Ryzen 3 3700x vs 3800x

[DuronBurgerMan]

And nope, The_Stilt wasn't the record holder, I misspoke on that. He was 72MHz from the world record.

https://www.kitguru.net/components/...ly-breaks-worlds-clock-rate-record-at-8-7ghz/

 

[Snowdog]

Good.
Remember when discussing power draw, we're not talking about just voltage. We are also talking about current and thermals, and thus implicitly resistance as well (remember Ohm's Law).

Of course, but in this case, lower supply voltage, equal less leakage current, equals less static power. Read the PDF you linked. The main ways to lower static leakage is lowering supply voltage (lowers clock speed), and increasing threshold voltage (also lowers clock speed). Better binned parts tend to run equal clock speeds at lower voltage, which should translate to lower leakage as well.

But of course measure to power to verify this. But I seriously doubt there is any consistent advantage to lower binned parts, even at lower clock speed. Low binned parts, aren't selected for their ability to use less power, they just contain more rejects that don't run at higher clock speeds. That might be different if these were specially binned mobile parts, but they aren't. It's simply 1800/X rejects limited in clock speed and clamped to lower power.

Also when I said there wasn't much binning. I should been clear that I meant at lower levels. Obviously a 1800X would need to be verified to reach it's stated speeds, so it eliminates the low performers, and raises the average, while rejects become 1700s lowering that average, but there were still cases where reviews received a 1700 that would OC marginally better than their 1800x. There would be much more variability at the 1700 level.

 

[chameleoneel]

The 3950x is two 8 core chiplets, for 105w.

In my opinion, the 3700x are probably 8 core chiplets which weren't quite power efficient enough to be in a 3950x, but still more power efficient than the worst chiplets. So they created a "low" power 8 core CPU. I bet basically every 3700x can match a 3800x.

The 3800x is probably whatever chips, fed a bunch of voltage to make sure they hit a certain peak. Kind of like what they have been doing with their videocards. And it will be a grab bag from there. Some 3800x will be able to undervolt and maintain 4.5Ghz. Some won't undervolt. Some will overclock a bit. etc.

As is, it looks like Zen 2 still has a bit of a power wall. Where at a certain point, it needs a whole lot more power, for not a large gain in MHZ.

 

[DuronBurgerMan]

Of course, but in this case, lower supply voltage, equal less leakage current, equals less static power.

Generally speaking, yes. But again, Ohm's Law applies with regard to overall resistance. The point is to eliminate as many variables as possible, and get us to the thing we really want to measure (re: my original claim): efficiency. For this, we should measure at the wall.

Read the PDF you linked.

I wouldn't link you to something I haven't read, my dude. All of the things I've been giving you are things I read back when I did my own deep dive into Ryzen.

The main ways to lower static leakage is lowering supply voltage (lowers clock speed), and increasing threshold voltage (also lowers clock speed). Better binned parts tend to run equal clock speeds at lower voltage, which should translate to lower leakage as well.

Here's where we have a disconnect. Higher leakage can sometimes unlock higher clocks in parts that may otherwise be somewhat less efficient in low power envelopes (i.e. also lower clocks). Or at least, that is The Stilt's original claim (to which I obviously agree).

But of course measure to power to verify this.

Agreed.

Also when I said there wasn't much binning. I should been clear that I meant at lower levels. Obviously a 1800X would need to be verified to reach it's stated speeds, so it eliminates the low performers, and raises the average, while rejects become 1700s lowering that average, but there were still cases where reviews received a 1700 that would OC marginally better than their 1800x. There would be much more variability at the 1700 level.

AMD could throw more CPUs that might otherwise pass as 1800Xs in the bin with the 1700s and 1700Xs, based on supply/demand. Old practice (remember when you could unlock the old extra cores on Phenoms?). But overall, the evidence of strong binning exists via the supplied HWBot link, and while exceptions to the trend may exist - for whatever reason - the trend is clear.

 

[DuronBurgerMan]

The 3950x is two 8 core chiplets, for 105w

In my opinion, the 3700x are probably 8 core chiplets which weren't quite power efficient enough to be in a 3950x, but still more power efficient than the worst chiplets. So they created a "low" power 8 core CPU. I bet basically every 3700x can match a 3800x.

The 3800x is probably whatever chips, fed a bunch of voltage to make sure they hit a certain peak. Kind of like what they have been doing with their videocards. And it will be a grab bag from there. Some 3800x will be able to undervolt and maintain 4.5Ghz. Some won't undervolt. Some will overclock a bit. etc.

As is, it looks like Zen 2 still has a bit of a power wall. Where at a certain point, it needs a whole lot more power, for not a large gain in MHZ.

Note the lower base clocks for the 3950X necessary to get within that power envelope. But I agree that these are better binned chiplets, probably. And I suspect the later release date for the 3950X is because AMD wants to build up a supply for Epyc.

The 6 core chiplets, OTOH, are fair game. And the 8 core chiplets that, for whatever reason, aren't good for Epyc wind up in the 3700X and 3800X (or, perhaps ones with higher static leakage, given the topic we're on).

And Zen 2 certainly has a power wall - all uarch's do. The question is figuring out where it is, exactly. We'll know when the official benchmarks and overclocking attempts are released. Can't wait! This is exciting shit (and I'm really considering dropping a 3900X in this box if the reviews are good).

 

[BrotherMichigan]

EPYC is already shipping to some customers, so it's probably more they're trying to guarantee that supply rather than trying to stockpile chiplets for use.

 

[DuronBurgerMan]

EPYC is already shipping to some customers, so it's probably more they're trying to guarantee that supply rather than trying to stockpile chiplets for use.

Makes sense. But that is stockpiling of a sort.

 

[Thatguybil]

I’ve learned a lot reading some of the above comments.

In my personal experience chips that can OC the best are typically the same chips that are stable at lower voltages.
IE a chip that is spec 4.2 ghz at 1.4 volts that is rock solid at 1.15 volts is more likely to be able to hit 5 ghz at 1.45 V.
If a chip is spec at 4.2 at 1.4 volts and can only be under volted to 1.35 volts is not likely going to OC very well.

Some of this gets turned on its head when you talk about extreme L2 overclocking. That chip that needed more voltage might actually work better at negative temperatures. Crazy.