kit guru 8970/50 in JUNE ???

[ShintaiDK]

Kitguru is anything but credible. But on the other hand its just a matter of time before refreshes gonna be 1½ year due to 3 year nodeshifts at TSMC.

 

[RussianSensation]

Sometimes they post false rumors, not unlike other sites. They did report that GTX660Ti would be delayed to August because NV had 120k inventory of 570s left to sell. That story proved to be true.

The current generation of consoles and general slump in PC DIY market could be reasons why NV and AMD push back their next generation and keep charging $350-500 for current ageing generation (esp. NV must be loving this since 670/680 didn't even have official price cuts once).

"The sources pointed out that Apple's iPad, in addition to notebooks and desktops, have strongly impacted the PC DIY market and could result in an on-year sales drop of more than five million motherboards in the branded motherboard market in 2012." ~ TechReport

 

[Subyman]

Tail profit optimization is why we may see AMD wait. They are hoping to maximize their profits as the cost of the manufacturing process relaxes over the life time of the product. They certainly need higher margins on their next few financial statements since they reported a lower than expected outlook on their upcoming statement.

 

[blackened23]

One interesting thing i've read is that the upcoming 8k series has the ability to use system memory - speculation is that these parts will be cheap due to that. How does system memory compare to GDDR5 in terms of latency?

 

[SirPauly]

GK110 K20 is not due until December 2012. So not a single K20 has shipped to customers. Where are you getting this info?

http://blogs.knoxnews.com/munger/2012/09/the-big-computing-change-is-ta.html

 

[RussianSensation]

Sharing system memory used to be a budget solution: http://www.amd.com/fr/products/technologies/ati-hyper-memory/Pages/ati-hyper-memory.aspx

Unified memory for CPUs and GPUs is not the same thing you are thinking about I bet (it's not taking a 512MB GPU and letting it use system RAM). Unified memory is actually a strategy for their APUs and HSA computing. The actual functionality may be present in the entire 8k series but I doubt high-end GPUs would benefit from this in games.
http://www.anandtech.com/show/5493/...ed-memory-for-cpugpu-in-2013-hsa-gpus-in-2014

"A unified address space that both AMD CPUs and GPUs can access (today CPUs and GPUs mostly store separate copies of data in separate memory spaces). In 2014 AMD plans to deliver HSA compatible GPUs that allow for true heterogeneous computing where workloads will run, seamlessly, on both CPUs and GPUs in parallel. The latter is something we've been waiting on for years now but AMD seems committed to delivering it in a major way in just two years." ~ AnandTech

http://blogs.knoxnews.com/munger/2012/09/the-big-computing-change-is-ta.html

There is no mention at all what Kepler compute parts were shipped, K10 or K20. K10 has been shipping and we know that. When K10 officially launched, it was on NV's website --> See link. Where is K20? It hasn't officially launched yet. Even if from that article they did get K20s, the article says they only got 32 of them for early development testing. 32 but 14,592 still not delivered. Again, that article does not explicitly state that they installed 32 K20 Tesla parts, but even if it does, that just goes to show you how far behind K20 volume shipments are --> Again November-December 2012 most reasonable.

Here is my source:

"Sumit Gupta, GM of Nvidia's HPC business unit told The INQUIRER back at ISC that the Tesla K20 would be out before November, and while that might be so for some select customers, Nvidia is now saying that Tesla K20 cards will appear in December." ~ Source, August 8, 2012

 

[Subyman]

One interesting thing i've read is that the upcoming 8k series has the ability to use system memory - speculation is that these parts will be cheap due to that. How does system memory compare to GDDR5 in terms of latency?

The 680's Memory bandwidth is rated at 192GB/s. PCIe 3.0 would be a bottleneck since its rated at 16GB/s.

