Slashdot Mirror
AMD Confirms Kaveri APU Is a 512-GPU Core Integrated Processor
MojoKid writes "At APU13 today, AMD announced a full suite of new products and development tools as part of its push to improve HSA development. One of the most significant announcements to come out the sessions today-- albeit in a tacit, indirect fashion, is that Kaveri is going to pack a full 512 GPU cores. There's not much new to see on the CPU side of things — like Richland/Trinity, Steamroller is a pair of CPU modules with two cores per module. AMD also isn't talking about clock speeds yet, but the estimated 862 GFLOPS that the company is claiming for Kaveri points to GPU clock speeds between 700 — 800MHz. With 512 cores, Kaveri picks up a 33% boost over its predecessors, but memory bandwidth will be essential for the GPU to reach peak performance. For performance, AMD showed Kaveri up against the Intel 4770K running a low-end GeForce GT 630. In the intro scene to BF4's single-player campaign (1920x1080, Medium Details), the AMD Kaveri system (with no discrete GPU) consistently pushed frame rates in the 28-40 FPS range. The Intel system, in contrast, couldn't manage 15 FPS. Performance on that system was solidly in the 12-14 FPS range — meaning AMD is pulling 2x the frame rate, if not more."
nVidia has at least three versions of the GT630, each fairly different from one another. None of these would be an amazing accomplishment to beat, although they are more powerful than Intel's normal integrated offerings.
I read the internet for the articles.
...how much faster does it mine Bitcoins?
I need to mine some so I can put them in a totally safe online wallet.
AMD has had those for a long time ever since they started their APU family post ATI acquisition.
What market is amd shooting for?
Haswell with iris pro will probably beat out amd for integrated graphics performance and will have better battery life.
On the top end, desktop users will always go for a dedicated graphics card.
On the mobile end, these things will eat up battery and have no reason to be on a tablet.
All that's left is the cheap oem side of things. Haswell is still fairly expensive on the low end. If intel can bring down the price a bit and make it competitive they will beat out amd in every category.
Yes, if you spend more money you can get more performance. The whole point of the APU is that you can spend less on a single piece of silicon than you would for "a better CPU and a decent graphics card."
Yes. One part with a middle of the road CPU and a middle of the road GPU.
The one advantage I see technically to this approach is you can get data from the CPU to the GPU without having to touch a trace on the motherboard. The over all complexity of the system goes down and the CPU to GPU performance goes up.
The disadvantages are many. More heat/power dissipation on the one part means it will run hotter (not that AMD doesn't do that anyway). Makes you pay for the GPU, even if you don't use/want it. Higher latency between the memory and the GPU which is KEY to a GPU performance. I'm sure there's more..
All this aside. Bully for AMD. These are great devices for low cost systems with reasonable performance.
Full Disclosure: I have a current low end AMD/GPU based system that I really like. It was CHEAP, and performs well enough for what I do.
"File to fit, pound to insert, paint to match" - Aircraft Maintenance 101
We need a CPU/GPU SoC based on the tech that's going in to the xbone and ps4. They both have multicore procs with a built in GPU that's capable of pushing next gen games at HD resolution.
We need that. A cheap PC built on a powerful single-chip solution. Not this wimpy shit.
Personally, I'd go for the the PS4 solution. 8 gigs of high speed GDDR5 that's both the main memory and graphics memory? Fuck yes. Give me that. I'd forgo the ability to expand memory if I could have GDDR5 as main memory. (The DDR3+128meg edram solution in the xbone is cheaper and clearly inferior)
So if I buy an AMD CPU, I can play games with low frame-rates at low detail settings (yeah, I know it says 'medium', but when almost all games now go at least up to 'ultra', 'medium' is the new 'low').
Or I could just buy a better CPU and a decent graphics card and play them properly.
Yes, but could you do that in a compact HTPC cabinet (breadbox sized or smaller) and have your total system draw less than 100W or so?
I'm really excited by this news - because it allows traditional desktops to reinvent themselves.
Think Steam Machines, think HTPC that lets you do full HD and 4k in the future, think HTPC that also lets you do light-weight or mid-level gaming.
Think of a replacement to consoles - a computing device that gives you 90% of the convenience of a dedicated console, but gives you full freedom to download and play from the app store of your choice (Steam or anything else), gives you better control of your hardware, and lets you mix and match controllers (Steam Controller, keyboard and mouse, or something else that someone invents a year down the line).
