Info on Maxwell?

[FordFanboy]

I don't think Maxwell will really get interesting until it has a die shrink, they've done pretty much all they can squeezing out performance on 28nm.


I'd rather have HDMI over DVI, DVI was only a stopgap from VGA. Personally I'm using all three interfaces in this PC and I don't get the hatred for HDMI. HDMI 2.0 is going to be VERY useful! (UHD @60fps)

I don't think you really understand what a die shrink is. You're not going to see any huge gains from a die shrink.

 

[xoleras]

I don't think you really understand what a die shrink is. You're not going to see any huge gains from a die shrink.

Yep. This. The main effect of a "die shrink" is simply to reduce costs; it allows more chips per wafer. The real gains come from the higher transistor budget and increased efficiency which a more advanced process allows.

 

[Emmanon]

I've read "informed" rumors stating that Maxwell was designed from the ground up for power efficiency. Their target market is mobile computing. This makes sense, since demand for tablets, phones, etc. has exploded recently. Nvidia might be trying broaden its product portfolio.

 

[LordEC911]

Yep. This. The main effect of a "die shrink" is simply to reduce costs; it allows more chips per wafer. The real gains come from the higher transistor budget and increased efficiency which a more advanced process allows.

*not really directed at you xoleras, just expanding on your comments.*

Unfortunately there is a lag time with the cost advantage of jumping to a new node, wafer pricing for the first 12-24months usually means an increase in cost over previous nodes.
While increasing transistor density is always exciting in terms of performance, GPUs have been at a power wall the last few generations. A 20nm ASIC could see a near doubling in transistor density over 28nm but 20nm only offers a 15-25% savings in power consumption at the same frequency.
This is why we are seeing Nvidia tweaking their architecture and working to squeeze all the power efficiency they can out of their new GPU. Hopefully their efficiency gains translate well to 20nm.

 

[Wiseguy2001]

I don't think you really understand what a die shrink is. You're not going to see any huge gains from a die shrink.

More transistors, lower power and higher clocks. The biggest gains come from switching to a new process rather than refinements. I think I get the gist of things thanks!

 

[FordFanboy]

More transistors, lower power and higher clocks. The biggest gains come from switching to a new process rather than refinements. I think I get the gist of things thanks!

Actually, I don't think you get the gist of things. Switching to a new process is a refinement. The biggest gains actually come from new architectures, for example, Nehalem bringing in an integrated memory controller and reintroducing HT, etc.

 

[xoleras]

More transistors, lower power and higher clocks. The biggest gains come from switching to a new process rather than refinements. I think I get the gist of things thanks!

Maybe i'm completely misunderstanding you, but more transistors requires a new architecture. Plopping an existing architecture onto a new process does little to nothing - it does not magically add transistors. The transistor count remains exactly the same without a new architecture.

Again, the main reason any company has done this in the past was to save money, essentially. That has not been the case as of late because wafer prices with new processes (28nm, 20nm) has been exponentially more expensive; "die shrinks" have not been economically feasible for any company in recent years. So, generally speaking, new process = new architecture. Die shrinks, from what i've read, make little to no sense these days. In the past it DID make sense, but no longer. It's not economically feasible for anyone now (except Intel) and it does not provide much of a benefit these days.

 

[hcforde]

AMD's R9 was meant for the 22mm process but it was not ready so it is on the 28. When it is able to be on what it was designed for the heat factor will go down because it will be at the correct index it was designed for.

 

[LordEC911]

AMD's R9 was meant for the 22mm process but it was not ready so it is on the 28. When it is able to be on what it was designed for the heat factor will go down because it will be at the correct index it was designed for.

No. No it wasn't.

 

[ep0x73]

A crazy idea but Intel competes with AMD and Intel has the fab's so why not contract Intel to make your GPU's?
I can't see how this would break federal rules, maybe it does?

 

[LordEC911]

A crazy idea but Intel competes with AMD and Intel has the fab's so why not contract Intel to make your GPU's?
I can't see how this would break federal rules, maybe it does?

Intel doesn't have to play nice. If that was ever offered to Intel they would simply just price AMD/Nvidia out of the market. That was the big worry at the time when AMD was spinning off GF, how could AMD truly compete with ChipZilla without their own fabs and what happens if GF eventually goes under? TSMC is/was a powerhouse but wasn't at the same level as AMD's fab nor had the correct process. UMC hasn't been the same, Chartered got gobbled up, IBM/Samsung/ST were already started down the specialized road, IBM/Samsung/UMC are really the only candidates and don't seem likely at this point.

The other side of the matter is that Intel's fabs are specifically tuned/tweaked to their architectures and designs. As an almost pure in-house fab, they don't have the same experience and knowledge about catering their process to other/multiple design houses.

