Tuesday, February 24th 2015

AMD's Excavator Core is Leaner, Faster, Greener

AMD gave us a technical preview of its next-generation "Carrizo" APU, which is perhaps the company's biggest design leap since "Trinity." Built on the 28 nm silicon fab process, this chip offers big energy-efficiency gains over the current-generation "Kaveri" silicon, thanks to some major under-the-hood changes.

The biggest of these is the "Excavator" CPU module. 23 percent smaller in area than "Steamroller," (same 28 nm process), Excavator features a new high-density library design, which reduces die-area of the module. Most components are compacted. The floating-point scheduler is 38% smaller, fused multiply-accumulate (FMAC) units compacted by 35%, and instruction-cache controller compacted by another 35%. The "Carrizo" silicon itself uses GPU-optimized high-density metal stack, which helps with the compaction. Each "Excavator" module features two x86-64 CPU cores, which are structured much in the same way as AMD's previous three CPU core generations.
The compaction in components doesn't necessarily translate into lower transistor-counts over "Kaveri." In fact, Carrizo features 3.1 billion transistors (Haswell-D has 1.4 billion). Other bare-metal energy optimizations include an 18% leakage reduction over previous generation, using faster RVT components, which enables 10% higher clock speeds at the same power-draw (as "Kaveri"). A new adaptive-voltage algorithm reduces CPU power draw by 19%, and iGPU power draw by 10%. AMD introduced a few new low-power states optimized for mobile devices such as >9-inch tablets and ultra-compact notebooks, which reduces overall package power draw to less than 1.5W when idling, and a little over 10W when active. In all, AMD is promising a conservative 5% IPC uplift for Excavator over Steamroller, but at a staggering 40% less power, and 23% less die-area.
The integrated-GPU is newer than the one featured on "Kaveri." It features 8 compute units (512 stream processors) based on Graphics CoreNext 1.3 architecture, with Mantle and DirectX 12 API support, and H.265 hardware-acceleration, with more than 3.5 times the video transcoding performance increase over "Kaveri." For notebook and tablet users, AMD is promising "double-digit percentage" improvements in battery life.

Now, if only AMD can put six of these leaner, faster, and greener "Excavator" modules onto an AM3+ chip.
Add your own comment

85 Comments on AMD's Excavator Core is Leaner, Faster, Greener

#51
Aquinus
Resident Wat-man
cheesy999That's because the ddr2 controller was on the motherboard and the ddr3 one was on the cpu, they could presumably do something similar again though...
Wrong, The Phenom II and Athlon AM3 chips had both a DDR2 and DDR3 controller on the CPU and you could put AM3 CPUs in most AM2+ boards with the proper BIOS. AM3+ CPUs had the DDR2 portion of the IMC removed, so as a result AM3+ was incompatible with AM2+ whereas AM2+ CPUs only had a DDR2 controller which is why you couldn't put an AM2+ CPU in an AM3 board with DDR3, it was not because the controller was on the motherboard. AMD hasn't had a memory controller on the motherboard for many, many, many years. Even going back to skt939 with DDR, the memory controller was on the CPU. There is absolutely no memory controller logic on the motherboard, just IO with PCI-E and all the peripherals on the south bridge.
Posted on Reply
#52
arbiter
The cpu space needed for both controllers, and given AMD financial state. Its just not worth it to do it. How many people go out and just replace their MB and keep the same cpu (not counting MB failure)? The time, effort, money, and what would be mostly no boost in performance less you use APU graphic's but might well use that money and get a dedicated gpu instead. Cost of the board + new ram it would be better bang for $ to just buy the current gpu in the price range of what you would spend.
Posted on Reply
#53
sergionography
I don't know what to think of this :/ it's great what amd is doing with this process node but to think they have only increased density yet retained the same die size? Same number of cores? And same number of graphics clusters? Meaning all that real estate is going towards hsa stuff. Now the question remains, is it really worth it AMD? Is this hsa really worth it for you to use up 1/3 of the chip towards it. AMD needs to stop trying to save the world with their "innovations" that will pave the way of the future and instead learn how to make money, and not just with this but everything they been doing so far. Look at mantle for example, yes it's great and yes it forced other apis in the right direction, but did it make amd any money or better their position in the marketplace? Well the market share loss this year doesn't think so unfortunately, making it a fail business project. And similar thing with hsa, it didnt benefit amd one bit because it's been half baked baby steps so far and meaningless because there is no point of an excellent fiber if the main components are lacking. AMD would've been better off building simpler apus with either more cpu/gpu to gain more performance, or with less of this hsa fabric to get smaller chips and therefore more profits, and only release hsa when it is completed and tested at its best form and when they have good components to make it shine
Posted on Reply
#54
sergionography
arbiterThe cpu space needed for both controllers, and given AMD financial state. Its just not worth it to do it. How many people go out and just replace their MB and keep the same cpu (not counting MB failure)? The time, effort, money, and what would be mostly no boost in performance less you use APU graphic's but might well use that money and get a dedicated gpu instead. Cost of the board + new ram it would be better bang for $ to just buy the current gpu in the price range of what you would spend.
i agree with you. But then is using more than 1/3 of the chip on hsa fabric worth it? At a time when amds x86 cores are forced to compete only in the mid-low end market I don't think hsa makes sence. Amd should've waited till they can compete in the high end and began releasing hsa from top to buttom on chips with high margins first, that's the only place where new untested secondary technologies make sense.
Posted on Reply
#55
Aquinus
Resident Wat-man
sergionographyi agree with you. But then is using more than 1/3 of the chip on hsa fabric worth it? At a time when amds x86 cores are forced to compete only in the mid-low end market I don't think hsa makes sence. Amd should've waited till they can compete in the high end and began releasing hsa from top to buttom on chips with high margins first, that's the only place where new untested secondary technologies make sense.
HSA makes sense for APUs because the general purpose X86 cores and the specialized GPU shaders will need to share a memory controller and as a result need a common interface into it. HSA makes this problem a whole lot easier to deal with because it is that piece that sits between the CPU and GPU cores and the memory controller. The motivation is efficient use of local resources (aka cache and the IMC) without wasting too much time trying to coordinate the two. So it is a side-effect that cores talking to each other is faster but the real focus was about ensuring that both components can take maximum advantage of the memory controller without trampling over each other. Without it the different parts of the APU will be wasting time trying to compete for shared resources.

