Wednesday, May 8th 2013
AMD's Answer to GeForce GTX 700 Series: Volcanic Islands
GPU buyers can breathe a huge sigh of relief that AMD isn't fixated with next-generation game consoles, and that its late-2013 launch of its next GPU generation is with good reason. The company is building a new GPU micro-architecture from the ground up. Codenamed "Volcanic Islands," with members codenamed after famous islands along the Pacific Ring of Fire, the new GPU family sees AMD rearranging component-hierarchy within the GPU, in a big way.
Over the past three GPU generations that used VLIW5, VLIW4, and Graphics CoreNext SIMD architectures, the component hierarchy was essentially untouched. According to an early block-diagram of one of the GPUs in the series, codenamed "Hawaii," AMD will designate parallel and serial computing units. Serial cores based on either of the two architectures AMD is licensed to use (x86 and ARM), could handle part of the graphics processing load. The stream processors of today make up the GPU's parallel processing machinery.
We can't make out text in the rather blurry block-diagram, but are rather convinced that if it's authentic, then AMD is making some big changes. Another reason for AMD's delay could be silicon fab process. "Tahiti" as implemented on Radeon HD 7970 GHz Edition, already poses high thermal envelope. AMD doesn't want the 28 nm process to restrict its next-generation architecture development, and is holding out till the 20 nm process is in place at TSMC. The fab set Q4 as its tentative bulk manufacturing date for the process.
The source that leaked the block-diagram also posted specifications of the chip that's codenamed "Hawaii," which appears to be the flagship part.
Source:
ChipHell
Over the past three GPU generations that used VLIW5, VLIW4, and Graphics CoreNext SIMD architectures, the component hierarchy was essentially untouched. According to an early block-diagram of one of the GPUs in the series, codenamed "Hawaii," AMD will designate parallel and serial computing units. Serial cores based on either of the two architectures AMD is licensed to use (x86 and ARM), could handle part of the graphics processing load. The stream processors of today make up the GPU's parallel processing machinery.
We can't make out text in the rather blurry block-diagram, but are rather convinced that if it's authentic, then AMD is making some big changes. Another reason for AMD's delay could be silicon fab process. "Tahiti" as implemented on Radeon HD 7970 GHz Edition, already poses high thermal envelope. AMD doesn't want the 28 nm process to restrict its next-generation architecture development, and is holding out till the 20 nm process is in place at TSMC. The fab set Q4 as its tentative bulk manufacturing date for the process.
The source that leaked the block-diagram also posted specifications of the chip that's codenamed "Hawaii," which appears to be the flagship part.
- 20 nm silicon fab process
- 4096 stream processors
- 16 serial processor cores
- 4 geometry engines
- 256 TMUs
- 64 ROPs
- 512-bit GDDR5 memory interface
145 Comments on AMD's Answer to GeForce GTX 700 Series: Volcanic Islands
This could have huge ramifications, such as, people modding linux/windows to run on these chips solely, completely bypassing the current schism in memory hierarchy, and much much more.
On a more current note: I wonder how this complete redesign will influence driver development.
With AMD going this route, it sounds like they're going the same way as Nvidia, who is also rumoured to include ARM cores in one of their next major revisions. After reading this, I've got the feeling that AMD will get to market first with their version, like they traditionally get to market first on a new proces node.
Edit:
I'm wondering what part of the graphics workload is actually better off being performed on serial cores? perhaps a part of the pipeline currently handled by the CPU via drivers?
www.pcper.com/reviews/Graphics-Cards/Frame-Rating-Dissected-Full-Details-Capture-based-Graphics-Performance-Testin
4096 stream processors??? That's double than a 7970 GPU
512-bit GDDR5??? Prepare for some records on the bandwidth
I think AMD is preparing a monster here, and I wont be off to say that this GPU will almost be as powerful as a 7990 card.
A new 5870 on the horizon?? (compared to 4870X2 that is...)
64-bit ARM cores (A57) are the new black. It's all about decreasing dependance/latency associated with modern systems. You can call it an APU, but it's really just enhancing the capabilities of GPGPU beyond a GPU being a dumb processor being thrown parallel workloads by a CPU. It needs it's own brain(s) to go with the brawn. These advancements plus shared memory = huge improvements.
Wonder if they will scale this (in the sense that 1 cpu block would go with 4 ROPs and 4 Compute Units)...I bet they do...because that would make a ton of sense (and be pretty efficient I reckon). Be it one ARM block + 1 said gpu block, or 2 (say x86 blocks + 2 gpu blocks...which sounds an awful lot like the supposed specs of Kaveri).
Also, I think we all kind of figured we'd see a 512-bit bus considering bandwidth is a seriously huge (and limiting) factor, even now. Ofc we all plan on seeing GDDR6 at some point, but no word of production has been spoken about afaik.
Figuring 4096 cores and 512-bit/7gbps on the current architecture..that's around 975mhz-ish, 8000mhz, closer to 1100mhz. That is not accounting for cpu cores and perhaps very likely a more robust cache structure though. Obviously with the cpu cores, at the bare minimum cache will need to be larger/faster to feed them...I'm just throwing it out there assuming they are supplemented in conjunction with one another...or more-or-less bolted on to the current gpu spec.
Sounds plausible as while the the 1.9x density improvements are possible and touted 1.3x speed improvements at the same voltage may be approximately correct (and hence a chip this spec lines up with being the size to accompany a fast 512-bit bus), 20nm will likely target a lower nominal voltage. IE, where 28nm may have been .85v/850mhz but in reality running 1.05-1.175 (and up to 1.3v) with yields settling in-between 850-1175, this will probably be closer to .85-1.05v in reality (up to 1.175) with yields starting around 1000-1100mhz. Given the typical ~10% clock cushion per voltage...it lines up.
So...in essence...AMD is doing the exact same thing as nvidia is with Maxwell/Denver...only (conceivably) earlier and in a more flexible manner.
hmmm...wtf?!??
www.amd.com/us/products/desktop/graphics/8000/pages/8000-series.aspx#2
Edit:
Or AMD is talking about the 9000 series??????????????????????????????????????????????
AMD: Please deliver a new architecture that can alleviate the micro stuttering, at a price point this side of reasonable. The 7970 was good, but Nvidea kinda ate your lunch with the 6xx series by having less issues on the driver side.
Nvidea: A rebranding of 6xx cards to the 7xx series cards might fool some people. The 6xx series is nice, but the pricing is still unreasonable. Look away from the Titan, and deliver a 7xx series without massive mark-ups and minimal performance increases.
All GPU producers: Resist the urge to rush to market. We will wait four months until you can secure reasonably priced chips, and a decent step forward on features. Just make sure whatever you release can actually be utilized. Two or three games make my current 6950 cry at moderate (1920x1080) resolution, I'm not paying more than $250 for an improvement in the two of those games I actually play.
My opinion on the architecture change is simple. Include x86 style cores for the express purpose of doing physics processing without using CPU resources. An octet of slow speed cores doesn't sound like much, but they could do everything that Physx on a CPU does. AMD offloading their shiny new Tressfx to a separate on-board processor means better performance while being transparent to the user. If the APU is adding GPU to a CPU then this move must be the opposite approach to the same goal, a decent SoC