![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
Intel has started benchmarking their high end GPU's
- Thread starter Lakados
- Start date
Krenum
Fully [H]
- Joined
- Apr 29, 2005
- Messages
- 19,193
lol, funny how that works.
Having a third player is not a bad thing.
Maybe it'll keep the other two on their toes and add a level of competition to drive cost back to some sense of normalcy.
Lord knows it has gotten out of hand in recent years.
A damn 1660 GPU should not cost 300 'anything'...and yet I'm seeing them in the $400 range.
That's just crazy.
Maybe it'll keep the other two on their toes and add a level of competition to drive cost back to some sense of normalcy.
Lord knows it has gotten out of hand in recent years.
A damn 1660 GPU should not cost 300 'anything'...and yet I'm seeing them in the $400 range.
That's just crazy.
NightReaver
2[H]4U
- Joined
- Apr 20, 2017
- Messages
- 3,799
A 3rd player literally means nothing in such an extreme sellers' market. If they have an equivalent product, they can price just as high and it will fly off shelves. There is ZERO incentive for them to price lower when people gobble up literally everything off the shelf like a swarm of locusts.
That's certainly one way of looking at it.
Another is.... If I have the capital and want my company to gain a foothold into a segment I have never competed in, I am going to 'flood' potential consumers.
If I have the capability to manufacture my own chips and am looking for return on a long-term investment, then I want my GPU in everybody's hands.
I'll offer a good beginning product...(even if it is not fastest out-of-the-box) with pricing to reflect that.
The hope (much like car sales) is that the consumer will see value and be a repeat buyer of each generation.
My emphasis would also be on giving the best product support and customer service.
Price can increase later when I am established and have taken a bite out of the other two players.
Does that selling technique...(less the big budget, C/S and P/S sound familiar?)
Re: AMD.
Last edited:
NightReaver
2[H]4U
- Joined
- Apr 20, 2017
- Messages
- 3,799
In a normal market, sure. They really wouldn't have to compete with market share right now if they have a well reviewed product at comparable prices to the competition. People AREN'T being picky atm. They're buying 3060s for $500 lol. They're so desperate for any hardware, they'd gladly scoop up 1st gen Intel hardware for near the same price if the performance was there.
If 7nm is the problem and intel has that problem, how would intel's FABs help that bottleneck?
Intel used to be the monopoly problem, but you now have another de facto monopoly in the supply chain - TMSC. Issue is everyone wants the best.
NightReaver
2[H]4U
- Joined
- Apr 20, 2017
- Messages
- 3,799
That's the other issue. More competition don't mean jack if they all share the same resource bottleneck.
To Nvidias credit... they gave Samsung a shot. Can't argue with the chips the NV GPUs seem solid.... production is probably better then we think with the increased demand on GPUs the last year.
I would say we have at least 2 viable high end fabs... when I said we need Intel fabbing. Yes we need Intel to get their 7nm and beyond stuff actually working. Otherwise Intel just becomes one more mouth for TSMC to feed. I believe their first round of GPUs is going to be fabed by TMSC. It would be good for everyone if Intel manages to get their own stuff working.
IF Intel can manage to get that running by the end of the year. It could mean the next generation. We could end up with AMD gpus from TMSC... NV from Samsung and Intel from their own fabs. I would say that would go a long way to ensure GPU supply. Assuming Intel can at least make a card that can compete in the high volume mid range market. I doubt they have a halo crown winner in them anytime soon... but a good mid ranger could relive some stock issues. As silly as it sounds I am more excited about Intel GPUs then much else going on right now. I'm a Linux user... and as much as I hate a lot of what Intel does. Their support of open source stuff is without a doubt the best in the industry. I have no doubt Intels Linux support would be by far the best of all the players. AMD has come a long way open source wise, however no doubt Intel still does the best job. A decent Intel GPU might give me a good excuse to switch over to Intels Clear Linux. lol
There are a couple of things people seem to be forgetting about when it comes to the current state of cutting edge production.
First off was that Global Foundries tapped out of the game not too long ago. This is what forced AMD to go to TSMC for the CPU side of their Zen 2 and Zen 3 chips (the IO die is still a GF affair). This leaves only three major players on the cutting edge of fabrication: Intel, TSMC and Samsung. Despite Intel attempting to open themselves up as a foundry, their failures at 10 nm and 7 nm caused an exodus of initial customers. (I've personally thought that the acquisition of Altera by Intel was modified by the contractual penalties and probable lawsuit Intel would lose as Altera was on of the few who signed up for their 10 nm and 7 nm nodes. It was cheaper in the long term to spend $16.7 billion to buy that company.) That realistically leaves TSMC and Samsung as the only two foundries open for 3rd party designs that companies could leverage.