Now leveraging HSA to actually use shared memory between both X86 and GPU shaders is an entirely different problem which has more to deal with software design as applications must be able to take advantage of it if the workload is conducive to such improvements.

All in all, I think AMD made a mistake with SMT by using dedicated shared components. It was a huge waste of time for the regular market. If AMD invested these resources into their monolithic X86 core design with AM2+ and AM3, improving the IMC and cpu cache latencies would have done more for performance than SMT ever could have for 95% of situations that most consumers actually care about but instead of trying to scale vertically, they decided to scale horizontally and the result has been, in my opinion, lackluster. I just don't see how AMD is going to get out of this hole they've gotten themselves into.
Posted on Reply
#56
xfia
you make some good points for sure but hsa is way more efficient for a lot of mobile devices and like what if the game consoles had these architectural improvements today.. probably be quite a bit nicer and I think they will have something great for console gamers next gen. should help with lower cost gaming laptops too but that is something they will have to sell hard to be widely realized.
Posted on Reply
#57
newtekie1
Semi-Retired Folder
arbiterThe cpu space needed for both controllers, and given AMD financial state. Its just not worth it to do it. How many people go out and just replace their MB and keep the same cpu (not counting MB failure)? The time, effort, money, and what would be mostly no boost in performance less you use APU graphic's but might well use that money and get a dedicated gpu instead. Cost of the board + new ram it would be better bang for $ to just buy the current gpu in the price range of what you would spend.
That is why on the first generation of CPUs you only have the DDR3 memory controller. Get something out in the high end market to start generating some money. AM4 CPUs would be DDR3 only, AM4+ would have DDR3 and DDR4 so AM4 users could upgrade their CPU without upgrading their motherboard.
Posted on Reply
#58
BiggieShady
AquinusIf AMD invested these resources into their monolithic X86 core design with AM2+ and AM3, improving the IMC and cpu cache latencies would have done more for performance than SMT ever could have for 95% of situations that most consumers actually care about but instead of trying to scale vertically, they decided to scale horizontally and the result has been, in my opinion, lackluster. I just don't see how AMD is going to get out of this hole they've gotten themselves into.
After splurging insane amount of cash on ATI acquisition, it seems they were saving on CPU R&D because scaling out horizontally is easier/cheaper. Given the lackluster results and the fact they are in huge debt now, they really should have taken a huge loan back then and put it all in R&D - at least they would have a chance to get out of the debt today. I'm not sure console chips manufacturing will be enough, because I have a feeling they offered very attractive price to ms and sony to seal the deal.
Posted on Reply
#59
HumanSmoke
sergionographyI don't know what to think of this :/ it's great what amd is doing with this process node but to think they have only increased density yet retained the same die size? Same number of cores? And same number of graphics clusters? Meaning all that real estate is going towards hsa stuff.
Yet, Kaveri already has integrated HSA...so it kind of begs the question, how different are they?
Posted on Reply
#60
ThE_MaD_ShOt
newtekie1That is why on the first generation of CPUs you only have the DDR3 memory controller. Get something out in the high end market to start generating some money. AM4 CPUs would be DDR3 only, AM4+ would have DDR3 and DDR4 so AM4 users could upgrade their CPU without upgrading their motherboard.
This is exactly my thoughts, get something out now then worry about ddr4 when it becomes mainstream. Amd has always been a tad late to the party when adopting new memory standards so why should this time be any different. Release something now using ddr3 then there next release can run either ddr3 or 4.
Posted on Reply
#62
HumanSmoke
_FlareHas someone seen this before ?
FP4 Socket with DDR4
Probably a development board for the Excavator-cored Toronto (server orientated BGA), which is specced for DDR4