I'm long on AMD for this reason. Maybe I'm a sucker. But there is a chance that desktops can find a place in your living room instead of your basement. And I'm quite excited about that.
The best benefit I find is the need for only a single heatsink. One giant radiator and a single large, slow case fan, and you have an extremely quiet system.
The problem is that we want what the consoles have. A LOT more cores, GDDR5 and hUMA.
Good-bye
The real deal here is you are purchasing an APU for roughly less than half the price when compared to a mid-range Intel & discrete graphics solution, and getting double the performance. Apple knows this is the way to go for price-performance and that is why the new entry-level 15' MBPs lost discrete GPU. OEMs like Apple are forcing Intel to catch up with integrated GPU technology. It's all about trade-offs: you place a less performing GPU in the same die as the CPU in order to get the best possible memory interface. While you won't reach enthusiast or prossumer performance levels without adding a high end GPU, you will definitely target the common user market.
AMD should at least try using Intel Iris Pro which is their highest end GPU. The 630 GT is a ok low end GPU depending on which version they use.
You can call the advantage "complexity", but in practice that really means price, heat, and size, all of which are critical to laptops. Additionally, putting them on the same die makes it easier to have unified memory, which can further simplify things (and be as fast or faster in some applications for the less money if designed correctly - for example, compute tasks that touch a lot of the same data on the CPU and GPU like video encoding, etc).
And it most definitely does not have to "run hotter" than *two* discrete parts (and is certainly easier to cool, anyway). Computers are *always* using a GPU these days, modern OSes do all sorts of 3D effects (even some mobile ones). If the GPU (and software/driver) is designed well, it would be a lot simpler, cheaper, and possibly even more power efficient than the dual-graphics design of Macbook Pros and some Wintel laptops...
For a desktop, this isn't anything all that exciting (except for those who want cheap PCs with reasonable performance). For a laptop/embedded system, it's a really interesting chip, even if it's not the cheapest.
Actually the clock speed for the 862GFLOPS figure is in the footnotes, see here: http://images.anandtech.com/doci/7507/amd_kaveri_specs-100068009-orig.png
So, even unintentionally, they are talking about clock speeds...
Violence is the last refuge of the incompetent. Polar Scope Align for iOS
How is "price/performance tradeoff" a hard concept to understand?
Because if you care about games, this is too slow. If you don't care about games, this is irrelevant. I can't see any tradeoff that makes any sense outside of tiny niche markets (e.g. people who want to play games but don't care if they look like crap).
The best benefit I find is the need for only a single heatsink. One giant radiator and a single large, slow case fan, and you have an extremely quiet system.
I have an i7 and GTX 660 in my gaming PC, and it's an extremely quiet system with a heck of a lot more performance than this thing.
it's fast enough to play though.
world was created 5 seconds before this post as it is.
Why the needlessly stringent power draw? You can get passively cooled discrete GPUs or low-noise active cooling which would give you a major bump in performance. APUs won't be able to do 4K for a loooong time for anything but video.
What's wrong with 28-40 FPS on BF4 at 1920x1080? That's a brand-new game with high end graphics.
These machines share the memory between CPUs and GPUs, and that's the advantage:
You can use the GPU cores to do otherwise forbiddingly expensive operations (such as detailed
physics, particle simulations, etc) very easily. Traditional systems need to copy data between vram and main memory over the bus system bus, which takes time.
Programming languages are already starting to support mixed CPU/GPU programming with through new language constructs. At the moment, mainly rendering and physics is done on the GPU, soon it will be easy to do anything that can be efficiently parallelized.
It also probably costs as much for the CPU and GPU as it would for the entire AMD-based system.
So I _COULD_ buy a car for under $20k that does 0-60 in a modest amount of time...
Or I could just buy a Bugatti Veyron with a better engine and drive properly.
Is that the argument you're making?
Because if you care about games, this is too slow.
Its really not, unless its not the game you care about but the eye-candy. IIRC professional Starcraft 2 gamers (who could be said to "care about games") turn the graphics way down anyways.
It also probably costs as much for the CPU and GPU as it would for the entire AMD-based system.
The highest end AMD APU you can currently build includes an A10-6800K, which is a whopping $140 for the CPU+GPU. Include the cost of RAM for the GPU so that it can be comparable with a discrete GPU setup...$22 to compensate for dedicating 2GB of DDR3 1866 to the GPU...
$140 + $22 = $166.
His GTX 660 is $190. His i7 is no less than $290 based on todays newegg prices.
