Who thinks Maxwell is getting a rebrand/rebadge?

[tviceman]

AMD has released it's new lineup of cards, and despite the various opinions about the current price points, performance levels, and power draw levels, new (or at least refreshed) products that are at least priced effectively vs. another company's products will drive higher sales. AMD is set to ride the 300 series until it's true next-generation parts are ready for production.

Unless Pascal comes in the first of 2016, I fully expect to see GTX 1000 series (or whatever nomenclature Nvidia decides upon) to be released before the end of this year. It would likely benefit Nvidia's bottom dollar in the long term to release "new" products rather than straight up price drops on it's existing lineup, PLUS a 2H release of Pascal would be pushing the GTX 900 series to 2+ years old (aka too long). I think a full fledged GM200 with 75-100mhz higher boost clock than Titan X will be the GTX 1080 TI @ $699, while the GM204 will get rebadged with higher boost clocks as the GTX 1070 @ $399-429 and so on and so forth.

 

[Subyman]

Moving this to general since the OP's topic is predicated on what AMD did. You'll probably want to be able to use AMD and its line up to discuss the issue.

Moderator Subyman

 

[BigDaveX]

They did re-spins for both Fermi and Kepler, wouldn't surprise me if they come up with another incremental improvement to hold the fort until Pascal arrives. At the very least, I'd expect them to release a 985 Ti with higher clockspeeds and a fully-enabled chip, just because the 980 Ti overclocks so well and AMD have the potential to eke some extra performance out of the Fury X with driver improvements.

 

[Qwertilot]

It was definitely very possible that they'd have done this - it'd probably have been the normal reaction if AMD had (for instance) launched Fury/X as the 390/X, nano as ~380 etc.

Not sure if they've really got the motivation to do it now to be honest. Depends a bit on their timescales for Pascal of course.

 

[Mako88]

Of course it will get a refresh, that's not rocket science.

Pascal is further away than most are assuming here, so there will absolutely be time to do a hugely clocked 14nm refreshed 980 prior to Pascal launching that offers +40-50% to a current 980 Ti. No doubt about it.

Nothing wrong with it at all, will drive prices down, which needs to happen badly.

 

[werepossum]

Anybody have any info on time frame? I'm trying to decide when to buy a new 4GB VRAM video card for Far Cry 4 and Fallout 4. Ideally I'd wait until just before the Fallout 4 launch, but I'm afraid that once stock is depleted on R9 290 and 290X and the GTX970 is competing only with the 390/390X the 970 prices may actually creep up. On the other hand, if NVidia releases a refresh then 970 prices will come down, and probably either way 390/390X prices will come down. Although I'm not at all satisfied that the 390 is faster in 1080p than a well-cooled factory overclocked 290X, so that might not be a winner if 390 prices settle down lower than current 390 prices but higher than current 290X prices.

 

[Head1985]

They dont need refresh.Why you think maxwell needs refresh?
we didnt hear single speculation about that so its not gonna happen.
Nv can easilly price GTX980 at 300USD if they want.Its same GPU as GTX970 only not cut-down and GTX970 sell for 300USD already.They will still make tons of money from 300USD gm204.
hawaii is bigger than GM204 and how much 290 cost?

 

[Mako88]

Anybody have any info on time frame? I'm trying to decide when to buy a new 4GB VRAM video card for Far Cry 4 and Fallout 4. Ideally I'd wait until just before the Fallout 4 launch, but I'm afraid that once stock is depleted on R9 290 and 290X and the GTX970 is competing only with the 390/390X the 970 prices may actually creep up. On the other hand, if NVidia releases a refresh then 970 prices will come down, and probably either way 390/390X prices will come down. Although I'm not at all satisfied that the 390 is faster in 1080p than a well-cooled factory overclocked 290X, so that might not be a winner if 390 prices settle down lower than current 390 prices but higher than current 290X prices.

Stop waiting, just buy. If you wait, you can make an argument to wait forever, limitlessly.

Always something cheaper/faster around the corner. Buy now, enjoy now.

They dont need refresh.Why you think maxwell needs refresh?
we didnt hear single speculation about that so its not gonna happen.

Yes it absolutely will happen, as Pascal is in a galaxy far far away and there's plenty of time to fit in a nice 14nm die-shrink refresh as I've already written. Money is the reason of course.

 

[Head1985]

But tell me why they need rebrand/refresh GTX970/980?They can sell GTX980 for 300USD if they want.Pascal is most likely only year ahead.GP100 already tapeout.
GTX970/980 are still more than competetive vs 390/390x.For 1080P gaming they are still better.

 

[werepossum]

Stop waiting, just buy. If you wait, you can make an argument to wait forever, limitlessly.

Always something cheaper/faster around the corner. Buy now, enjoy now.



Yes it absolutely will happen, as Pascal is in a galaxy far far away and there's plenty of time to fit in a nice 14nm die-shrink refresh as I've already written. Money is the reason of course.

True, but I have a deadline - Fallout 4. So it's not like I'm waiting for something which will never arrive.

