AMD Q4/2013 Desktop Roadmap

[sniffin]

Granted it doesn't spend much time there, but it does otherwise stay pegged at 3.2GHz in games, I have seen this myself. Not bad for a desktop power hog

It is virtually never at 2.5GHz base. 2.8GHz is as low as it will ever go

The 25W sku has a base of 2.1 and turbo of 2.9. I have no experience with it but I wouldn't be surprised if it also regularly sits above base. Kind of makes the power hog label pretty silly doesn't it, especially considering that this is shared with 384 VLIW shaders

 

[mrmt]

Granted it doesn't spend much time there, but it does otherwise stay pegged at 3.2GHz in games, I have seen this myself. Not bad for a desktop power hog

It is virtually never at 2.5GHz base. 2.8GHz is as low as it will ever go

Could you please provide data on both clocks and power consumption? AMD isn't the most reliable company when talking about power consumption.

 

[sniffin]

Yeah let me just go spend 3 hours putting together data just to prove you wrong. Lol no thanks. Praise be to intel!

 

[ShintaiDK]

FX4300 vs 5750M in Cinebench MT.

FX4300 ~3.25
5750M ~2.20

Seems like the 5750M cant turbo under that 4T load.

Its funny enough no different in x264 as well, both pass 1 and pass 2.

 

[Enigmoid]

Granted it doesn't spend much time there, but it does otherwise stay pegged at 3.2GHz in games, I have seen this myself. Not bad for a desktop power hog

It is virtually never at 2.5GHz base. 2.8GHz is as low as it will ever go

The 25W sku has a base of 2.1 and turbo of 2.9. I have no experience with it but I wouldn't be surprised if it also regularly sits above base. Kind of makes the power hog label pretty silly doesn't it, especially considering that this is shared with 384 VLIW shaders

Know for a fact the a10-6800k runs at 4.4 ghz (max turbo). The a10-5750m is the exact same die with all the same parts enabled but different speeds. Yet the a10-5750m only gets 64% the performance. The a10-5750m that AT tested was probably running around 2.8 ghz in that test. An that is a test where short term turbo would be used.

Notebookcheck confirms

http://www.notebookcheck.net/Review-HP-Pavilion-17-e054sg-Notebook.105008.0.html

Cinebench's single-thread tests measure the processor at 3.2 GHz. Occasionally the core accelerates to 3.5 GHz. In multi-thread tests, the cores achieve a speed of 2.8 GHz but continually slow down to 2.5 GHz.... As both the Cinebench tests and our own stress test show, the CPU cannot deliver its full speed in multi-core usages.

http://www.notebookcheck.net/Review-MSI-GX70H-A108972811B-Notebook.99516.0.html
(GX 70)

We checked the CPU overclocking (Turbo Core 3.0) with Cinebench R10 and the Unigine Heaven 2.1 benchmark. Tools like HWiNFO and CPU-Z quickly show that AMD's technology is not working as well as Intel's version. 3.2 GHz in the Single-Core test and 2.8 GHz in the Multi-Core test are decent but not really overwhelming. The clock also fluctuates between 2.8 and 3.2 GHz and can even drop to 2.5 GHz for short periods with simultaneous GPU load. Intel's Turbo Boost technology usually enables a significantly higher clock in our experience.

 

[sniffin]

Why are you using synthetic benchmarks that place 100% load on every thread as examples for me being wrong? You've proved my point. 2.8GHz in synthetics makes 3.2/3.5GHz in user workloads like gaming pretty plausible. Unless people are buying mobile parts for Cinebench, encoding is probably the only time somebody wold ever experience base clocks

 

[blackened23]

Yeah let me just go spend 3 hours putting together data just to prove you wrong. Lol no thanks. Praise be to intel!

Oh. So the burden of proof for your argument is on us. Well, that makes perfect sense.

That has to be most eyeroll worthy statement i've seen here.

 

[sniffin]

Oh. So the burden of proof for your argument is on us. Well, that makes perfect sense.

That has to be most eyeroll worthy statement i've seen here.

Except that he asked for proof because AMD are "unreliable" regarding TDP, an assertion I've seen mentioned here many times but without any form of hard evidence. Random arguments without factual basis is the trend, I'm just trying to fit in

edit: The burden of proof falls to him actually. The clocks have been basically verified as above, if he believes the 35W TDP is unrealistic and AMD have lied, that is up to him to prove not me.