 

[sontin]

There is no mention at all what Kepler compute parts were shipped, K10 or K20. K10 has been shipping and we know that. When K10 officially launched, it was on NV's website --> See link. Where is K20? It hasn't officially launched yet. Even if from that article they did get K20s, the article says they only got 32 of them for early development testing. 32 but 14,592 still not delivered. Again, that article does not explicitly state that they installed 32 K20 Tesla parts.

They got 1000 more last week. GK110 is ready for mass production. You can deny it but it will not chance anything.

 

[SirPauly]

There is no mention at all what Kepler compute parts were shipped, K10 or K20. K10 has been shipping and we know that. When K10 officially launched, it was on NV's website --> See link. Where is K20? It hasn't officially launched yet. Even if from that article they did get K20s, the article says they only got 32 of them for early development testing. 32 but 14,592 still not delivered. Again, that article does not explicitly state that they installed 32 K20 Tesla parts.

Expands on the info:


http://www.hpcwire.com/hpcwire/2012...h_of_kepler_gpus_for_titan_supercomputer.html

 

[blackened23]

The heck?

This will be achieved with 18,688 nodes running the latest 16-core AMD CPUs and 14,592 K20 GPUs. Each node will have 32 GB of memory, supplying 2GB per CPU core.

AMD CPUs in a super computer? Does it run crysis?

 

[RussianSensation]

They got 1000 more last week. GK110 is ready for mass production. You can deny it but it will not chance anything.

Volume shipments in December 2012. You can deny it, but it's coming straight from NV. :thumbsup: Wishful thinking if you think GK110 will launch as consumer part with a 1Ghz 2880 SP, 240 TMUs as a sub-250W TDP part by March 2013. NV is fully committed to delivering K20 first to its professional customers. 1000 units going out is nothing if GK110 needs to be in the GTX770/780 line. At this pace, it'll take at least 1-2 quarters to build up sufficient inventory of left over GK110 chips that haven't already been sold to customers who pre-ordered them for months in advance for compute.

 

[SirPauly]

It just changed from not a single GPU to volume shipments!

 

[RussianSensation]

Expands on the info:

http://www.hpcwire.com/hpcwire/2012...h_of_kepler_gpus_for_titan_supercomputer.html

Ok, now expand how this relates to GK110 launch in large volumes in the consumer space in 2013? This is what we are talking about. NV has nothing ready in the form of GK110 for the consumer market at reasonable profitability target, until at least Q1 2013. AMD on the other hands is already sampling board production for HD8970.

It just changed from not a single GPU to volume shipments!

No, it didn't change. That's because you need volume shipments to launch GK110 as a consumer GPU. You can't connect the 2 points? Shipping 32 or even 1000 K20 parts to select customers who preordered these GPUs 6 months ago has little to do correlating that GK110 will be ready for launch by even January 2013. You have hundreds of orders to fulfill first. Again, there is no evidence to support the view that GK110 will launch as a consumer part before HD8970. Since we are talking about consumer parts, HD8970 appears to be much closer to launch than a GTX780 GK110 part is. Which is why the whole discussion of HD8000 series being delayed to June came up.

Why would AMD have HD8970 parts laying around and ready in January-February 2013, but give NV a 2-3 months head start then by purposely delaying them to June 2013? This is why the June 2013 time frame is very odd unless GK110 launches even later than that. Are you implying that GK110 will beat 8970 to market in the consumer space? Personally this is highly doubtful unless AMD is behind schedule/has ran into issues or NV pulls a rabbit out of a hat with GK110 production. This is why volume shipments is key. GK110 has to first fill all those pre-orders, then NV needs to build up inventory of these chips for GTX780. They'll be lucky to launch GTX780 by March 2013.

 

[sontin]

Volume shipments in December 2012. You can deny it, but it's coming straight from NV. :thumbsup: Wishful thinking if you think GK110 will launch as consumer part with a 1Ghz 2880 SP, 240 TMUs as a sub-250W TDP part by March 2013. NV is fully committed to delivering K20 first to its professional customers. 1000 units going out is nothing if GK110 needs to be in the GTX770/780 line. At this pace, it'll take at least 1-2 quarters to build up sufficient inventory of left over GK110 chips that haven't already been sold to customers who pre-ordered them for months in advance for compute.