$190 + $290 = $480
So he is $314 in the hole. Clearly he doesnt want to talk about semantics such as cost.
"His name was James Damore."
APUs won't be able to do 4K for a loooong time for anything but video.
The main bottleneck for on-die GPU's is memory bandwidth. Intel "solved" the bandwidth problem in Iris Pro by including a massive L4 cache that cannot be manufactured cheaply. AMD hasn't solved the problem, but is far enough ahead in GPU design that the Iris Pro is only on par with AMD's 6800K APU.
"His name was James Damore."
Won't happen. Integrated devices like "smart" TVs and dedicated media streaming hardware have already obsoleted HTPCs, and as much as I like to play some PC games on a big screen, the market is clearly satisfied with the walled gardens known as game consoles, most of which can serve as media streaming clients as well.
https://www.eff.org/https-everywhere
Brand new AMD APU with 512 GPU cores beats Discrete NVidia card with 128 cores that's more than a year old now.
Hang on, was this supposed to be impressive?
APUs won't be able to do 4K for a loooong time for anything but video.
..if a "looooong time" means as soon as AMD and Intel support DDR4, which is in 2014... sure.
I think by "anything but video" he was referring to gaming, even the 780 Ti and R9 290X struggle with 4K. What do you think DDR4 would change? As far as I know they already support the 4K resolution but it'll play like a slide show.
Live today, because you never know what tomorrow brings
Actually this is a bigger deal than you think. I remember when you had to pay extra to get a floating point processor. Most software worked really hard to use integers when ever they could since they could not depend on an FPU being in most systems.
By having a GPU as part of the CPU more software will start to use GPU computing to speed up things like transcoding and even spreadsheets http://slashdot.org/story/13/07/03/165252/libreoffice-calc-set-to-get-gpu-powered-boost-from-amd.
We all know that GPUs can speed up a lot of operations but developers don't want to put in the work because not everyone has them.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
Higher latency between the memory and the GPU which is KEY to a GPU performance.
Latency doesn't matter very much for GPU. A little, but not much. Bandwidth is what matters for GPU. (Latency matters for CPU tasks.) As long as the GPU has enough buffer depth to cover the latency to and from memory (which it certainly does), the memory bandwidth is what will keep the GPU pipelines completely full.
Makes you pay for the GPU, even if you don't use/want it.
This is like saying the disadvantages of apples is that they don't taste like oranges.
It also probably costs as much for the CPU and GPU as it would for the entire AMD-based system.
If you'd read the thread, you'd see that the GP didn't mention cost at all, only noise.
What's wrong with 28-40 FPS on BF4 at 1920x1080? That's a brand-new game with high end graphics.
It's not 'high-end graphics' when you're playing on a low graphics setting.
Turn it up to Ultra and see what it runs at.
Why the needlessly stringent power draw? You can get passively cooled discrete GPUs or low-noise active cooling which would give you a major bump in performance. APUs won't be able to do 4K for a loooong time for anything but video.
You make a valid point - and I don't know *all* the options that exist.
It would actually be a very interesting exercise to do this kind of a comparison. Say, take some HTPC like constraints such as space and heat, identify the options available - both CPU+discrete graphics and CPU+GPU integrated, and compare the options using price and performance.
Back to your point, it is not just power draw - space and cooling are also factors. A reasonably strong integrated CPU+GPU system lets you build a cabinet that can be very slim - say, something that resembles a compact blue ray player.
I would also imagine that an integrated solution like this will allow better airflow.
Finally there's price. Undoubtedly discrete graphics will always have the performance crown. However, if you think of Moore's law, CPUs have already reached the point of diminishing returns in terms of size of individual cores or even number of cores in a chip. From now on, IMHO, Moore's law will be all about integrating as much of the motherboard as possible into a single chip or package. And GPU is the most obvious starting point.
To put it another way, in terms of price-performance-heat, discrete GPUs will not be able to compete with a highly integrated solution - over time. They will keep getting pushed into smaller and smaller niches. An integrated solution will generally be cheaper and cooler for equivalent performance. It wasn't a viable solution in many cases until now only because the performance was sub-par - but Kaveri is the first viable chip that gives you enough horsepower to play last gen games at full HD with reasonable frame rates. In two years, Kaveri will be at 2 teraflops - the same as a PS4.
Hey, anyone one who can't read a thread before jerking their knee.
That, and the fact that programming for GPUs now is analogous to having to choose between different processor- and language-specific floating-point libraries.
