AMD Carrizo APU Details Leaked

[mrmt]

And this is a highly technical manual dug up from the depths of their website, not a "marketing slide".

Seems that I was wrong. I confused Compute Core with Compute Unit:

http://www.pcper.com/news/Processors/AMD-CES-2014-Presentation-Kaveri-Goes-Official

 

[Xpage]

Well the statement "next year [insert some random Intel CPU] will finish off what's left of AMD's "big core"." has been said for years, and so far it has not come true.

Of course the same thing has been said about AMD CPUs going to finally match intel CPUs.


Always gets my hopes up before they are dashed with benchmarks. I hope socketed kabini will be nice and cheap.

 

[erunion]

Well the statement "next year [insert some random Intel CPU] will finish off what's left of AMD's "big core"." has been said for years, and so far it has not come true.

The FX line sure seems dead.

 

[NostaSeronx]

The FX line sure seems dead.

If there is no news for FX by June 12, 2014. Then, you and others can consider the FX brand dead for this year.

---
I think within the 40h-4Fh families we have a rather high chance of getting an FX.

"Warsaw" / "Centurion" replacing "40h-4Fh with Steamroller" for both the "Server" / "Enthusiast" markets. Warsaw being the stop gap for the servers and Centurion being the stop gap for enthusiasts.

I think we forum goers will definitely see an unnoticed skip to Excavator with the next FX.

Very vague information of this new FX;
E3 2014

Possible paper launch of this new FX;
October to November

Potential retail availability of this new FX;
December to January

 

[Fjodor2001]

Would anyone care to estimate how an 8 core Excavator based FX CPU would measure up against an Intel 4770K in multi-threaded workloads?

 

[AtenRa]

Would anyone care to estimate how an 8 core Excavator based FX CPU would measure up against an Intel 4770K in multi-threaded workloads?

Win some, loose on others. Just like Piledriver vs Sandy/Ivy.

edit: It would win in the majority of Multitasking/Throughput though.

 

[Enigmoid]

Would anyone care to estimate how an 8 core Excavator based FX CPU would measure up against an Intel 4770K in multi-threaded workloads?

It would need to be about 15%+ to be equivalent in MT workloads.

 

[Fjodor2001]

Ok, so in this next-gen-console era of gaming where multi-threaded games are becoming more common, an 8 core Excavator based FX CPU would likely be a better buy than a 4770K (or the Broadwell replacement for that). Especially if we're considering the price aspect too.

Then I find it hard to see why AMD would not release such an FX CPU in 2015.

 

[NostaSeronx]

Would anyone care to estimate how an 8 core Excavator based FX CPU would measure up against an Intel 4770K in multi-threaded workloads?

If there was a theoretical 8-core Excavator based FX CPU.

It would have;
8-cores Disabled
8 MB of L2, 16 MB of L3.
32(GFX +12? GPP) possible lanes of PCIe 3.
If placed on a theoretical AM4 socket, it would have dual-channel / If on the same platform as the 16-core part, it would have quad-channel.
Support for memory unification to discrete GPUs.

P.S. I know you asked for MT. I just wanted to add that 02h(Abu Dhabi/Warsaw/Vishera/Centurion) to 40h-4Fh(unknown) will have a huge platform improvement gap.

 

[norseamd]

Ok, so in this next-gen-console era of gaming where multi-threaded games are becoming more common, an 8 core Excavator based FX CPU would likely be a better buy than a 4770K (or the Broadwell replacement for that). Especially if we're considering the price aspect too. Then I find it hard to see why AMD would not release such an FX CPU in 2015.

they might be making a non module based processor

 

[norseamd]

If placed on a theoretical AM4 socket, it would have dual-channel / If on the same platform as the 16-core part, it would have quad-channel. Support for memory unification to discrete GPUs.

why do they still make seperate enthusiast lines? why not just merge the enthusiast and business lines and save some money. they could just bin the cpus for more clock speed and lower the core count if they need to

 

[Phynaz]

Would anyone care to estimate how an 8 core Excavator based FX CPU would measure up against an Intel 4770K in multi-threaded workloads?