Tape-out of GK110 was in March. nVidia got early silicon back in May. With another revision of the chip they need 6-7 months for a launch. TSMC is starting production of 28nm in a second fab in Q4 2012.

We know that you are in full "AMD marketing mode" but the reality is that GK110 is ready for mass production and nVidia is only waiting for more wafers from TSMC. And if you think that they can not put out BigK after 7 months: They showed GF100 early October 2009 and reviews went online end of March 2010 - that is less than 6 months.

But yeah it must be "Wishful thinking" to believe that nVidia doesn't need 12 months from tape-out to a market launch...

 

[SirPauly]

I don't desire to get into a subjective, semantic debate -- you clearly offered there wasn't one GPU!

So not a single K20 has shipped to customers.

Offered a link that proves other wise -- that's all.

 

[RussianSensation]

SirPauly, yes, I was wrong on the "not a single K20 part was shipped to consumers" but NV is still on track to start volume shipments by November-December 2012. That hasn't changed. Good thing you follow NV so closely that we have the latest news of 32 K20 and 1000 K20 parts shipping to Oak Ridge. Thanks for correcting me.

We know that you are in full "AMD marketing mode" but the reality is that GK110 is ready for mass production and nVidia is only waiting for more wafers from TSMC.

AMD marketing mode? Ok w/e. There is no marketing mode when all I did is question the volume available of K20 and how that translates to GK110 consumer part. I guess you only see bias when I pinpointed that volume shipments of GK110 are still nowhere near necessary to launch a GK110 consumer party by January 2013.

The official target date for Titan to be available to end users is March 2013 at which point 14,592 K20 GPUs will be utilized in the supercomputer. Please then tell us why it would take Oak Ridge until March 2013 to get all of those K20 chips if NV is in mass producing stages right now. Waiting for mass production and being in mass production mode are not the same thing. K20 will start shipping in volume by November-December 2012 based on NV's HPC management's comments. So when do you expect GTX780 GK110 to ship then? Before HD8970?

Based on what you are implying, NV will launch with GK110 consumer part first before 8900 series, but AMD would then purposely delay HD8900 series to June 2013 knowing this? Why do you think the June 2013 time frame is reasonable for HD8970 but GK110 can somehow be ready to go if NV just snaps its fingers?

 

[RussianSensation]

I think AMD would have been best served by releasing the 7970GE in January at the $550 price point the regular hd7970 started at.

1050mhz 7970 in January 2012 was probably impossible in sufficient volumes. The reasoning for lower clock speeds was addressed due to yields. If 1050mhz 7970 was easily doable in January 2012, why would AMD have waited 6 months for XT2 revision of the chip and 28nm node maturity to be able to launch a 1050mhz part? That does not compute. The reason to launch HD7970 925mhz was purely strategic since delaying it 2-3 months wouldn't have allowed AMD to sell even a single 7900 product and then by the time GTX680 launched, they might have been able to launch a 1Ghz 7970 part, but that part would only have been as fast as a 680 anyway, forcing AMD to drop prices from $550 to $500 anyway. Launching 2.5 months early was smart.

That's what happens in GPU space, you either launch first and get beaten later (5870 vs. GTX480) or you launch first and hope your competitor flops somehow (9700 Pro vs. 5800U), or you launch a much faster part allowing you to maintain high price throughout the generation (GTX280 vs. 4870). We've seen all these scenarios in the GPU space before. It's actually a normal state of affairs if we see a part that launches later to outperform the early part and subsequent price drops for the slower part. That's how the entire tech industry is supposed to work. Interestingly enough you said nothing about HD7970 having a 2.5 months lead for barely any more $ than GTX580 or put the blame on NV at all for allowing AMD to charge such high prices on the HD7750-7870 line for at least 6 months of this generation or said anything about a similar strategy that NV used with its initial high prices of GTX280/260. It's obvious GPU makers raise prices if the opportunity presents itself and drop them when competition arises. If anything the anomaly is HD5850/5870 pricing. That is the one generation that makes no sense at all why AMD didn't launch those 2 cards at $399-549. We got lucky.