"[Regarding the 'cloud,'] ownership was what made America different than Russia." -- Woz
I agree that it's about through put and my "latency" term was a poor word choice. You got to keep the GPU's busy, which means you have to get data into and out of memory as quickly as possible.
"File to fit, pound to insert, paint to match" - Aircraft Maintenance 101
Clearly you didn't bother to read the thread, or you'd have noticed that it wasn't about cost.
The person I replied to specifically talked about cost, and in fact its the only thing they talked about.
Maybe YOU should read the thread.
"His name was James Damore."
And it most definitely does not have to "run hotter" than *two* discrete parts
No but it concentrates all the heat into the space of one of those discrete parts, so the cooler needs to be more efficient with diffusing heat from the tiny little surface.
If expensive solutions count, why not a PC version of the PS4 board? :-)
8Gbyte would be enough for an average gaming PC, and with DDR5 the bandwidth would be decent too
C - the footgun of programming languages
even the 780 Ti and R9 290X struggle with 4K.
30+ FPS in Far Cry 3 with Ultra settings at 4K resolution doesnt sound to me like "struggling"
49 FPS Grid 2 with Ultra settings at 4K resolution doesnt sound to me like "struggling"
45+ FPS in Just Cause 2 with Very High settings at 4K resolution doesnt sound like "struggling"
30+ FPS in Total War: Rome II with HQ settings at 4K resolution doesnt sound like "struggling"
If you arent right about existing hardware.. how could you possibly be a good judge of future hardware? Are you unaware that people are already doing 1080P gaming on their APU? My guess is that you were thinking that APUs "struggle" with 480P or some shit....
"His name was James Damore."
Bassed on the requirements I'm guessing this is for HTPC purposes.
In that case the 'needlessly stringent' power draw is because
A. The case is probably tiny, it may not even have space for a discrete GPU. Less power input = less heat to dissipate.
B. For a completely fanless solution you want a picoPSU. These max out at around 160watts.
C. Most people looking for quiet HTPC could care less if you can run Gears of Warcraft 5 on it.
Makes you pay for the GPU, even if you don't use/want it.
This is like saying the disadvantages of apples is that they don't taste like oranges.
Also, you can get GPU-less versions if that's what you want (socket FM2 Athlon X2 3xx and Athlon X4 7xx for the current generation, guess they'll make similar Steamrolled-based ones sooner or later)
Well hopefully OpenCL can solve this.
See my blog http://ilovecookes.blogspot.com/ for light hearted technical information.
As SOC become more capable, less need for dedicated GPU. Lower cost and smaller devices will drive this. If Kaveri plays BF4, it could take a big chunk of video card sales. AMD is going to catch up with Intel offloading parallel work to the GPU. Nvidia GPU will die out since they do not have a CPU to pair with it. Would have been good for Intel to by Nvidia. I don't think Intel will now that Nvidia is making Tegra ARM chips. Tegra is uphill battle against qualcomm and dozen other ARM companies.
Games don't come with just two settings: Ultra and Shit. There have always been a large number of people who will pay to play games not on the lowest setting, but don't want to pay way more for the diminishing returns of going from the second highest to the highest settings. And while there have always been some snobs trying to show off their e-penis or otherwise want to make themselves feel better because they spent more on their computer, it seems to be a recent trend for those people to insist others are must be running the game on lowest setting. What do you think you will gain by doing so? I can clearly see in the settings menu that I've set games to high or better, am I supposed to just assume I'm schizophrenic and have been hallucinating the game settings menu?
If expensive solutions count, why not a PC version of the PS4 board?
Maybe because nobody is selling such a thing?
The deal here is that memory controllers are now integrated into CPU''s (for performance reasons), so you can really only effectively use memory that the CPU was specifically designed for. Even if the mobo provided some emulation so that you could drop gddr5 into it instead of ddr3, your throughput would still be limited to 64 times the cpu's base memory clock, which is already attainable with ddr3.
"His name was James Damore."
Yeah, that's the point! It's easier to make a cooling system to diffuse heat from one die than two...
My HTPC has a passively cooled Geforce 210. Works like a charm, quiet enough that the wife hasn't thrown it out of the living room. Sure, it's slower than the Intel HD 4000, but it cost me $30 when the old GPU died.
for i in `facebook friends "=bday" 2>/dev/null | cut -d " " -f 3-`; do facebook wallpost $i "Happy birthday!"; done