 

[Keysplayr]

You mean something like G80 -> G92 ish? Anything is possible. If Pascal is indeed that far away, we just might see respun Maxwell on 14nm.

 

[tviceman]

You mean something like G80 -> G92 ish? Anything is possible. If Pascal is indeed that far away, we just might see respun Maxwell on 14nm.

No I don't mean shrinks of existing chips. I don't think that will happen unless Pascal is and has been behind schedule. I think a fully unlocked GM200 (aka Titan minus 12gb vram) with higher boost speeds will come out as the new flagship (GTX 1080 TI) of a new lineup of Maxwell rebrands, pushing previous tiers of performance down to slightly cheaper prices.

So GM204 goes from being GTX 980 at 1200mhz boost to gtx 1070 at 1300mhz boost, etc.

 

[Silverforce11]

Definitely a full GM200 6GB SKU.

 

[jpiniero]

I would not expect a Maxwell "16"/"14" nm, except maybe at the low end. You have to remember the lack of a $/transistor benefit so the only real benefit would be lower power consumption and it's not like Maxwell is doing poorly there.

A 28 nm rebadge to fill in the lineup to go with the Pascal Titan, yes.

 

[tviceman]

I would not expect a Maxwell "16"/"14" nm, except maybe at the low end. You have to remember the lack of a $/transistor benefit so the only real benefit would be lower power consumption and it's not like Maxwell is doing poorly there.

A 28 nm rebadge to fill in the lineup to go with the Pascal Titan, yes.

We may see small Pascal come out and round out the bottom of Maxwell's rebadge lineup, like GM107 did with Kepler.

 

[Silverforce11]

NV has shown they learnt from their mistakes and improve, so its likely they will test the next node with some Maxwell shrink to figure out potential issues, before going gung ho with Pascal and HMB2 on an unproven node.

 

[tviceman]

NV has shown they learnt from their mistakes and improve, so its likely they will test the next node with some Maxwell shrink to figure out potential issues, before going gung ho with Pascal and HMB2 on an unproven node.

I'm fairly sure next gen small die GPU's won't be HBM enabled on first gen. Nvidia hasn't done a shrink since GM200b in Jan. '09. I'm betting Nvidia doesn't shrink Maxwell, rather they introduce GP107 6-7 months ahead of any other Pascal part like they did with GM107 and Maxwell to wet their feet with a new node while also beefing up their low-power GPU for high volume notebook sales.

 

[JDG1980]

We may see small Pascal come out and round out the bottom of Maxwell's rebadge lineup, like GM107 did with Kepler.

That's certainly a possibility. I've suggested that AMD should try something similar, debuting Arctic Islands on 14nm FinFET with a small chip using GDDR5 as soon as they possibly can.

If Nvidia did bring out GP107 in late 2015 or early 2016, it would make GM206 one of the shortest-lived GPUs we've seen in the past couple of years. It only came out this January, but a small-die 16nm FinFET chip would definitely outclass it. 16nm/14nm is expected to have about double the transistor density of 28nm, so we can expect a Pascal chip with a ~150 mm^2 die size to have 3.8 billion transistors, which is considerably more than GM206. Doubling the shader count of GM107 would give 1280 "CUDA cores", which would beat the GM206 by 25%. Nvidia could release it with 4GB of RAM and put it at the same price point as AMD's 4GB R9 380, and it would beat AMD's offering in performance and absolutely demolish it in perf/watt. We'd be looking at a card using no more power than GTX 750 Ti, but offering better performance than a GTX 960. It would be the perfect HTPC card or upgrade for OEM systems. And even on a new process node, it shouldn't be difficult to manufacture a chip this small. Even ShintaiDK's pessimistic charts show the cost-per-transistor for 16nm to be only marginally more than 28nm, so silicon manufacturing costs shouldn't be much more than GM206.

 

[tviceman]

That is exactly what will happen, but not quite as soon as you say. I am guessing a late Q1 release, or perhaps very early Q2 2016. But yes, it would kill GM206 and make it a short-lived chip that had limited product release in only 1-2 skus.

 

[Enigmoid]

That's certainly a possibility. I've suggested that AMD should try something similar, debuting Arctic Islands on 14nm FinFET with a small chip using GDDR5 as soon as they possibly can.

If Nvidia did bring out GP107 in late 2015 or early 2016, it would make GM206 one of the shortest-lived GPUs we've seen in the past couple of years. It only came out this January, but a small-die 16nm FinFET chip would definitely outclass it. 16nm/14nm is expected to have about double the transistor density of 28nm, so we can expect a Pascal chip with a ~150 mm^2 die size to have 3.8 billion transistors, which is considerably more than GM206. Doubling the shader count of GM107 would give 1280 "CUDA cores", which would beat the GM206 by 25%. Nvidia could release it with 4GB of RAM and put it at the same price point as AMD's 4GB R9 380, and it would beat AMD's offering in performance and absolutely demolish it in perf/watt. We'd be looking at a card using no more power than GTX 750 Ti, but offering better performance than a GTX 960. It would be the perfect HTPC card or upgrade for OEM systems. And even on a new process node, it shouldn't be difficult to manufacture a chip this small. Even ShintaiDK's pessimistic charts show the cost-per-transistor for 16nm to be only marginally more than 28nm, so silicon manufacturing costs shouldn't be much more than GM206.

