Athlon X4 845 ExcavatorIPC benchmarks

[nenforcer]

This AMD XV x4 845 is a nice tease for the current FM2+ boards out there. Correct me if I'm wrong, but isn't Bristol Ridge suppose to release 2Q 2016 still on .28nm with a socket FP4 mobile / AM4 desktop?

I'm still concerned about AMD upcoming AM4 "Promontory" chipset - I've heard rumors its been designed by ASMedia which would seem unnecessary given that A88X (Bolton) had full support for PCI-E 3.0 but just with a 3rd party(Renaesys?) designed USB 3 interface.

 

[TheELF]

what graph? there's no graph in arachnotronic's post.

Did I quote arachnotronic?

Maybe,next READ what the guy was saying instead of reacting first. He tried to say that AMD was changing its claims. Don't try to defend his error.

Guy who?
YOU posted a graph and YOU said that amd is factual,I responded by saying that releasing a graph with no numbers on it(other then the 40% ) is not factual.

 

[ElFenix]

Did I quote arachnotronic?


Guy who?
YOU posted a graph and YOU said that amd is factual,I responded by saying that releasing a graph with no numbers on it(other then the 40% ) is not factual.

arachnotronic posted the image which was being discussed. there's no graph on it. can you point me to a graph posted by USER8000?

 

[looncraz]

I have prepared a place for my interrogation of the x4 845 when it arrives in a few days. I am just running the Steamroller benchmarks, and have completed Sandy Bridge and K10 benchmarks, everything at 3Ghz.

There's nothing to see at this point in time, other than my general theme and the construction notice, however Wednesday or Thursday should see the raw data portion come online, and Friday should see the Zen simulation data and theoretical performance estimation come online.

http://excavator.looncraz.net/

BTW, I'm looking for which games I should benchmark, I'm looking at Hitman:Absolution and Batman: Arkham City, for starters, as these games are already installed on my HTPC. I will also benchmark 3dMark Vantage & 11, as well as Unigine's Heaven and Valley. Game benchmarks will just be comparing Steamroller to Excavator - both at 3Ghz.

I will also look to see if I can get the two processors to reliably stick to 3.8Ghz without throttling, and I will be testing the voltage limits of the x4 845 as well as comparing power consumption at 3GHz and 3.8Ghz. My voltage data only has 6W granularity, but it could still tell us something :thumbsup:

 

[jhu]

I have prepared a place for my interrogation of the x4 845 when it arrives in a few days. I am just running the Steamroller benchmarks, and have completed Sandy Bridge and K10 benchmarks, everything at 3Ghz.

There's nothing to see at this point in time, other than my general theme and the construction notice, however Wednesday or Thursday should see the raw data portion come online, and Friday should see the Zen simulation data and theoretical performance estimation come online.

http://excavator.looncraz.net/

BTW, I'm looking for which games I should benchmark, I'm looking at Hitman:Absolution and Batman: Arkham City, for starters, as these games are already installed on my HTPC. I will also benchmark 3dMark Vantage & 11, as well as Unigine's Heaven and Valley. Game benchmarks will just be comparing Steamroller to Excavator - both at 3Ghz.

I will also look to see if I can get the two processors to reliably stick to 3.8Ghz without throttling, and I will be testing the voltage limits of the x4 845 as well as comparing power consumption at 3GHz and 3.8Ghz. My voltage data only has 6W granularity, but it could still tell us something :thumbsup:

Would you mind running Blender with these two files?

 

[looncraz]

Would you mind running Blender with these two files?

Sure thing!

 

[jhu]

Sure thing!

Thank you!

 

[tweakboy]

Intel destroys amd.

 

[MajinCry]

-snip-
BTW, I'm looking for which games I should benchmark, I'm looking at Hitman:Absolution and Batman: Arkham City, for starters, as these games are already installed on my HTPC. -snip-

What I'd suggest, is testing out Bethesda's games; badly coded, lots of background processing, lots of draw calls, and easily modifiable.