 

[Wiseguy2001]

Actually, I don't think you get the gist of things. Switching to a new process is a refinement. The biggest gains actually come from new architectures, for example, Nehalem bringing in an integrated memory controller and reintroducing HT and larger L3.

Thanks for pointing out one of the biggest reasons to add new transistors before your ninja edit. I agree, the more cache the better.

Maybe i'm completely misunderstanding you, but more transistors requires a new architecture. Plopping an existing architecture onto a new process does little to nothing - it does not magically add transistors. The transistor count remains exactly the same without a new architecture.

Again, the main reason any company has done this in the past was to save money, essentially. That has not been the case as of late because wafer prices with new processes (28nm, 20nm) has been exponentially more expensive; "die shrinks" have not been economically feasible for any company in recent years. So, generally speaking, new process = new architecture. Die shrinks, from what i've read, make little to no sense these days. In the past it DID make sense, but no longer. It's not economically feasible for anyone now (except Intel) and it does not provide much of a benefit these days.

Of course they utilise the extra silicon from moving to a new process, you're just arguing semantics now. Yes one reason is that they save money, another is they're restricted in how much power they can use by thermal spec's and what customers want.

Enough of this, let's focus on the thread topic, Maxwell.

 

[Emmanon]

In about two weeks, we get the first taste of Maxwell.
http://www.tomshardware.com/news/nvidia-gtx-750-ti-gm107,25974.html

 

[xorbe]

In about two weeks, we get the first taste of Maxwell.
http://www.tomshardware.com/news/nvidia-gtx-750-ti-gm107,25974.html

That's a repost of the info from post #12 (both source videocardz).

 

[Liger88]

No. No it wasn't.

If he's referring to Volcanic Islands, yes it actually was. Sea Islands, the Radeon HD 7000 Series refresh was expected to be 28nm, but it for some reason got canceled and they just called it Volcanic Islands. It was why a lot of screwy information was coming out leading up to the months of the R 200 cards coming out.

 

[Unknown-One]

You do realize that the standards body for DVI disbanded over a decade ago? I like DVI as well, but it isn't going to be around long term - I give it about 5 years or so before it disappears even for legacy reasons.

And I'm still not seeing your point...

DVI is the more capable interface:
- Dual-link DVI supports more bandwidth and higher resolutions than HDMI.
- Any DVI port can be adapted to HDMI.
- Adapted DVI ports support HDCP and audio, just like a physical HDMI port.
- Adapted DVI ports can be made to support all HDMI levels (including HDMI 2.0).
- DVI comes with no per-port license fee attached to it.

You gain NOTHING (in fact, you lose some capabilities) by putting physical HDMI ports on a graphics card instead of physical dual-link DVI ports.

Main thing is, HDMI 2.0 will be able to do 4k on a single cable, while DVI-D tops out at 2560*1600.

Wrong, wrong, wrong. The 2560x1600 limitation ONLY applies to Dual-Link DVI implementations using two 165 MHz clock sources. Modern clock sources support much faster rates than this.

Dual-Link DVI, according to the spec, has no set maximum at all. If you want to run higher resolutions over DVI, you simply need a faster source generator.

HDMI works under exactly the same principal. It's a single-link DVI port, and they keep attaching faster-and-faster source generators. It's just a crappier version of DVI.

HDMI 2.0 emulates dual-link DVI by running a single-link at (up to) 680 MHz. Dual-Link DVI supports two such clock generators, and nothing in the DVI spec says you can't attach dual 680MHz clock generators (which would allow double the bandwidth of HDMI 2.0)

That's the entire problem with DVI right now, it is not 4k capable - theoretically, 2 DVI cables *could* do it, but that is a theory. I have not seen a single screen using double DVI implementation, while nearly every 4k screen is using HDMI 1.4a with 2 cables or displayport with 1 cable.

You wont see a dual-cable DVI implementation because DVI doesn't need a dual-cable implementation

Dual-Link DVI has double the number of data lines as HDMI. Plugging in one dual-link DVI cable can offer the same (or more) bandwidth as plugging in two HDMI cables.

HDMI 2.0 can do 4k with 1 cable.. Heck, 4k wasn't even thought of when the last iteration of DVI was created, and it's a moot point now since the standards body is disbanded. Both HDMI 2.0 and displayport are able to do 4k on a single cable.

As said, Dual-Link DVI can do 4k with one cable as well. Not an issue. The DVI spec has built-in allowances for this (the same allowances HDMI is using to increase its maximum supported resolution).

