Athlon X4 845 ExcavatorIPC benchmarks

[jhu]

I think he's pointing out that having an L3 does help.

 

[PG]

I just got an 845 in myself to play with. My previous cpu was an A8-7650K. With the A8 I normally disable turbo and set the multiplier to 41 for 4.1 Ghz. For the sake of some benchmarks I ran it at stock and also at the 4.1 Ghz. Then I swapped in the 845. I guess I was oblivious to the fact that the 845 doesn't have the 2400Mhz memory multiplier, so I had to run my RAM at 2133. I did not go back and run the A8 benchmarks again with the ram at the lower 2133 because it would be too much work and I was lazy. So this isn't a completely fair comparison.
Note I am not even remotely a professional reviewer. I did not do a clean install of windows before running benchmarks with each cpu. I also ran things once and that was it. No multiple runs.
Anyway, here are the numbers. Sorry about the bad formatting:

Hardware:
A88XM-A
16GB DDR3 HYPERX BEAST, 2400Mhz, 11-13-14-32 for A8-7650K
16GB DDR3 HYPERX BEAST, 2133Mhz, 11-13-30-30 for 845
EVGA GTX 960 SSC 2GB
A8-7650K at stock and also 4.1 GHz
AMD Athlon 845 at stock
Auto voltages for both cpus

CPU-Z SINGLE MULTI
A8 STOCK 1073 4004
A8 4.1 1193 4738
845 1203 4563

RISE OF THE TOMB RAIDER
OVERALL
A8 STOCK 53.28
A8 4.1 54.5
845 53.77

ARKHAM KNIGHT
MIN MAX AVG
A8 STOCK 45 79 63
A8 4.1 48 87 67
845 39 74 57


Note Rise of the Tomb Raider just got an update with a built in benchmark which is nice. There are more and specific results for the 3 areas in the map that are benchmarked, but I just went with the Overall since it was easier.

 

[looncraz]

My x4 845 vs A8 7600 vs i5 2500k vs Phenom II X4 showdown is mostly available online in the form of a comprehensive performance interrogation.

http://excavator.looncraz.net/

Mind you, my interests in generating these numbers is in creating as thorough of a Zen performance guesstimate as possible.

I did all of the tests at 3Ghz, but also did some stock-speed testing. I did not even attempt overclocking, because I don't care about that. I did three games, over two dozen benchmarks, and a lot of graphics work

I will be retesting some of my x4 845 power consumption numbers because they seem far too low after looking at them, but I should have that done within 24 hours.

 

[dark zero]

My x4 845 vs A8 7600 vs i5 2500k vs Phenom II X4 showdown is mostly available online in the form of a comprehensive performance interrogation.

http://excavator.looncraz.net/

Mind you, my interests in generating these numbers is in creating as thorough of a Zen performance guesstimate as possible.

I did all of the tests at 3Ghz, but also did some stock-speed testing. I did not even attempt overclocking, because I don't care about that. I did three games, over two dozen benchmarks, and a lot of graphics work

I will be retesting some of my x4 845 power consumption numbers because they seem far too low after looking at them, but I should have that done within 24 hours.

So it's confirmed... eliminating the L3 cache and lowering the L2 cache HEAVILY hurt AMD to the point that in some test is below Piledriver (I checked some test and yeah... even Piledriver is best)

However, AMD managed to improve their ST performance BIG TIME.
If only Bristol Ridge were having some L3 cache if they can't deliver more L2 cache.... it would improve notably the results.... at least is an improved Carrizo and the GPU won't have any issues (defeating again Skylake iGPU without edRAM).

Definately the worst aspect is the Memory performance... is beyond mediocre.

 

[el etro]

So it's confirmed... eliminating the L3 cache and lowering the L2 cache HEAVILY hurt AMD to the point that in some test is below Piledriver (I checked some test and yeah... even Piledriver is best)

However, AMD managed to improve their ST performance BIG TIME.
If only Bristol Ridge were having some L3 cache if they can't deliver more L2 cache.... it would improve notably the results.... at least is an improved Carrizo and the GPU won't have any issues (defeating again Skylake iGPU without edRAM).

Definately the worst aspect is the Memory performance... is beyond mediocre.

Memory always was bad on CON cores. It just got better and better at each new uarch, even with latency going back up.

On-Die L3 takes much space from where's supposed to be GPU on the APUs, but surely would help a little in most benches. But it was cut from SR/Carrizo just because is slow compared to another designs(Phenom II, Intel big cores) cache.