 

[f1sherman]

The heck?



AMD CPUs in a super computer?

Cray's Gemini interconnect uses Hypertransport 3.0.
Intel bought Cray's interconnect business back in April.
PCI-E 3.0 will replace HT in Cray's next gen Cascade supercomputer and Xeons will replace Opterons. Only Nvidia will stay D:

 

[Silverforce11]

GK110 is ready for mass production and nVidia is only waiting for more wafers from TSMC.

No doubt its "ready", R&D and design and testing would be long completed by now since we are so long into this gen already.

Everyone is waiting on TSMC to increase 28nm capacity because GPUs have to compete with so many smartphone chips for production.

The question is not that they cannot have a product ready, the question is on TSMC's ability to deliver the volumes required to even fill their HPC niche market with orders, so far one lab will get their K20 with much fanfare. How many else are in line waiting? After all those are filled, we can start to talk about consumer graphics from GK110.

Plus the biggest factor you are all not discussing: BigK is really big (one would be completely delusional to not expect a massive die given the accurate specs). Do you think TSMC can even make it with decent yields on an immature node they are struggling to ramp up? Think about it. You either have too much blind faith and believe so, or a realist and judge based on their previous track record: no, they aren't likely to deliver yields that would meet both HPC and consumer demands. Until they have respectable yields on BigK, NV has NO incentive to sell a costly low yield product for $500-$700 (a "small die" gtx680 already goes for $500+) when they can sell tesla cards for ten times that.

@Russian: Until AMD feels that NV is close to launching the refresh or bigK for consumers, they dont even need to launch their refresh. A 360mm2 Tahiti die going for $300-450 is actually quite good. They could refresh with a 280mm2 at the $300 pricepoint, but its going to dislodge 78xx (much smaller die) even lower .. or 8970 at ~420mm2 (my guess) for $500 isnt a big improvement over 79xx in terms of die mm2/$.

 

[thilanliyan]

If anything, it's NV that continues to rip off consumers with $400-420 670 and $500-550 680s.

That's the power of a better brand presence for ya...look at what Apple is able to charge for their phones even though they may not be technically superior to other less expensive phones.

 

[Silverforce11]

NV has a right to charge more for an equivalent or lesser product, because many years of sustained delivery and good marketting should be rewarded.

Physx is the bomb. Didn't you all hear?

A few games a year with Physx is all the masses need to pay more for NV.

 

[tviceman]

There is no mention at all what Kepler compute parts were shipped, K10 or K20.

Yes there is.

http://www.hpcwire.com/hpcwire/2012...h_of_kepler_gpus_for_titan_supercomputer.html

They are getting K20's.

 

[RussianSensation]

@Russian: Until AMD feels that NV is close to launching the refresh or bigK for consumers, they dont even need to launch their refresh. A 360mm2 Tahiti die going for $300-450 is actually quite good. They could refresh with a 280mm2 at the $300 pricepoint, but its going to dislodge 78xx (much smaller die) even lower .. or 8970 at ~420mm2 (my guess) for $500 isnt a big improvement over 79xx in terms of die mm2/$.

HD78xx prices have fallen a lot though (7850 is going for $180-190 on Newegg). It creates room to introduce 8850/8870 for $299/349 with ~GTX670 level of performance (of course I don't know this but just going off rumors).

1Ghz 7970 is $380 on Newegg, far down from $550 MSRP of January 2012. Assuming 8950 is faster, AMD could sell that for $450, and push 8970 to $550 (up from $450 for most 7970 GEs). AMD would maintain profit margins by virtue of raising prices at each price bracket to offset the die size increase because they'll once again have an early head start and performance leadership. But the upside is that they would once again beat NV to the market by 2-3 months.