HDMI is having issues running high resolutions because it's based on single-link DVI. This means HDMI needs a much faster source to achieve the same thing as dual-link DVI.

edit: nevermind. I suppose you're talking in the context of Maxwell, i'm 100% sure Maxwell cards will have DVI. DVI will disappear at *some* point, but that will be many years from now. I'm getting ahead of myself again - DVI will be around for some time to come simply for legacy reasons. Also, if we're talking HDMI 1.4a (which is honestly a joke) i'd use DVI-D over HDMI 1.4a any day. I can't even get 2560x1600 working on most HDMI 1.4 implementations.

And I reiterate, why would DVI go away when it's the superior interface?

Think of it this way:
- You can adapt any DVI port to HDMI. It's effectively the same thing as single-link DVI + audio. Even HDMI 2.0 signaling can be run through a DVI port if a faster source generator is included on the card.
- HDMI ports? You're stuck. You can't adapt that to dual-link DVI no matter what. There's simply no way for HDMI to outstrip dual-link DVI's bandwith, because the faster source generator applies to DVI as well (and dual-link DVI gets to use two of them instead of just the one that HDMI is limited to).

Also consider, HDMI uses exactly the same specs for its TMDS and VESA/DDC links as DVI. If you implement HDMI, you've already implemented (at the very least) single-link DVI.

 

[LordEC911]

If he's referring to Volcanic Islands, yes it actually was. Sea Islands, the Radeon HD 7000 Series refresh was expected to be 28nm, but it for some reason got canceled and they just called it Volcanic Islands. It was why a lot of screwy information was coming out leading up to the months of the R 200 cards coming out.

No, while they might have hoped TSMC would hit their original roadmap/timeline for 20nm, their previous experience with Ibiza made them much more cautious regarding node jumps.
AMD has been changing/moving/merging/adding/removing internal code names due to what happened over a year ago. Not much point trying to read into the codenames from the last year or two.
I can guarantee you though that Hawaii was always meant to be 28nm, there was never anything planned for 22nm. According to some they had taped out a couple 20nm ASICs a bit ago.

 

[Sunin]

So with a 75 watt design will we see the 880 having 4 gpus on a single board? Could the 750ti be 1/4 the power of the 880? If so that could get interesting. 10k+ fire strike extreme from a single card. Or will they just incorporate more transistors in the 880"? This is the first time they might could go for scale ability and offer a range of cards not limited by the amount of unlocked stream/cuda cores. So base model 1 processor scaling to 4 for the GTX. Thoughts? Wouldn't this be cheaper for the company overall. Mass produce one chip for scale of economy. Just seems like something they might do with the low wattage rating

 

[FordFanboy]

Thanks for pointing out one of the biggest reasons to add new transistors before your ninja edit. I agree, the more cache the better.



Of course they utilise the extra silicon from moving to a new process, you're just arguing semantics now. Yes one reason is that they save money, another is they're restricted in how much power they can use by thermal spec's and what customers want.

Enough of this, let's focus on the thread topic, Maxwell.

It's pretty clear you don't understand the engineering.

 

[LordEC911]

So with a 75 watt design will we see the 880 having 4 gpus on a single board? Could the 750ti be 1/4 the power of the 880? If so that could get interesting. 10k+ fire strike extreme from a single card. Or will they just incorporate more transistors in the 880"? This is the first time they might could go for scale ability and offer a range of cards not limited by the amount of unlocked stream/cuda cores. So base model 1 processor scaling to 4 for the GTX. Thoughts? Wouldn't this be cheaper for the company overall. Mass produce one chip for scale of economy. Just seems like something they might do with the low wattage rating

No, that has been talked about since AMD went with their "small" core strategy. It is simply too much work and too many cons to that type of implementation. They would all have to be on the same substrate, all have to have a ridiculous amount of traces/interfaces so that all the dies could communicate as needed while also lengthening all the traces that normal would be on a single die but now spread out unevenly across 4. They could save loads on manufacturing but the R&D cost and then the cost of actually implementing and packaging the cores like that would probably negate any sort of benefits.

Now after saying all that, an implementation of what you are talking about that is probably something we will see in the future would be die stacking. Once the semiconductor industry figures out how to effectively cool the stacked dies for normal consumers, then we could see simply making a single ASIC and scaling up. GPUs as we know them have been doing something similar the last few generations, placing most of the actual GPU parts in "building blocks" and then scaling up or down.

28nm at best we see around 3x the specifications of 750Ti, maybe a bit more at lower clocks, 20nm maybe around 4x.

 

[Sunin]

So what will be their Approach for 880 just make a larger chip then?

 

[ep0x73]

Intel doesn't have to play nice.

They don't but if they want to make $$ they would. Intel could easily pry Nvidia away from TSMC if they gave them a great offer to make chips on cutting edge fabs.
Intel has the resources to take on contracts if they desire.
TSMC could then concentrate on AMD's desires almost exclusively.