 

[Yuriman]

My x4 845 vs A8 7600 vs i5 2500k vs Phenom II X4 showdown is mostly available online in the form of a comprehensive performance interrogation.

http://excavator.looncraz.net/

Mind you, my interests in generating these numbers is in creating as thorough of a Zen performance guesstimate as possible.

I did all of the tests at 3Ghz, but also did some stock-speed testing. I did not even attempt overclocking, because I don't care about that. I did three games, over two dozen benchmarks, and a lot of graphics work

I will be retesting some of my x4 845 power consumption numbers because they seem far too low after looking at them, but I should have that done within 24 hours.

Thanks for your testing!

I understand you've chosen your min/max values to make differences more visible, and not to be deceptive, but I'm wary of using 90-110% scales like this:

It suggests to the reader that even a 4% improvement is big, when it's almost certainly imperceptible. Visually, 10% faster is being represented with a bar that's twice as tall, while that's also very likely still in the range of imperceptibility.

Just my 2 cents, but I feel a sense of scale is very important when making comparisons.

 

[Insert_Nickname]

So it's confirmed... eliminating the L3 cache and lowering the L2 cache HEAVILY hurt AMD to the point that in some test is below Piledriver (I checked some test and yeah... even Piledriver is best)

However, AMD managed to improve their ST performance BIG TIME.
If only Bristol Ridge were having some L3 cache if they can't deliver more L2 cache.... it would improve notably the results.... at least is an improved Carrizo and the GPU won't have any issues (defeating again Skylake iGPU without edRAM).

Definately the worst aspect is the Memory performance... is beyond mediocre.

Thanks for the write-up. I got my 845 yesterday, but haven't had time to put it thoroughly though its paces.

One thing I -did- notice. Running the built-in CPU-Z benchmark, the 845 managed to get within striking distance (4580 vs 5100) of of the provided reference Bulldozer 8150 system. That's actually pretty impressive with half the modules, half the L2 cache and no L3 cache. Single thread isn't even a contest (1219 vs 803). Again quite impressive.

If only AMD had launched something like Excavator back in 2011. Excavator really is what the construction cores should have been from the start... :sneaky:

 

[DrMrLordX]

I think he's pointing out that having an L3 does help.

Yes, but it helps based on how fast it is/what is its latency etc. Stars chips had l3 that overall did more to help than Vishera's (for example). So using them as a comparison to explain why L3 is good on Vishera is a bad idea.

My x4 845 vs A8 7600 vs i5 2500k vs Phenom II X4 showdown is mostly available online in the form of a comprehensive performance interrogation.

Thanks for the data! The more you know . . .

edit: not a bad comparison, but it probably would have been better to use a Stars x2 and a SB/IB/HW i3 for this comparison if you were trying to get some Zen-like data? A stars quad or a SB i5 would be comparing directly against a Zen 4c/8t chip, for example, while the XV you've got is 2m/4t . . . and at least based on the early analysis of Zen that I've seen, it appears that each Zen core will be wider than an XV module.

 

[looncraz]

Thanks for your testing!

I understand you've chosen your min/max values to make differences more visible, and not to be deceptive, but I'm wary of using 90-110% scales[...]

You're welcome :thumbsup:

Sadly, if I use the full scale, the results are very difficult to even distinguish. I tried to use the same range for most of the charts to help reduce the perception impact, but I was much aware of it.

 

[looncraz]

Thanks for the data! The more you know . . .

edit: not a bad comparison, but it probably would have been better to use a Stars x2 and a SB/IB/HW i3 for this comparison if you were trying to get some Zen-like data? A stars quad or a SB i5 would be comparing directly against a Zen 4c/8t chip, for example, while the XV you've got is 2m/4t . . . and at least based on the early analysis of Zen that I've seen, it appears that each Zen core will be wider than an XV module.

I had what I had

That said, I assume that the removal of the module penalty and other factors are all included in AMD's 40% IPC figure, so I'm treating it like a cure-all for IPC estimates, with the exception of the performance profile itself (which benchmarks are likely to see the most benefit).

I think Zen will have a performance profile somewhere between K10 and Sandy Bridge, based on various aspects of its known design. The removal of internal bottlenecks that impact certain benchmarks more than others could result in a near doubling of performance in those specific cases while still only seeing a 40% IPC increase. In some cases, a 40% IPC increase will not give a 40% increase.

It's an interesting experiment into what IPC actually means for performance. How wrong or right I am will be determined by Zen's performance profile, which I will also be testing once I get one (I won't be an early adopter, though, unless people donate the big bucks :biggrin::whiste::sneaky

 

[jhu]

I had what I had

That said, I assume that the removal of the module penalty and other factors are all included in AMD's 40% IPC figure, so I'm treating it like a cure-all for IPC estimates, with the exception of the performance profile itself (which benchmarks are likely to see the most benefit).