 

[AtenRa]

Both are ~$50 desktop chips, they will compete against each other no matter how hard you try to pretend they wont.
Power comsumption does matter when we are talking about power hogs like Vishera that draw up to 160W more power than Haswell and still deliver worse performance. However, Haswell Celeron vs Kabini Sempron/Athlon power comsumption difference is completely insignificant by comparison (especially ''T'' 35W TDP models), total costs at the end of the month will be similar. You also fail to realize that Celerons are overall faster chips so they should finish any task faster than the competition, making the gap even smaller at the end of the day.

So then, Intel ATOM BayTrail based Pentium J2850 and J2900 10W TDP with a list price of $94.00 will directly compete against Haswell Pentium G3430 53W TDP at $93.00.

Nice, according to you guys the Baytrail ATOM will be smoked to kingdom come by Haswell Based Pentiums and Celerons. And because now no one cares about power consumption in Desktops, ATOM Pentium Baytrail products are doom to fail.
Why buy the J2900 when you can get the Haswell Pentium at the same price and have higher performance with a "few more" watts in power consumption ???

Now that we have cleared that, we can continue talking about AMD Desktop Roadmap.

 

[sniffin]

Interesting how that graph from hardware.fr pretty much confirms that AMD are reliable with TDPs too. Similar wattage to a 3930K (130W TDP), especially seeing as AM3+ is a old and inefficient platform (990FX NB alone pulls ~19W).

The 9590 pulls ~90-95W more than the 8350, which is exactly what you'd expect a 220W part to pull over a 125W part.

So I'd love to hear about how you came to the conclusion that they're unreliable re power mrmt. It would seem to me as though AMD are pretty strict about abiding by their published numbers.

 

[Enigmoid]

Why are you using synthetic benchmarks that place 100% load on every thread as examples for me being wrong? You've proved my point. 2.8GHz in synthetics makes 3.2/3.5GHz in user workloads like gaming pretty plausible. Unless people are buying mobile parts for Cinebench, encoding is probably the only time somebody wold ever experience base clocks

2.8 ghz in cinebench (which isn't synthetic as the rendering software runs exactly the same- Maxon cinema 3D) with 0 IGP load. They also noted that 3.5 ghz was rarely seen, 3.2 ghz was the effective singlethread turbo.

In games when you load up the igp, expect similar clockspeeds.

And this really isn't a high powered chip. 100% load on this chip is like 33% load on an i5-4570k.

Needless to say, under heavy load 2.8 ghz is the maximum speed.

 

[monstercameron]

so a laptop specced chips isnt as fast as a desktop specced chip?

also a10-7850k listing on amazon
http://www.amazon.com/AMD-AD785KXBJA...7639045&sr=1-2

 

[sniffin]

2.8 ghz in cinebench (which isn't synthetic as the rendering software runs exactly the same- Maxon cinema 3D) with 0 IGP load. They also noted that 3.5 ghz was rarely seen, 3.2 ghz was the effective singlethread turbo.

In games when you load up the igp, expect similar clockspeeds.

And this really isn't a high powered chip. 100% load on this chip is like 33% load on an i5-4570k.

Needless to say, under heavy load 2.8 ghz is the maximum speed.

3.2GHz is what you'd expect from dual threaded turbo. Heavy load across 4 cores -> 2.8GHz. Absolute 100% load -> 2.5GHz.

Playing games like PoE, Skyrim etc, it sits at 3.2GHz constantly. This will obviously vary depending on the notebooks cooling capabilities, but mine has no trouble sitting at 3.2GHz. BF4 is an example of a game that will pull it down to 2.8GHz.

I don't understand the comparison to the i5. How is that relevant

 

[Sweepr]

Nice, according to you guys the Baytrail ATOM will be smoked to kingdom come by Haswell Based Pentiums and Celerons

Performance-wise they certainly will.

And because now no one cares about power consumption in Desktops, ATOM Pentium Baytrail products are doom to fail.

Some people might priorize the lowest possible power consumption (there's lots of cheap 10W Bay Trail-Ds for those people). What I said (and you purposedly distorted) is: many desktop users would be ok with a few watts difference as long as there is a noticeable performance jump (Haswell Celeron is >2x as fast as 10W Bay Trail-D and 25W Kabini in 1-2 core CPU tasks).

Why buy the J2900 when you can get the Haswell Pentium at the same price and have higher performance with a "few more" watts in power consumption ???

Good question. Personally I'd take a 35W Haswell-based Celeron ''T'' over Bay Trail-D/Kabini any day. Bay Trail fits the mobile market perfectly (from 8'' fanless iPad mini-like Android/Windows tablets to 2-in-1 convertibles and small notebooks), but it's hard to keep up with <$50 Haswell-based CPUs in the desktop space.