For example, take Oblivion:

Jack all the settings up to max
Install an intensive ENB mod (Direct3D 9 wrapper for better visuals)
Enable water reflections - ~15 fps hit on my Phenom II, even when there's no water (e.g, in the Imperial City Market District.)
Go to the Market District @ noon
Press `, type in "TGM", then press enter
Type in "player.placeatme 000055BD 20", then press enter.
Press ' again.
Watch the framerate die.

Oblivion is almost entirely single-threaded (assigning Oblivion.exe to two cores gives 5 more fps over one core, assigning three or four cores gives no performance gains), so it'd make for a good test for ST performance. And since it uses forward rendering, whenever there are lights in the scene, every object is rendered again. Draw calls go through the roof when the NPCs and player holds a torch.

Skyrim would be another good one, due to the ease of jacking up the AI.

Jack all the settings up to max
Install an intensive ENB mod
Install the More Spawns - Scriptless mod made by yours truly (3x as many bandits, for example)
Press ', type in "TGM", then press enter
Type in "player.setlevel 50"
Press ' again
Go to any bandit fort (Fort Greymoor is close to Whiterun)
Press ' again
Type in 'player.placeatme 000F811C 5", then press enter.
Press ' again
And watch the framerate die.

ENB's GUI, in post-Oblivion games, allows you to monitor the draw calls made for one frame, every second or so, if that piques yer interest.

 

[PG]

Here is a review: http://www.jagatoc.com/2016/03/hands-on-review-overclocking-athlon-x4-845-carrizo-murah-tanpa-igp/

 

[The Stilt]

From Jagat: "The most efficient AMD processor, but only can be overclocked a bit
What you did to me was ... NASTY"

Since AMD has always struggled with the L2 caches, I bet they still do despite reducing the L2 size. All 15h designs the Fmax is limited by the L2 cache.

On Steamroller the L2 latency was 19 cycles and the absolute Fmax of the design is in 4.6GHz range. Excavator meanwhile has 17 cycle L2 latency... (17 / 19) * 4600 = ~4.1GHz :sneaky:

Most likely HDL and GF28HPP instead of GF28A reduced the Fmax too, but I don't think the Fmax reducing by the same percentage as the L2 got faster is a coincidence.

I estimated 4.2GHz Fmax for Carrizo without even testing D:

 

[Abwx]

The chip is not unlocked, they simply increased the BCLK...

Also i once warned that voltage means mothing in isolation, and that power can be lower with higher voltages if parasitic capacitances are reduced accordingly, prove is brought by the 845 althouh it s was already the case for their competitor..

 

[naukkis]

Since AMD has always struggled with the L2 caches, I bet they still do despite reducing the L2 size. All 15h designs the Fmax is limited by the L2 cache.

On Steamroller the L2 latency was 19 cycles and the absolute Fmax of the design is in 4.6GHz range. Excavator meanwhile has 17 cycle L2 latency... (17 / 19) * 4600 = ~4.1GHz :sneaky:

Most likely HDL and GF28HPP instead of GF28A reduced the Fmax too, but I don't think the Fmax reducing by the same percentage as the L2 got faster is a coincidence.

I estimated 4.2GHz Fmax for Carrizo without even testing D:

1MB L2-cache latency is also 17 cycles for Piledriver and Steamroller. It takes that two extra cycles to check that extra MB.

 

[The Stilt]

1MB L2-cache latency is also 17 cycles for Piledriver and Steamroller. It takes that two extra cycles to check that extra MB.

Source?
AMD documents state >= 20 cycles for Bulldozer and Piledriver, >= 19 cycles for Steamroller and >= 17 cycles for Excavator.

 

[el etro]

HDL doing well.

Power score

http://www.jagatoc.com/2016/03/hands-on-review-overclocking-athlon-x4-845-carrizo-murah-tanpa-igp/7/

 

[naukkis]

Source?
AMD documents state >= 20 cycles for Bulldozer and Piledriver, >= 19 cycles for Steamroller and >= 17 cycles for Excavator.

Didn't find source now, but if I don't remember wrong AMD has documented 2 cycles lower latency for 1MB cache modules for whole dozer family.