I think Zen will have a performance profile somewhere between K10 and Sandy Bridge, based on various aspects of its known design. The removal of internal bottlenecks that impact certain benchmarks more than others could result in a near doubling of performance in those specific cases while still only seeing a 40% IPC increase. In some cases, a 40% IPC increase will not give a 40% increase.

It's an interesting experiment into what IPC actually means for performance. How wrong or right I am will be determined by Zen's performance profile, which I will also be testing once I get one (I won't be an early adopter, though, unless people donate the big bucks :biggrin::whiste::sneaky

Clearly you're a shill for showing Excavator in such a bad light compared with i5 2500k. Also Cinebench+Winrar is the definitive, end-all-be-all of benchmarks.

BTW, could you do some Blender benchmarks?

 

[VirtualLarry]

One thing I -did- notice. Running the built-in CPU-Z benchmark, the 845 managed to get within striking distance (4580 vs 5100) of of the provided reference Bulldozer 8150 system. That's actually pretty impressive with half the modules, half the L2 cache and no L3 cache. Single thread isn't even a contest (1219 vs 803). Again quite impressive.

Considering how my G4400 does in CPU-Z benchmark (2250 / 4000), that's not too shabby for just a dual-core, if it's getting close to an 8150 in MT.

Edit: Sorry for the OT, hope it doesn't derail the thread.

I'm waiting for AM4 to mess with XV.

 

[myocardia]

nvm

 

[looncraz]

Clearly you're a shill for showing Excavator in such a bad light compared with i5 2500k. Also Cinebench+Winrar is the definitive, end-all-be-all of benchmarks.

BTW, could you do some Blender benchmarks?

I don't control what the data says, I just generate it ()

I wanted to do some Blender, but I forgot to do it with the A8 7600, so I won't be able to compare. I can run the numbers just on the x4 845, though, if you want. I'm not putting the A8 7600 back in, I plan to sell it

 

[jhu]

I don't control what the data says, I just generate it ()

I wanted to do some Blender, but I forgot to do it with the A8 7600, so I won't be able to compare. I can run the numbers just on the x4 845, though, if you want. I'm not putting the A8 7600 back in, I plan to sell it

I'd appreciate it. Thanks!

 

[deasd]

X4 845 Linux result

http://openbenchmarking.org/result/1603119-GA-1401227SO30

 

[Insert_Nickname]

Considering how my G4400 does in CPU-Z benchmark (2250 / 4000), that's not too shabby for just a dual-core, if it's getting close to an 8150 in MT.

If anything that just goes to show just how bad the original Bulldozer core really was.

Slightly OT, but I think of the construction cores as dual-core-with-super-HT. If you view a module as a "core", they're not bad for what they are. Also another reason to stay away from the single module variety.

I'm waiting for AM4 to mess with XV.

XV+GCN1.3 IGP+DDR4 could be a good budget alternative to Pentium/Celeron and especially the Atom derived variety of those. We'll see.

I think I saw rumours somewhere that AMD is intending to replace the Cat cores with single module XV. If they do that and can clock them high enough (>3GHz with a 6-10W TDP) the budget market served by Cat cores and Bay Trail/Braswell derivatives could be in for a bit of revolution.

 

[The Stilt]

X4 845 Linux result

http://openbenchmarking.org/result/1603119-GA-1401227SO30

4.42% faster than Steamroller w/ extremities removed, 8.52% without. Pretty much as advertized by AMD (+5% IPC).

 

[Shehriazad]

Somehow the gaming results are disappointing...I kind of expected it to actually somewhat beat a 3.7/4Ghz Athlon 860k.

 

[dark zero]

Somehow the gaming results are disappointing...I kind of expected it to actually somewhat beat a 3.7/4Ghz Athlon 860k.

Remember that Bristol Ridge is about to come and is the FULL Excavator.

 

[DrMrLordX]

I had what I had

Right, the point being that if you have a 2m/4t chip, running it against a 4c/8t chip with SMT permanently disabled (read: an i5) doesn't really tell you much about what Zen is going to be like. It only tells you about XV itself, in a 2m configuration, which makes it a competitor for newer i3s. It might tell you what Zen will be like in a 2c/4t configuration (read: 40% faster than 2m/4t XV), but that's it.

That said, I assume that the removal of the module penalty and other factors are all included in AMD's 40% IPC figure

The "module penalty" is mostly, if not completely, gone in Steamroller and XV. You don't get superficial slowdown from loading up 4 threads on 2m SR or XV.