 

[Enigmoid]

I don't understand the comparison to the i5. How is that relevant

Simply to show that it is much easier to put a heavy (relatively speaking) load on the chip because it is not very powerful.

 

[USER8000]

This was posted on the semi-accurate forums:

http://insidehpc.com/2013/12/22/hsas-clusters/

It seems APUs with Excavator cores will appear in 2015,have DDR4 compatability and will be SOCs.

I cannot see that happening on 28NM,so it probably is a 20NM part.

 

[inf64]

Looks like Excavator will support both DDR3 and DDR4. I hope they don't change their mind at the last second regarding the socket compatibility-this would suck for those buying Kaveri in hope they will get another CPU gen. out of their motherboards. But if that happens at least Excavator will be better APU since AMD desperately needs DDR4 or stacked memory to make any improvements in iGPU performance. Without memory BW improvements they will be wasting die space adding more stream processors on-die...

 

[ShintaiDK]

DDR4 support might be certain SKUs only, laptop or server for example. Specially since they list it as both FM2+, SoC and BGA. Or DDR4 could simply be an optional support they havent decided on yet. I doubt you see chips that supports both. Since DDR4 is not even pincompatible with DDR3.

It however seals the deal that Excavator is 2M/4T as well. And no HBM yet either.

 

[mikk]

DDR4 doesn't help AMD. Kaveri comes with DDR3-2400 and DDR4 isn't much faster 2015. Bandwidth bottleneck probably will be bigger than today.


DDR4 for mobile makes sense to save some power, also in mobile only DDR3-1600 is feasible.

 

[inf64]

It however seals the deal that Excavator is 2M/4T as well. And no HBM yet either.

That's not such a bad thing if you take into account how much faster it will be than Piledriver core. SR will make some rather solid IPC improvements, somewhere spectacular,somewhere rather small-it all averages out in the end. If it happens that Excavator in turn will greatly improve on SR (2x SIMD throughput for example is a massive improvement), then 2M/4T Excavator parts will probably beat ~FX8350 level of performance- greatly beating its ST performance and probably match or exceed its MT performance.

When you look at how FX8xxx stacks against Haswell you can see it has much lower ST performance and comparable MT performance- both these areas are targeted for improvement with both SR and even more so with Excavator.
Also, if Excavator comes on 20nm node, 3 module parts will most probably be launched in its lifetime. The modules will be really small. PD module on 32nm is ~31mm^2 with 2MB of L2; SR module @ 28nm will be even smaller; Excavator module @ hypothetical 20nm node will probably take up 50% less die area than PD on 32nm, even if it has some serious uarchitectural additions/changes.

 

[ShintaiDK]

That's not such a bad thing if you take into account how much faster it will be than Piledriver core. SR will make some rather solid IPC improvements, somewhere spectacular,somewhere rather small-it all averages out in the end. If it happens that Excavator in turn will greatly improve on SR (2x SIMD throughput for example is a massive improvement), then 2M/4T Excavator parts will probably beat ~FX8350 level of performance- greatly beating its ST performance and probably match or exceed its MT performance.

When you look at how FX8xxx stacks against Haswell you can see it has much lower ST performance and comparable MT performance- both these areas are targeted for improvement with both SR and even more so with Excavator.
Also, if Excavator comes on 20nm node, 3 module parts will most probably be launched in its lifetime. The modules will be really small. PD module on 32nm is ~31mm^2 with 2MB of L2; SR module @ 28nm will be even smaller; Excavator module @ hypothetical 20nm node will probably take up 50% less die area than PD on 32nm, even if it has some serious uarchitectural additions/changes.

Kaveri still needs to deliver before we even talk Excavator performance. And 3M/6T is as likely as it was with Kaveri. Remember max TDP for Excavator chips will be 65W

 

[mrmt]

That's not such a bad thing if you take into account how much faster it will be than Piledriver core. SR will make some rather solid IPC improvements, somewhere spectacular,somewhere rather small-it all averages out in the end. If it happens that Excavator in turn will greatly improve on SR (2x SIMD throughput for example is a massive improvement), then 2M/4T Excavator parts will probably beat ~FX8350 level of performance- greatly beating its ST performance and probably match or exceed its MT performance.

That on specific workloads or in every single workload out there? Because you are projecting a 100% IPC increase from Piledriver to Excavator, which is rather unlikely.

 

[ShintaiDK]

Haswell is for example 80% faster than Ivy Bridge in "specific workloads".

 

[daxzy]

Any reason why AMD won't release a Jaguar/GCN combo for consumers? They've already completed most of the work with X-One and PS4.

 

[Shivansps]

They will, those are the FS1B based APU, but dont even think for a moment they will be any good for gaming.