[AMD Processors] The future CPUs

[AtenRa]

IPC and ST-performance (taking clockspeed into account) are nearly synonymous. And all I'm saying is, it needs to be a LOT more than 40% more ST/MT performance.

Well, give me a measurement of Excavator ST performance today to compare against SandyBridge and then we can talk about ZEN.

 

[Fjodor2001]

IPC and ST-performance (taking clockspeed into account) are nearly synonymous. And all I'm saying is, it needs to be a LOT more than 40% more ST/MT performance.

How have you arrived at that conclusion? Any calculations to base it on?

And how did you arrive at the estimate that Zen should be at SB level of performance per core?

 

[YBS1]

Seems people are expecting it to be in IB/Haswell performance range per core. Since it'll be competing against Broadwell-E, Zen should be pretty close. At least far closer to Intel 8 core HEDT than 6 core HEDT.

We'll see, the faster it is the better off they will be. Unless we are just stunned by the performance increase though I feel pretty certain we're going to see that 8c/16t chip being sold at mainstream pricing though. If not at introduction, soon after (FX9590).

 

[Yuriman]

Clocks are approximately the same, so you can just about compare them 1:1.

7850K vs 2500K:

Sandy Bridge is 63% faster

Sandy Bridge is 61% faster

Sandy Bridge is 94% faster

Sandy Bridge is 32% faster

Sandy Bridge is 43% faster

Sandy Bridge is 50% faster

Sandy Bridge is 53% faster

Sandy Bridge is 37% faster

Sandy Bridge is 54% faster

Sandy Bridge is 28% faster

 

[Yuriman]

Sandy Bridge is 34% faster

Sandy Bridge is 27% faster

Average advantage of Sandy Bridge over Kaveri at the same clocks: 48%

I didn't cherry pick; that's every test from Anandtech's 7850K review. So, assuming we'll have similar clockspeeds, we'll need around 48% higher ST and MT performance to reach Sandy Bridge levels.

Note that they're both 95w TDP chips, too. I can't speak for their actual power consumption, but it looks unfavorable.

 

[Yuriman]

How have you arrived at that conclusion? Any calculations to base it on?

And how did you arrive at the estimate that Zen should be at SB level of performance per core?

Talk on the forum has been that we'll be looking at similar to SB performance. I can't cite where that originated, but that's scuttlebutt.

Based on Anandtech's testing, if Zen is 40% faster than Kaveri, it will still be short of Sandy Bridge (clock for clock).

 

[AtenRa]

Clocks are approximately the same, so you can just about compare them 1:1.

7850K vs 2500K:

I said Excavator ST performance,

All those benchmarks you posted are about Steamroller Dual Module Quad Core MT performance vs Sandybridge Quad Core Performance.

 

[Yuriman]

I said Excavator ST performance,

All those benchmarks you posted are about Steamroller Dual Module Quad Core MT performance vs Sandybridge Quad Core Performance.

An almost irrelevant distinction. From what I've been able to gather (following the Carrizo thread), Excavator brought very little in the way of performance improvements outside of a power constrained environment; it was a significant improvement in power consumption. Also, the module penalty in Kaveri is almost non-existant, as you frequently point out, which makes ST performance more easily relatable to MT performance. So, comparing a 2500K to a 7850K is, within perhaps single digit percents, valid.

 

[AtenRa]

An almost irrelevant distinction. From what I've been able to gather (following the Carrizo thread), Excavator brought very little in the way of performance improvements outside of a power constrained environment; it was a significant improvement in power consumption. Also, the module penalty in Kaveri is almost non-existant, as you frequently point out, which makes ST performance more easily relatable to MT performance. So, comparing a 2500K to a 7850K is, within perhaps single digit percents, valid.

Comparing the 7850K with no L3 cache and with a module MT penalty (smaller than PileDriver but still present) to quad Core Sandy i5 is oranges to apples.
Just to give you an example why your comparison is wrong, take FX8350 (PileDriver) vs Sandy Core i7 2600 and you will come to a different conclusion.

Again, we have no ST performance of Excavator in order to be able to safely talk about ZEN yet.

 

[Yuriman]

How much L3 cache does Excavator have?

The best I can come up with is this comment:

These changes, according to AMD, relate to a 4-15% higher IPC for Excavator in Carrizo compared to Steamroller in Kaveri.

Source

Now, I'll admit that this is all rough figuring, but it's definitely ballpark. Sandy Bridge still looking to be 33-43% faster per clock.

 

[Fjodor2001]

Talk on the forum has been that we'll be looking at similar to SB performance. I can't cite where that originated, but that's scuttlebutt.

I've heard anything from Nahalem to Skylake. Most seem to place it at IB/Haswell level. But of course it's all speculation at this point.

Based on Anandtech's testing, if Zen is 40% faster than Kaveri, it will still be short of Sandy Bridge (clock for clock).

What Anandtech testing?

Also, remember that Sandy Bridge was 32 nm. Zen will be 14 nm. So the Zen clocks are likely to be significantly higher.

 

[AtenRa]

How much L3 cache does Excavator have?

The best I can come up with is this comment:



Source

Now, I'll admit that this is all rough figuring, but it's definitely ballpark. Sandy Bridge still looking to be 33-43% faster per clock.

There is no L3 cache Excavator CPU today, and im not expecting one in 2016 since those will be APUs as well.