Posted on Reply
#63
anolesoul
As "slow" as AMD has been to PUT anything out for "competition" against Intel, AND to actually support DDR4...I'm NOT holding my breath!
Posted on Reply
#64
HumanSmoke
anolesoulAs "slow" as AMD has been to PUT anything out for "competition" against Intel, AND to actually support DDR4...I'm NOT holding my breath!
Probably wise. You'll note that the above slide states that AMD's ARM-cored Seattle is a 2014 product, yet wont be launched until at least Q2 2015. Given the fabrication time for an SoC , you would have to think that the A1100 would already be in production but that might not be the case.
Posted on Reply
#65
RealNeil
I have enough CPU power to do me for the foreseeable future. (just got a FX-9590 CPU and board that I won, also have a i7-4790K to install)
So I can wait to see what they release, and read reviews on it when it's time.
I really hope that they nail it with something.
I hope that their next series of GPUs kick ass too!

I'll wait and see.
Posted on Reply
#66
arbiter
anolesoulAs "slow" as AMD has been to PUT anything out for "competition" against Intel, AND to actually support DDR4...I'm NOT holding my breath!
DDR4 will help the amd APU's gpu side but that is it. Doubt AMD will go ddr4 least for now as prices are still high and not worth the cost for performance it would give. Might well put the extra $ in to a dedicated gpu.
Posted on Reply
#67
RealNeil
I've been told that APUs respond well to faster RAM.
Posted on Reply
#68
newtekie1
Semi-Retired Folder
RealNeilI've been told that APUs respond well to faster RAM.
Only really on the GPU side of it. Think of using faster system RAM as akin to overclocking the memory on a graphics card.
Posted on Reply
#69
refillable
These slides are almost always very interesting but that doesn't mean it will always as interesting as the processors/benchmarks themselves. Let's all just wait.
Posted on Reply
#70
hairyfeet
Jorge...and as documented by the past three Intel node reductions, no tangible performance is gained with die shrinks below ~32Nm. The die shrink is primarily for lower power consumption, increased transistor density and lower unit costs. That's why AMD is in no hurry to rush to smaller node sizes.

AMD is doing a very good job on their APU designs. Carrizo should make a lot of people very happy. AMD is scheduled to officially release these chippies in Q2 of '15. They are actually in production now.

For those not up top speed on die shrink sizes, there is almost zero performance gain now that we are below 32Nm. The biggest gains in die shrink are lower power and higher transistor density, which reduce CPU/APU costs. With AMD's advanced power management on Carrizo, power consumption is reduced even further than in prior low power APU models.

Carrizo APUs will be available in both mobile and desktop versions with mobile models being released first.
Except they still have a VERY large amount of sales of 6 and 8 core AM3+ budget gamers and NONE of them are gonna downgrade to a 4 core APU?

My 6 year old Phenom II X6 paired with an HD7790 is gonna kill those 4 core APUs, badly. And if I were building new, as my youngest recently did, you can get an FX8300 for $115 and pair it with an HD7790 or R260 and SLAUGHTER the most expensive APU in their line up thanks to more cores and GDDR 5.

So if my choice is a 4 core APU from AMD or a 4 core CPU from Intel, why would I buy AMD? After comparing doing A/V work and gaming with 6 cores I'll go 8, even 10, but I won't downgrade to just 4, not when you can get 4 WITH HT from Intel, it just wouldn't make any sense.
Posted on Reply
#71
xfia
where is that 16 compute core apu with a strong hold on 1080p
Posted on Reply
#72
hairyfeet
xfiawhere is that 16 compute core apu with a strong hold on 1080p
You should probably make it clear you want 16 actual COMPUTE cores, because I think AMD is being really disingenuous with their adverts saying "12 cores" when IRL its a lousy 4 CPU cores and 8 MUCH weaker and more specialized (and also memory constrained) GPU cores. I'm sorry but a GPU unit is NOT a core, I can NOT just run any program I want on it like I can a CPU and until then? I'd argue their ads are seriously deceptive.

Of course this is why my youngest never even thought about their APUs, as he already HAD 4 Phenom II cores and had just gotten an HD7790 so there would have been absolutely no point in any chip in their APU lineup...and that is the problem. Still plenty of tasks where more REAL CPU cores can be a benefit, but if you have even the lowest end GDDR 5 card (which is currently a whopping $60) then nothing in their APU lineup makes ANY sense at all.