And if they have same latency for 1MB and 2MB caches their designers really are incompetent.

 

[The Stilt]

Didn't find source now, but if I don't remember wrong AMD has documented 2 cycles lower latency for 1MB cache modules for whole dozer family.

And if they have same latency for 1MB and 2MB caches their designers really are incompetent.

What 1MB L2 caches are you talking about? No 15h desing prior Excavator have 1MB L2 cache by the design. Some harvested models have less, but that's because 256 (of 512 within a CU) 4K blocks have been disabled. Disabling the cache blocks should have no effect on the latency.

I think it is safe to rely on #47414 document in this matter. Some older versions should be available in public too.

 

[naukkis]

What 1MB L2 caches are you talking about? No 15h desing prior Excavator have 1MB L2 cache by the design. Some harvested models have less, but that's because 256 (of 512 within a CU) 4K blocks have been disabled. Disabling the cache blocks should have no effect on the latency.

I think it is safe to rely on #47414 document in this matter. Some older versions should be available in public too.

AMD documents have had both 1MB and 2MB L2 per module versions from early Bulldozers, and there has always been 2 cycle variation for L2 latencies. Sure, there ain't have been 1MB versions in production until Carrizo but 2 cycle latency decrease from halving L2 should be considered minimum, Intel Prescott has very similar L2-latencies, 23 cycle for 1MB and 27 cycle for 2M.

 

[dark zero]

Was supposed to be a dissaster... It ended...

Quite decent.
Call me crazy, but AMD managed to get Phenom II levels without L3 cache.

If AMD were smart... They should included L3 cache on Bristol Ridge... It would reach Core i3 levels again, at least on MT and not OCed.

Also, I want to see a Sempron Quad Core XV at 3.2Ghz at 50 dollars and a Dual Core Duron at 2.8Ghz at less than 40 dollars. It would sell decently.

 

[The Stilt]

The L3 in 15h is so slow that it makes basically no difference. Including the L3 makes the NB significantly more complex and inflates the die size quite significantly.

Hard to imagine that the L3 would make a difference large enough to justify the extra die area, even on the cheap 28nm process. Most likely it would help Excavator based design more than any of the previous iterations due the smaller L2 cache, however the difference would still be very minor.

The 15h architecture is finally dead and it should be let to turn back to ashes.

 

[SPBHM]

The L3 in 15h is so slow that it makes basically no difference. Including the L3 makes the NB significantly more complex and inflates the die size quite significantly.

Hard to imagine that the L3 would make a difference large enough to justify the extra die area, even on the cheap 28nm process. Most likely it would help Excavator based design more than any of the previous iterations due the smaller L2 cache, however the difference would still be very minor.

The 15h architecture is finally dead and it should be let to turn back to ashes.

I remember seeing tests where the FX 4300 (4MB l3) was beating the APUs with the same cores(trinity/richland) clearly in games, the AM3+ l3 is still probably better than using just ram.

 

[monstercameron]

Meh.

 

[itsmydamnation]

I remember seeing tests where the FX 4300 (4MB l3) was beating the APUs with the same cores(trinity/richland) clearly in games, the AM3+ l3 is still probably better than using just ram.

the L3 is an eviction cache, the fact that it is there and significantly faster then main memory still provides a big advantage to workloads that are bigger then the L2 size.

CON cores cache hasn't performed well but the L3 is the least of the problems, it doesn't operate like intels L3 does, so there is no point comparing its numbers directly.

Now it looks like the L3 in Zen will act far closer to intel then CON cores does* so the latency and throughput numbers will be more comparable.

*there are still pretty big differences, one cares about data locality the other doesn't, ones tied to a core the other to a compute unit etc.

 

[Abwx]

The L3 in 15h is so slow that it makes basically no difference.

Only difference between the X4 840 and the X4 955 on the average at the bottom of the page is the L3 cache, frequency is the same...

http://www.hardware.fr/articles/913-7/cpu-performances-applicatives.html

Even in FP there s a difference if we look at 3DS Max.

 

[DrMrLordX]

Those are not 15h CPUs.