 

[Yuriman]

I've heard anything from Nahalem to Skylake. Most seem to place it at IB/Haswell level. But of course it's all speculation at this point.

What Anandtech testing?

Also, remember that Sandy Bridge was 32 nm. Zen will be 14 nm. So the Zen clocks are likely to be significantly higher.

The testing that I linked in the post right before the one you quoted.

There is no L3 cache Excavator CPU today, and im not expecting one in 2016 since those will be APUs as well.

Then why did you bring up L3 cache as an issue when using Kaveri as an analogy to compare Carizzo to SB?

 

[Fjodor2001]

The testing that I linked in the post right before the one you quoted.

SB IPC at 14 nm clocks should place it as IB to Haswell performance level, right?

 

[Yuriman]

SB IPC at 14 nm clocks should place it as IB to Haswell performance level, right?

Sandy Bridge performance would perform exactly the same as Sandy Bridge, if it were shrunk to 14nm. It would just be smaller, and probably draw less power.

 

[mrmt]

SB IPC at 14 nm clocks should place it as IB to Haswell performance level, right?

Why do you assume Samsung SoC optimized process will clock as high as Intel server optimized process?

 

[ShintaiDK]

FX7600P 2.7Ghz base, 3.6Ghz turbo. (Steamroller)
FX8800P 2.1Ghz base, 3.4Ghz turbo. (Excavator)

Both 35W.

 

[MajinCry]

Ooh! OOH! Time for the new guy to blaze the ol' trails.

Then why did you bring up L3 cache as an issue when comparing SB to Kaveri?

Because yon new fangly-bangly-danglederry Zen is gon' have 'em L-threesies, aye? He's pointin' out one o' the ol' deficits ta be aware o', milawdship.

And he does have a point 'bout comparin' the ol' multithreaderies o' yore; El Bulldozeraro Familia De Cepusados has that 'ere module-multithreadin' penalty o' 10-15% minimum, if ah does rememberin' correctly. Zen winnae ha'e naen o' tha' lark.

 

[Abwx]

From what I've been able to gather (following the Carrizo thread), Excavator brought very little in the way of performance improvements outside of a power constrained environment

Lol, i guess that more gathering is necessary, you just proved that you didnt even made the effort to learn about Excavator, and are just throwing random assumptions..

 

[Yuriman]

Follow my logic here. I'll put it all in one place, to make it easier:

If

Average advantage of Sandy Bridge over Kaveri at the same clocks: 48%

I didn't cherry pick; that's (the average of) every test from Anandtech's 7850K review.

And

And

These changes, according to AMD, relate to a 4-15% higher IPC for Excavator in Carrizo compared to Steamroller in Kaveri.

Then we can expect

... it will still be short of Sandy Bridge (clock for clock).

And even if it's not, it's still ballpark. Sandy Bridge is well behind Skylake, as can be seen here.

Therefore

AMD will need more than 40% over their current uarch to compete with the same number of cores/threads. Because of this, I'm hoping the 40% figure is incorrect.

 

[Abwx]

FX7600P 2.7Ghz base, 3.6Ghz turbo. (Steamroller)
FX8800P 2.1Ghz base, 3.4Ghz turbo. (Excavator)

Both 35W.

So they had an APU that was 35W at 2.7GHz and, according to you, they decided that one that would eat those 35W at 2.1GHz would be better..

Did i understand well the fud you re spreading so others can take exemple and spread the "good" word..?..

 

[MiddleOfTheRoad]

Do you want to compare a delay to launch a product (Intel) with getting out of entire market segments (AMD)? So much for your actual objectivity at work here.

Damn nice. You should earn a Ph.d in moving the goal posts.

You were talking about delays -- and Intel has had more of them in recent history. So get off the soapbox and deal with it.

BTW, it's not like Intel is abandoning the Motherboard market or anything.... oh wait..... EPIC FAIL on your part.
http://www.zdnet.com/article/intel-to-leave-desktop-motherboard-business-by-2016/

Just more babble and fanboism's.

 

[Fjodor2001]

Sandy Bridge performance would perform exactly the same as Sandy Bridge, if it were shrunk to 14nm. It would just be smaller, and probably draw less power.

14 nm would also mean higher clocks. See e.g.

3770K@22nm: 3.5/3.9 Ghz
6700K@14nm: 4.0/4.2 Ghz

So the net performance will be higher of Zen too @ 14 nm. You'd end up somewhere between SB and Skylake, i.e. IB/Haswell.

 

[ShintaiDK]

So they had an APU that was 35W at 2.7GHz and, according to you, they decided that one that would eat those 35W at 2.1GHz would be better..

Did i understand well the fud you re spreading so others can take exemple and spread the "good" word..?..

You completely missed the point in your fury.

IPC increase doesnt come for free.

 

[Yuriman]

14 nm would also mean higher clocks. See e.g.

3770K@22nm: 3.5/3.9 Ghz
6700K@14nm: 4.0/4.2 Ghz

So the net performance will be higher of Zen too @ 14 nm. You'd end up somewhere between SB and Skylake, i.e. IB/Haswell.

On the other hand, we saw a decline of max clockspeed (when overclocked) from SB to IB to HW to BW.

But yeah, lots of assumptions there. My point is, with this "40%" AMD has given us, they're not going to be in parity with Intel.