This is why I'm seriously hoping they either come out with new AM3+ chips or come out with 8-12 real CPU core AM4s soon, because right now I have a sinking feeling they are just gonna throw in the towel WRT all the budget gamers out there and strictly build APU granny boxes and if that is the case? Stick a fork, they are as good as dead as its all the hexa and octo AM3+ guys that are keeping the fanbase alive, recommending their chips to friends, even with no new chips in this long there still isn't anything on the Intel side that can compete with 8 real cores for multithreading until you spend over double the price. But again if the ONLY choices we have is a 4 core APU from AMD, stuck with a GPU that will never EVER get used if you have even a $60 card, versus a 4 core from Intel with higher single core performance AND the option of getting a 4 core with HT down the line when the price drops..why would I stick with AMD when my hexacore gets too long in the tooth?
Posted on Reply
#73
xfia
I meant 8 cpu and 8 gpu.. they wont release anything new for am3+ but 8 excavator threads (4 modules) on fm2+ would be nice to see.

the whole compute core thing makes since to me but yeah misunderstood by a lot of people as well as hsa for the most part.

it may seem strange that something more advanced and efficient than other desktop platforms started out on the lower end with no new chipset but kaveri was the first generation. apu's by design have the cpu and gpu on the same chip so there was nothing high end to start without further research and refinements. kaveri is pretty good at 900p so I would expect these new apu's to be knocking on the 1080p door. due to the way that technology advances a separate cpu and gpu will remain a better performing option but after another generation or two people will be saying 1080p gaming is a job for one those amd apu's so you can save money if you want.

amd's got plenty of granny boxes coming! zen.. excavator.. carrizo.. playstation.. xbox.. wii..
Posted on Reply
#74
arbiter
hairyfeetYou should probably make it clear you want 16 actual COMPUTE cores, because I think AMD is being really disingenuous with their adverts saying "12 cores" when IRL its a lousy 4 CPU cores and 8 MUCH weaker and more specialized (and also memory constrained) GPU cores. I'm sorry but a GPU unit is NOT a core, I can NOT just run any program I want on it like I can a CPU and until then? I'd argue their ads are seriously deceptive.
Well GPU cores in software that can be accelerated by them runs much faster but that is not much software to start with. Most software that people would use that is GPU accelerated in such a way usually more pro end software and likely those people will spend a lot on hardware to begin with. Look at the slides AMD put out claiming their apu match's an i7 in mobile side, All those benchmarks the compare with are all GPU accelerated ones which AMD has big edge on.
Posted on Reply
#75
hairyfeet
arbiterWell GPU cores in software that can be accelerated by them runs much faster but that is not much software to start with. Most software that people would use that is GPU accelerated in such a way usually more pro end software and likely those people will spend a lot on hardware to begin with. Look at the slides AMD put out claiming their apu match's an i7 in mobile side, All those benchmarks the compare with are all GPU accelerated ones which AMD has big edge on.
But you just pointed out the POINTLESSNESS of the whole thing as those running the kind of high end AV and specialty software like 3D CAD that can actually USE those GPU cores are NOT gonna be running AMD APUs for such CPU intensive tasks, not when they can get a MUCH more powerful GPU for MUCH cheaper and pair it with more real CPU cores!

The simple fact is IMHO the entire APU concept makes ZERO sense with the exception of mobile. In a laptop where space is a premium and power is severely limited? Then sure having the CPU and GPU on one die cuts down the costs and power usage, but on a desktop? Even if you get the lowest end APU you are still getting ripped off, I mean look at the prices, you can get a dual core APU with an HD8300 for $69 OR you can go to some place like Biiz and pick up an FX4300 with four REAL cores that will do any task (not just the extremely limited GPU accelerated ones) for the same money. By the time you figure the increased cost of the APU over the CPU, the need for much faster RAM compared to the CPU as the GPU side of an APU is ALWAYS starved for memory bandwidth? you will simply never come out ahead as even the lowest end GPU with dedicated GDDR 5 memory (which as I said is just $60, the GeForce 610 or 710 IIRC) with just slaughter the thing without effort!

Look I'm about as hardcore an AMD supporter as they come, I have 6 AMD PCs in my family going back to my father's Phenom I quad all the way up to the FX8300 of the youngest, but if you are not running a laptop? There just isn't a selling point for their APUs, there just isn't. If you build a machine with the least expensive APU and I build the least expensive CPU+GPU the increased cost of the APU is gonna make it a losing proposition,just compare the lowest ACTUAL quad core APU with the same on the AM3+ side and its not even funny how lopsided it is, you can get 4 REAL cores AND a GPU for less than the APU quad so no matter how you slice it? It just doesn't make sense.
Posted on Reply
Add your own comment
Jul 4th, 2024 04:18 EDT change timezone

New Forum Posts

Popular Reviews

Controversial News Posts