It wouldn't be an estimate, it would be a wild-assed guess.

 

[Homeles]

It wouldn't be an estimate, it would be an wild-assed guess.

Yep. I tried to guesstimate Steamroller's performance and failed miserably. I was off by a good 100%+ on IPC. The architecture is just so broken that what would normally be a huge improvement results in a headfirst slam into another bottleneck. At this point, the issue is either with the stupid high cache and memory access latencies, or the whole thing is unsalvageable.

 

[NostaSeronx]

- Steamroller -

Improvements;
Memory Access and Bandwidth (L1, L2, RAM)
Decode/Dispatch Rate
More Stable Load/Store

Regression;
4 macro-op instruction dispatch to 3 macro-op instruction dispatch
4 macro-op retire to 3 macro-op retire
No improvements for the fetch unit to match up with dispatch/decode.
Pipe 0 and Pipe 2 of the FPU are now one pipe.

-
Fixed the memory rate but ruined the instruction rate. It is more efficient but it is also equally slower than predecessors.

 

[erunion]

Ok, so in this next-gen-console era of gaming where multi-threaded games are becoming more common, an 8 core Excavator based FX CPU would likely be a better buy than a 4770K (or the Broadwell replacement for that). Especially if we're considering the price aspect too.

Then I find it hard to see why AMD would not release such an FX CPU in 2015.

Negative.

The number of threads on the next gen consoles are a red herring. The CPU side of their APUs are weak, an i5 is much more powerful. Creating extra threads won't change that.

You won't need an i7 to play next gen ports, more like an i3.

 

[NostaSeronx]

The CPU side of their APUs are weak, an i5 is much more powerful.

You can't compare Temash and Kabini to the processors in the Playstation 4 and Xbox One.

Temash and Kabini => One 64-bit Memory Bus.
Playstation 4 and Xbox One => Four 64-bit Memory Buses.

Remove the bottleneck caused by slow memory from Temash and Kabini. You'll have a processor that is faster than; (Clock to Clock)
Physical Quad-core Ivy Bridge. (4C/4T)
Virtual Quad-core Haswell. (2C/4T)

The six cores available for gaming is vastly faster than what was provided by the; Xenon and Cell processor.

If the improvement between each generation of the cat family is maintained. It would not be surprising if the Playstation 5 and Xbox Two would be using Leopard derived cores. It would also not be surprising if the mainstream performance (Trinity/Richland/Kaveri/Carrizo/Dozer family) gets replaced as well.

The cat family from AMD is the most competitive architecture for x86.

 

[erunion]

You can't compare Temash and Kabini to the processors in the Playstation 4 and Xbox One.

Temash and Kabini => One 64-bit Memory Bus.
Playstation 4 and Xbox One => Four 64-bit Memory Buses.

Remove the bottleneck caused by slow memory from Temash and Kabini. You'll have a processor that is faster than; (Clock to Clock)
Physical Quad-core Ivy Bridge. (4C/4T)
Virtual Quad-core Haswell. (2C/4T)
.

Kabini doesn't have a CPU memory bottleneck.

Even IF they did match Core clock for clock, their clocks speeds of 1.6 and 1.75 GHz, put them at half the performance. I assume the actual deficit is greater.

 

[Enigmoid]

You can't compare Temash and Kabini to the processors in the Playstation 4 and Xbox One.

Temash and Kabini => One 64-bit Memory Bus.
Playstation 4 and Xbox One => Four 64-bit Memory Buses.

Remove the bottleneck caused by slow memory from Temash and Kabini. You'll have a processor that is faster than; (Clock to Clock)
Physical Quad-core Ivy Bridge. (4C/4T)
Virtual Quad-core Haswell. (2C/4T)

The six cores available for gaming is vastly faster than what was provided by the; Xenon and Cell processor.