GP107 is much more likely to be in the ~120 mm^2 range due to the high costs of 14 nm.

 

[NTMBK]

There are cheaper high-capacity GDDR5 chips becoming available- the same memory chips that are on the R9 3XX series- and I suspect that NVidia will use these to boost the memory capacity of their lineup in a refresh. 8GB 980 refresh, maybe even a 24GB Titan X Black (or whatever they call it).

 

[JDG1980]

GP107 is much more likely to be in the ~120 mm^2 range due to the high costs of 14 nm.

IIRC, Nvidia said they would be sticking with TSMC. So they'll be using 16FF+ for the next node, not 14nm.

I don't see why 150 mm^2 would be too much even on an immature node. It's a very reasonable chip size. Keep in mind that the very first chips released on 28nm were AMD's Tahiti (352 mm^2) and Nvidia's first 28nm release was GK104 (294 mm^2). Compared to that, 150 mm^2 should be no problem. Even if there was a 30% higher per-transistor cost compared to 28nm, the chip shouldn't cost any more to manufacture than GM204 (398 mm^2 on 28nm), and the low power requirements would save on PCB costs, so it could compete in the $250 retail price range while maintaining profit margins for Nvidia. Keep in mind that AIB GTX 970 cards can go as low as $300 after rebate, and those have bigger coolers, PCBs, and power stages than a hypothetical 150 mm^2 FinFET GP107 would require. And the GP107 would probably be a much higher volume part, which would act as a "pipe cleaner" for the new node and help tide it through until it reaches maturity for larger chips. The best bins would also serve as premium offerings for mobile.

 

[bryanW1995]

You mean something like G80 -> G92 ish? Anything is possible. If Pascal is indeed that far away, we just might see respun Maxwell on 14nm.

What are you hearing, if anything, on 16/14nm? I don't have any concrete data, but the impression that I'm getting is that "it's gonna be a really long time". I wouldn't be surprised to see both NV and AMD release another refresh based upon their current architectures before going down a node (or, um 2 nodes now I guess). Well, NV at least, for AMD they will just bring the new process down to the midrange probably.

28nm needs to go...

 

[ShintaiDK]

IIRC, Nvidia said they would be sticking with TSMC. So they'll be using 16FF+ for the next node, not 14nm.

I don't see why 150 mm^2 would be too much even on an immature node. It's a very reasonable chip size. Keep in mind that the very first chips released on 28nm were AMD's Tahiti (352 mm^2) and Nvidia's first 28nm release was GK104 (294 mm^2). Compared to that, 150 mm^2 should be no problem. Even if there was a 30% higher per-transistor cost compared to 28nm, the chip shouldn't cost any more to manufacture than GM204 (398 mm^2 on 28nm), and the low power requirements would save on PCB costs, so it could compete in the $250 retail price range while maintaining profit margins for Nvidia. Keep in mind that AIB GTX 970 cards can go as low as $300 after rebate, and those have bigger coolers, PCBs, and power stages than a hypothetical 150 mm^2 FinFET GP107 would require. And the GP107 would probably be a much higher volume part, which would act as a "pipe cleaner" for the new node and help tide it through until it reaches maturity for larger chips. The best bins would also serve as premium offerings for mobile.

Cost, cost and cost. You will be greatly dissapointed in 14/16nm.

 

[JDG1980]

What are you hearing, if anything, on 16/14nm? I don't have any concrete data, but the impression that I'm getting is that "it's gonna be a really long time". I wouldn't be surprised to see both NV and AMD release another refresh based upon their current architectures before going down a node (or, um 2 nodes now I guess). Well, NV at least, for AMD they will just bring the new process down to the midrange probably.

28nm needs to go...

GloFo's 14nm FinFET processes are licensed from Samsung; they will be a direct copy of Samsung's methods. And Samsung already has a shipping product with FinFET (the SoC in the Galaxy S 6 phone). There are reports that GloFo reached "volume production" with the LPE (efficiency-focused) FinFET process early this year, and the LPP (performance-focused) FinFET process, which AMD will presumably be using, is expected to be "certified" in Q3 2015 and reach actual volume production in Q1 2016. With any luck, Arctic Islands GPUs should therefore be released in H1 2016 - and they need it ASAP. It would be suicide for AMD to try another rebadge; it would just reinforce the impression that AMD can't do anything right and can't bring new products to market.

As for TSMC and Nvidia, the foundry reported achieving "risk production" of 16FF+ seven months ago. There have also been rumors that they've been looking at Samsung as a backup plan. In either case, I think we will see something from Nvidia on FinFET as early as Q1 2016, even if it's just a pipe cleaner part and not a full lineup.