Personally, I think if AMD will delay the launch to June 2013, it'll probably be because of other factors such as yields or maybe waiting for another revision of Venux XT chip to try and get similar clocks at lower voltages to curb the power consumption associated with a 410-420mm^2 die. Allowing NV to launch GK110 closely to HD8970 could prove to be a major mistake if GK110 is 15-20% faster than HD8970. That would mean AMD would have missed a 2-3 months window selling HD8950-8970 at $500+. We know NV makes very strong GPUs and with a history of 500-576mm^2 dies, NV could really put the hurting on the 8970 if AMD waits too long.

 

[tviceman]

1050mhz 7970 in January 2012 was probably impossible in sufficient volumes. The reasoning for lower clock speeds was addressed due to yields. If 1050mhz 7970 was easily doable in January 2012, why would AMD have waited 6 months for XT2 revision of the chip and 28nm node maturity to be able to launch a 1050mhz part? That does not compute. The reason to launch HD7970 925mhz was purely strategic since delaying it 2-3 months wouldn't have allowed AMD to sell even a single 7900 product and then by the time GTX680 launched, they might have been able to launch a 1Ghz 7970 part, but that part would only have been as fast as a 680 anyway, forcing AMD to drop prices from $550 to $500 anyway. Launching 2.5 months early was smart.

That's what happens in GPU space, you either launch first and get beaten later (5870 vs. GTX480) or you launch first and hope your competitor flops somehow (9700 Pro vs. 5800U), or you launch a much faster part allowing you to maintain high price throughout the generation (GTX280 vs. 4870). We've seen all these scenarios in the GPU space before. It's actually a normal state of affairs if we see a part that launches later to outperform the early part and subsequent price drops for the slower part. That's how the entire tech industry is supposed to work. Interestingly enough you said nothing about HD7970 having a 2.5 months lead for barely any more $ than GTX580 or put the blame on NV at all for allowing AMD to charge such high prices on the HD7750-7870 line for at least 6 months of this generation or said anything about a similar strategy that NV used with its initial high prices of GTX280/260. It's obvious GPU makers raise prices if the opportunity presents itself and drop them when competition arises. If anything the anomaly is HD5850/5870 pricing. That is the one generation that makes no sense at all why AMD didn't launch those 2 cards at $399-549. We got lucky.

My post computes just fine if you quit trying to think about their situation from a perspective that people who know WTF is going on are in charge of their marketing department.

Half of what AMD does outside of engineering GPU's makes no sense whatsoever. You think yields weren't there; I think yields were just fine. All they did to get the 7970GE is bump the voltage. 7970's have had very little to no problems hitting 1050mhz right from the start. Yields were already there, they were just being stupidly conservative. We're talking about a $550 video card, those don't exactly sell by 10's of millions. They didn't need five miillion hd7970GE's at launch. Less than a hundred thousand over the course of 2-3 months prior to GK104 would have sufficed just fine. Their launch was strategic, in that they didn't think GK104 would soundly beat a vanilla hd7970. They failed with their theory, failed to react quickly, and since have been forced to readjust pricing several times and face the not being the good guy in the market.

I completely disagree with just about everything you are saying in pretty much your entire post (except the hd5800's launch prices) and there has been plenty of instances to easily refute your view of the market, but I'm not going to take the time to start 10 page long arguments/discussions with you. Three easy instances are GTX580 and GTX570 launched before Cayman and both were faster, and 8800GT launched to no competition and undercut it's own lineup significantly.

 

[tviceman]

I think most people in this thread put too much stock into the idea that each company is releasing products based on the other company. I think both companies release products for the market and to get more customers more than to compete with each other.

If anyone here thinks that a company should never release a new product, or to delay in order to see what the competition might bring, would fail hard. If you have a new product that is ready to go, you release. Newer products can always command higher prices and that is the best thing for a company. Waiting until there is competition in the market to release a product gives the customer choices BESIDES your product AND a premium can't be charged.