) displayed worse results than what i got, for instance here a test with several apps :

Even in Prime 95 the chip is at barely 15W....

 

[Shivansps]

You can overclock a lot on AM1 if the board supports OC and CPU vcore increase, just disable AHCI and you can go very high.

But that destryo the "25W" thing and im not sure how long VRMs will hold.

 

[The Stilt]

You can overclock a lot on AM1 if the board supports OC and CPU vcore increase, just disable AHCI and you can go very high.

But that destryo the "25W" thing and im not sure how long VRMs will hold.

How are your PCI-E devices working at those speeds?

EDIT: 1.632V, HOLY D: