The Official Kaveri Review Thread (A10-7850K, etc)

[venturizhou]

Is there any speculation about the release of non k version of the A10? A8-7600 is suppose to be near end of Q1 right? or sooner?

 

[oceanside]

Sold. Read the last review over at phoronix and will be retiring my old P4 for a Kaveri on my Slackware machine. Motherboard, processor and ram are all on the way.

 

[Fjodor2001]

yeah, they havent mentioned a peep about anything lower than 45W tdp. Maybe this will have the surprise that we have been waiting fir...maybe quad channel ddr3/4

I just found this:

http://www.anandtech.com/show/7645/amd-press-conference-highlights

"Kaveri is coming out on desktops next week with laptop versions to follow in Q2/Q3, as Ian discussed already."

So maybe that means we'll see 15/25W Kaveri mobile CPUs in Q2/Q3? Maybe that (mobile CPUs) is the primary target for Kaveri, and where it's intended to shine?

 

[Erenhardt]

Those bioshock results look really weird, for instance the low score of the gtx660. Also that you get the same performance for the 7750 and 7770. In any case, I am not seeing 7770 levels of performance from the configuration tested, in fact it is 20% lower than 7750. I think we need to wait for the proper drivers and see the scaling and stutter results with more cards. It does look promising however. OTOH, as I stated before, I probably would not buy a card to use for crossfire, but instead, a cheaper CPU and more powerful gpu. It could be a plus in mobile, and also in cheap pre-builts with a weak power supply.

Here is the thread with review that focused on my point. It shows better kaveri in DualGraphics competing against other dgpus:
http://forums.anandtech.com/showthread.php?t=2364337

It is still without dualgraphics driver and with 2133 RAM.

 

[frozentundra123456]

Here is the thread with review that focused on my point. It shows better kaveri in DualGraphics competing against other dgpus:
http://forums.anandtech.com/showthread.php?t=2364337

It is still without dualgraphics driver and with 2133 RAM.

Whatever, that is not the information you cited in your original post. In any case, at its current price, you could buy a cheap athlon X4 and a HD 7770 both for the price of the 7850k alone, so I still don't see the point.

 

[monstercameron]

Whatever, that is not the information you cited in your original post. In any case, at its current price, you could buy a cheap athlon X4 and a HD 7770 both for the price of the 7850k alone, so I still don't see the point.

to be fair at its current price[ read a10-7850k] does come with BF4 in most cases. prices will go down and it will be even more competitive but you are right, for gaming a cpu+dgpu combo will perform better.

 

[Erenhardt]

Whatever, that is not the information you cited in your original post. In any case, at its current price, you could buy a cheap athlon X4 and a HD 7770 both for the price of the 7850k alone, so I still don't see the point.

Just for the sake of argument, you know that 7770 requires 6pin power connector.

Athlon, while amazing value, is a bit slower than kaveri. I agree that 7850k is a bit overpriced, but the included bf4 makes the deal more appealing.

 

[frozentundra123456]

Just for the sake of argument, you know that 7770 requires 6pin power connector.

Athlon, while amazing value, is a bit slower than kaveri. I agree that 7850k is a bit overpriced, but the included bf4 makes the deal more appealing.

That is true for a prebuilt system. You could possibly get HD7770 levels of performance without changing the power supply. I would like to see those results confirmed with an evaluation of micro stutter however.

For someone building their own system, I don't really see the power supply as an issue.

 

[Torn Mind]

As long as there are enough Molex connectors, the 6 pin adapter is an insignificant issue as many 7770s do include a 6-pin PCI-e adapter in the box.

 

[cytg111]

So .. much like finding the Higgs Boson, the universe would have been so much cooler if we had not .. much as this place would be so much more exciting if kaveri didnt pan out just as expected. Is it a correct assesment that AMD needs hsa-huma AND mantle to take off to even get into competetive terretory? I dont see that rocket lifter anywhere.

 

[Dresdenboy]

So .. much like finding the Higgs Boson, the universe would have been so much cooler if we had not .. much as this place would be so much more exciting if kaveri didnt pan out just as expected. Is it a correct assesment that AMD needs hsa-huma AND mantle to take off to even get into competetive terretory? I dont see that rocket lifter anywhere.

Would the adaption rate be higher if Intel would have included an HSA like capability? How many applications and games supported AVX right after the launch of SB?

Mantle is also a way to circumvent the problems caused by an inefficient graphics API which have been solved by more and more CPU performance in the past. Mantle also helps Intel CPUs a lot.

 

[ShintaiDK]

Would the adaption rate be higher if Intel would have included an HSA like capability? How many applications and games supported AVX right after the launch of SB?

Mantle is also a way to circumvent the problems caused by an inefficient graphics API which have been solved by more and more CPU performance in the past. Mantle also helps Intel CPUs a lot.

There is a huge difference. AVX is universal supported. HSA and Mantle are both are vendor limited and not universal. Also Intel had full compiler support long before the products was released. Same with AVX2.

HSA is also simply another attempt at the STREAM, CUDA, OpenCL, etc case. And its the same limited amount and type of apps that benefit. So it doesnt matter if Intel supports it or not. It just never had a chance to be anything important. Even worse, HSA is now keeping AMD from using eDRAM/eSRAM solutions. Something they desperately need. And stacked memory is too far away still. Not to mention HSA is still nowhere to be seen. Not to mention the context switch is first supported in Carizzo.

Mantle is 100% vendor specific and locked to a minority player, both in dGPU and iGPU. There is no incentive to do Mantle without sponsor money. And we already see this. AMD should atleast have focused their effort in making a better DX driver that actually supported better scaling like nVidia. Its now 4 years without this. And their driver team leaves much to be desired. Fracturing it further with more support and complexity in terms of mantle is a bad thing. Mantle with a beta driver is also now...Q1.

For me it simply seems AMD is trying the shotgun approach. Blind attempts to make a killer product everyone needs. Problem is there is no such product and all the money is wasted. And AMD is running out of resources and time quickly. Kaveri is a flop CPU wise and IGP is evolutionary at best. The price increase however further destroys any hope of a performance/$ worthy product with regression there.

AMD needs to focus their R&D. And they needed to do that years ago. The FX line is dead, server segment is dead. Big core line is on life support. Small core line falls between 2 chairs. The ARM attempt is simply foolish. dGPU are the only part that isnt in direct disaster.

And do I even have to mention AMDs ability to underdeliver on all the overestimations?

 

[Arkaign]

AMD should really go in battle mode and bring them out at $49, $79, and $99. At current price points they are not going to make more than a tiny 'dink' in the pond.

It is sadly ironic, because the people who would most benefit from better APU performance (gamers) already have a plethora of budget options that perform better, so long as they're not stuck on an ITX build. The extreme SFF idea is a little goofy to me, it's cake to stuff a mATX build to the side of a desk out of sight, and you get a ton more flexibility and options.

The next wave of $100 dGPUs will make this even more dramatically unappealing to most. :\

 

[ShintaiDK]

AMD should really go in battle mode and bring them out at $49, $79, and $99. At current price points they are not going to make more than a tiny 'dink' in the pond.

It is sadly ironic, because the people who would most benefit from better APU performance (gamers) already have a plethora of budget options that perform better, so long as they're not stuck on an ITX build. The extreme SFF idea is a little goofy to me, it's cake to stuff a mATX build to the side of a desk out of sight, and you get a ton more flexibility and options.

The next wave of $100 dGPUs will make this even more dramatically unappealing to most. :\

When you say ITX, you mean the extreme small ITX

You can have a GTX780TI/290X in an ITX system :biggrin:

But as you say, its supposed to be a mainstream product, not a limited niche product.

 

[mrmt]

AMD should really go in battle mode and bring them out at $49, $79, and $99. At current price points they are not going to make more than a tiny 'dink' in the pond.

AMD cannot do that with a huge 245mm^2 die.

 

[sontin]

AMD cannot do that with a huge 245mm^2 die.

Sure they can.
Look at the next gen consoles. They sell such SoCs for the $99 or so.

Mantle is also a way to circumvent the problems caused by an inefficient graphics API which have been solved by more and more CPU performance in the past. Mantle also helps Intel CPUs a lot.

AMD doesn't support DX's Driver Command List. So they are one reason why many DX games are not better optimized for more draw calls.

 

[itsmydamnation]

AMD doesn't support DX's Driver Command List. So they are one reason why many DX games are not better optimized for more draw calls.

and lots of devs have said they are useless, but lets not let that get in the way of anything. find a dev saying mantle is useless.

 

[Lonbjerg]

and lots of devs have said they are useless, but lets not let that get in the way of anything. find a dev saying mantle is useless.

You mean DICE have said it "useless" right?

 

[NTMBK]

Even worse, HSA is now keeping AMD from using eDRAM/eSRAM solutions. Something they desperately need.

There's no reason why HSA would prevent eDRAM caches. If AMD were to copy Intel's approach and have a large eDRAM L4 cache, that would be entirely transparent to the HSA software level.

 

[ShintaiDK]

There's no reason why HSA would prevent eDRAM caches. If AMD were to copy Intel's approach and have a large eDRAM L4 cache, that would be entirely transparent to the HSA software level.

If we look at the Xbox. Then hUMA is not supported for the same reason.

 

[Gideon]

If we look at the Xbox. Then hUMA is not supported for the same reason.