I think Zen will have a performance profile somewhere between K10 and Sandy Bridge, based on various aspects of its known design.

If that's the case, then Keller should just hang himself now.

It's an interesting experiment into what IPC actually means for performance.

If AMD is telling the truth, then it means +40% throughput for Zen over XV on a per-thread basis. If they are lying, then it could possibly lead to Zen having lower throughput than XV on a per-thread basis.

 

[looncraz]

Right, the point being that if you have a 2m/4t chip, running it against a 4c/8t chip with SMT permanently disabled (read: an i5) doesn't really tell you much about what Zen is going to be like. It only tells you about XV itself, in a 2m configuration, which makes it a competitor for newer i3s. It might tell you what Zen will be like in a 2c/4t configuration (read: 40% faster than 2m/4t XV), but that's it.


The "module penalty" is mostly, if not completely, gone in Steamroller and XV. You don't get superficial slowdown from loading up 4 threads on 2m SR or XV.


If that's the case, then Keller should just hang himself now.


If AMD is telling the truth, then it means +40% throughput for Zen over XV on a per-thread basis. If they are lying, then it could possibly lead to Zen having lower throughput than XV on a per-thread basis.

The module penalty is very much still there and can be seen in the relative multi-threaded performance in reference to Phenom II or the i5.

With Phenom II as the 1.0 reference, the x4 845 does the following:

CB R10 ST: 0.94
CB R10 MT: 0.87
CB R11 ST: 1.02
CB R11 MT: 0.94
CB R15 ST: 1.01
CB R15 MT: 0.88

That trend continues for every ST/MT result aside from CPU-Z and the Peacekeeper browser test.

Some of that is because of the shared FPU, specifically, but the integer results show it as well. The reduced cache versus Kaveri also has a role, but it is smaller (2% or so).

(posted early, sorry)

In regards to the performance profile, it is an offset from the 40% baseline.

However, adding 40% to Excavator blindly, puts the results all over the place. Applying a profile that adjusts them in a more likely manner, but forcing the total net performance gain to be a mere 40%, looks shockingly like Sandy Bridge - in fact, the scores are often nearly identical.

Even with this, we see some FPU/SIMD tasks with a 60%+ improvement, and pure integer being a mere 20%. That was necessary to keep things at 40% total... which is not likely from a 40% IPC module-free design.

I am currently working on the profile that assumes that 40% is, genuinely, an IPC increase - for both integer and floating point - and then I will take away the module penalties, and the share FPU penalties, and then add a bonus to SIMD-heavy workloads. Finally, I will provide an assumed boost from the process itself.

After all that, I will have a range for every benchmark and I will create an averaged performance estimate, then call it a quits and see where things land when Zen shows up at my door, provided I don't go broke before then

 

[NTMBK]

then call it a quits and see where things land when Zen shows up at my door, provided I don't go broke before then

So you're in the same boat as AMD :awe:

Great work, thanks a lot for digging into this! You should try selling your article to a tech news site.

 

[Yuriman]

When the "40%" figure first came out, I glanced through benchmarks and came to the same conclusion that Zen IPC would likely average somewhere between Sandy Bridge and Haswell, tending toward Sandy Bridge.

It's been said before, and everyone here knows it, but I'll say it again for context: Until we also know clocks and core counts, we really have no idea about overall performance.

My performance needs are not actually all that high lately. If Zen can deliver better-than-Sandy IPC, even at relatively low clocks (let's say 2.5-3.2ghz) with 14nm-competitive performance-per-watt, comes with a minimal iGPU that has AMD's software and drivers, and assuming it's priced similarly to equivalently performing Intel chips, it would be a very attractive chip to me.

 

[Ancalagon44]

My x4 845 vs A8 7600 vs i5 2500k vs Phenom II X4 showdown is mostly available online in the form of a comprehensive performance interrogation.

http://excavator.looncraz.net/

Mind you, my interests in generating these numbers is in creating as thorough of a Zen performance guesstimate as possible.

I did all of the tests at 3Ghz, but also did some stock-speed testing. I did not even attempt overclocking, because I don't care about that. I did three games, over two dozen benchmarks, and a lot of graphics work

I will be retesting some of my x4 845 power consumption numbers because they seem far too low after looking at them, but I should have that done within 24 hours.

Thanks, very interesting.

I still find it amazing that even with the 3rd and 4th revision to the Bulldozer architecture, AMD has not matched the IPC of Thuban. Let alone anything Intel has.

I really really hope Zen delivers.