If the improvement between each generation of the cat family is maintained. It would not be surprising if the Playstation 5 and Xbox Two would be using Leopard derived cores. It would also not be surprising if the mainstream performance (Trinity/Richland/Kaveri/Carrizo/Dozer family) gets replaced as well.

The cat family from AMD is the most competitive architecture for x86.

Except for the god awful ST performance.

Its about equal to a P4 in many cases.

Kabini CPU performance is not RAM limited.

 

[NTMBK]

Seems that I was wrong. I confused Compute Core with Compute Unit:

http://www.pcper.com/news/Processors/AMD-CES-2014-Presentation-Kaveri-Goes-Official

Hah, yeah, that one is pretty confusing. Especially when

1 GPU compute core = 1 GPU compute unit
1 CPU compute core = 1/2 CPU compute unit

:S :S :S

I guess "compute unit" is just "minimum reproducible unit"?

 

[monstercameron]

Except for the god awful ST performance.

Its about equal to a P4 in many cases.

Kabini CPU performance is not RAM limited.

so many things wrong with this statement.

1. it is more efficient
2. it uses much lower power
3. it has more perf per clock
4. much smaller die size

 

[mrmt]

Hah, yeah, that one is pretty confusing. Especially when

1 GPU compute core = 1 GPU compute unit
1 CPU compute core = 1/2 CPU compute unit

:S :S :S

I guess "compute unit" is just "minimum reproducible unit"?

Seems so, but what's a compute core then, and, changing the question a bit, what can we expect when we read that AMD is selling a 8 CC unit? As much as there might be a technical argument to terminology, I don't think marketing is doing a good job when bringing that to the marketing slides. Seems more like an arbitrary convention to give APU some big numbers to show in marketing slides than anything that might give the consumer any clue on what to buy.

 

[NTMBK]

Seems so, but what's a compute core then, and, changing the question a bit, what can we expect when we read that AMD is selling a 8 CC unit? As much as there might be a technical argument to terminology, I don't think marketing is doing a good job when bringing that to the marketing slides. Seems more like an arbitrary convention to give APU some big numbers to show in marketing slides than anything that might give the consumer any clue on what to buy.

The compute core seems to be the minimum independently controllable unit- one core on the CPU, or one whole shader cluster on the GPU. So you could in theory have 12 different HSA tasks being operated on simultaneously, 4 on the CPU and 8 on the GPU. I am actually glad that they went with a vaguely sensible definition of "core" on the GPU- I'm sick of marketing claiming that each individual SIMD lane on a GPU is a core. Tegra K1 having "192 cores" is a bad example of that.

But yeah, "compute core" in more of a marketing term. "Compute unit" makes sense in terms of repeatable blocks in silicon, and I suspect comes from the engineers.

 

[el etro]

1) Its about equal to a P4 in many cases.

2) Kabini CPU performance is not RAM limited.

1) No. Test cinebench to get it. Kabini performance/thread is close to Trinty performance/thread.

2) DC could help more the performance of Kabinis(Kabini is single-channel)...

 

[Enigmoid]

so many things wrong with this statement.

1. it is more efficient
2. it uses much lower power
3. it has more perf per clock
4. much smaller die size

None of which have anything to do with ACTUAL REAL WORLD performance.

I don't care about the efficiency,die size, etc. My statement has everything to do with performance.

1) No. Test cinebench to get it. Kabini performance/thread is close to Trinty performance/thread.

2) DC could help more the performance of Kabinis(Kabini is single-channel)...

Kabini does well but its generally at half the clockspeed. Having half the BW is generally not a problem.

DC won't help anything other than maybe igp.

 

[Insert_Nickname]

At this point, the issue is either with the stupid high cache and memory access latencies, or the whole thing is unsalvageable.

I'm inclined to believe that. AMD's memory controller is a lot less efficient then Intel's. Case in point, same 2133MHz dual channel memory kit yields ~32GB/s read/write on an i7-3770, and only ~21GB/s read and 10.5GB/s write on a 6800K...

AMD's DDR3 controller could definitely do with an update.