I'm pretty sure it has more to do with the specific implementation, Microsoft wanted to enforce as well as time constraints to design the chip. I can't see anything in HSA or hUMA itself, that would stop it from using eDRAM, when the memory were treated as a L4 cache.

 

[mrmt]

Sure they can.
Look at the next gen consoles. They sell such SoCs for the $99 or so.

The business model is completely different.

With consoles, AMD is bound to a very stringent contract where their margins are meant to be low, they don't have to spend money on sales/marketing efforts, they don't have to support a R&D pipeline around the product because the follow-on, if build by them will be demanded by Sony or MSFT, which btw aren't really interested in making money with their consoles, just covering costs and the life cycle of the product is way big.

With consumer chips, they are not bound by a contract that ties down their margins, they have to maintain a R&D pipeline by themselves, the have a lot of sales/marketing efforts (which costs a lot of money), the life cycle of the product is far smaller than console chips, the replacements of their products is their own responsibility, and everyone on the supply chain is interested in making a lot of money from the product itself.

So while AMD might be able to sell console chips relatively cheap, they would not be able to reproduce this model with their consumer chips.

 

[mrmt]

If we look at the Xbox. Then hUMA is not supported for the same reason.

I can't remember where but I read that the problem with hUMA and esRAM for the XBO isn't the esRAM itself, but that the communication channels are only from the GPU. The CPU can't access the esRAM. So you could get unified memory addressing on other EDRAM implementations as long as you don't replicate the XBO self imposed limitations.

 

[Dresdenboy]

There is a huge difference. AVX is universal supported. HSA and Mantle are both are vendor limited and not universal. Also Intel had full compiler support long before the products was released. Same with AVX2.

HSA is a system architecture, kind of API+drivers, supporting existing compilers, but taking over the HSAIL code generation. HSA is not an ISA like AVX. Please excuse the TLAs.

HSA is also simply another attempt at the STREAM, CUDA, OpenCL, etc case. And its the same limited amount and type of apps that benefit. So it doesnt matter if Intel supports it or not. It just never had a chance to be anything important. Even worse, HSA is now keeping AMD from using eDRAM/eSRAM solutions. Something they desperately need. And stacked memory is too far away still. Not to mention HSA is still nowhere to be seen. Not to mention the context switch is first supported in Carizzo.

It's not simply another attempt, it's (to the programmer) a simpler attempt. It's actually even OpenCL agnostic. You can use C++ AMP, OpenMP etc.
The amount of apps that benefit increases, since one main problem of using OpenCL/CUDA etc. is the currently involved overehead, which of course zeroes out many use cases on embedded/mobile heterogeneous processors. We use OpenCL at work and know enough use cases.
AVX is an ISA to increase throughput in vectorizable and/or parallelizable code with little effort. OpenCL/CUDA etc. help in parallelizable code. This includes for example image processing, big data, search algorithms, particle filters, genetic programming, big neuronal networks, number crunching, databases, etc. This in combination with (much) lower overhead leaves a lot of use cases.
And think mobile/embedded: in such cases, you get more performance for the same power.

As others said, fast local memory is simply a cache. And there are enough caches on both the CPU and GPU parts. Any not that easily usable solution like the one in the XBox One just adds complexity.

Carizzo will reduce that overhead even further. The multitasking shouldn't be a problem yet - due to the amount of available HSA apps.

Mantle is 100% vendor specific and locked to a minority player, both in dGPU and iGPU. There is no incentive to do Mantle without sponsor money. And we already see this. AMD should atleast have focused their effort in making a better DX driver that actually supported better scaling like nVidia. Its now 4 years without this. And their driver team leaves much to be desired. Fracturing it further with more support and complexity in terms of mantle is a bad thing. Mantle with a beta driver is also now...Q1.

The minority in dGPU market is of other dimension than in iGPU market. And since Mantle also helps Intel CPUs with AMD dGPUs, the market for Mantle currently is 1/3 of the whole dGPU market. The API is another step of vendor specific game optimizations (not driver, but by direct game studio support), which already can make a difference of 10-20%.

And how similar are Mantle concepts to low level console APIs, allowing to use similar engine architectures?

 

[Dresdenboy]

Sure they can.
Look at the next gen consoles. They sell such SoCs for the $99 or so.

I was going to answer that, but it's already covered.

AMD doesn't support DX's Driver Command List. So they are one reason why many DX games are not better optimized for more draw calls.

Since when does a ~35% dGPU marketshare company prevent developers from optimizing their engine for the other ~65%?
Here's Nvidia's view:
https://developer.nvidia.com/sites/...dev/docs/GDC_2013_DUDASH_DeferredContexts.pdf

There is still a part of DX not influenced by AMD.