Going back a few more years, Intel had a massive excess of production when their 22 nm node was in full swing and 14 nm was on the horizon. Instead of continuing to have massive excess, they decided to postpone the opening of fab 42 in Arizona which was originally targeted for 14nm. Now Intel is in the middle of a self created shortage and has just now brought fab 42 online. Of course this isn't with another calibration as it was re-planned for 7 nm but finally scaled back to 10 nm. After the fiasco that was 10 nm, Intel was hoping to transition quickly to 7 nm but that appears to be even worse than their 10 nm line up. The current expectation is that Ice Lake-SP will the first products out of this facility. While the rest of the foundries would be having issues right now, Intel would have weathered things fine and wouldn't have had to resort to using TSMC for PC chipset production. (One of the under reported things is that through various acquisitions of the past decade, Intel inherited a massive amount of contracts with TSMC. Intel's long term goals for their acquisitions was to move everything in house where it'd ultimately be cheaper and presumably on a better node anyway.) This would also have had the secondary effect of freeing up some production in TSMC as well depending on where Intel's acquisition were in the transition to Intel's own fabs.
Of course one way to increase production is to use larger wafers. 300 mm wafer were expected to be replaced by 450 mm wafers by now (Intel's Fab 42 was expected to be their first 450 mm facility.) Despite the silicon industry sinking billions into the development of 450 mm wafer standards, the main consortium behind it collapsed after numerous delays. There is no technical reason preventing the development of 450 mm wafers appearing in the future but it will be years away in the unlikely event it will ever happen now. However, had 450 mm production started as originally envisioned in the 2015-2016 time frame, there wouldn't be the severe shortage of chips we are currently facing. The arrival of 450 mm wafers would have likely offset the volume bottleneck that multipattern processes created. One interesting aspect of 450 mm wafers is that large monolithic dies like nVidia's >800 mm^2 GV100 and GA100 monsters would be far more economical to produce.
The only thing that will ease the current crunch a bit is EUV which is just now being adopted and effectively a requirement for processes post 7 nm. This reduces the number of steps in the production line as 14 nm class node relied heavily on multi-patterning to achieve their feature size. However the roll out of EUV machinery has been slow. The arrival of additional EUV machinery and the transition to those nodes using it will likely be the most immediate resolution to the current manufacturing crunch.
While not directly affecting production, large designs are quickly moving to the direction of chiplets. This will allow designs to leverage mixed nodes and have higher yields overall due to less wasted silicon.
First off was that Global Foundries tapped out of the game not too long ago. This is what forced AMD to go to TSMC for the CPU side of their Zen 2 and Zen 3 chips (the IO die is still a GF affair). This leaves only three major players on the cutting edge of fabrication: Intel, TSMC and Samsung. Despite Intel attempting to open themselves up as a foundry, their failures at 10 nm and 7 nm caused an exodus of initial customers. (I've personally thought that the acquisition of Altera by Intel was modified by the contractual penalties and probable lawsuit Intel would lose as Altera was on of the few who signed up for their 10 nm and 7 nm nodes. It was cheaper in the long term to spend $16.7 billion to buy that company.) That realistically leaves TSMC and Samsung as the only two foundries open for 3rd party designs that companies could leverage.
Going back a few more years, Intel had a massive excess of production when their 22 nm node was in full swing and 14 nm was on the horizon. Instead of continuing to have massive excess, they decided to postpone the opening of fab 42 in Arizona which was originally targeted for 14nm. Now Intel is in the middle of a self created shortage and has just now brought fab 42 online. Of course this isn't with another calibration as it was re-planned for 7 nm but finally scaled back to 10 nm. After the fiasco that was 10 nm, Intel was hoping to transition quickly to 7 nm but that appears to be even worse than their 10 nm line up. The current expectation is that Ice Lake-SP will the first products out of this facility. While the rest of the foundries would be having issues right now, Intel would have weathered things fine and wouldn't have had to resort to using TSMC for PC chipset production. (One of the under reported things is that through various acquisitions of the past decade, Intel inherited a massive amount of contracts with TSMC. Intel's long term goals for their acquisitions was to move everything in house where it'd ultimately be cheaper and presumably on a better node anyway.) This would also have had the secondary effect of freeing up some production in TSMC as well depending on where Intel's acquisition were in the transition to Intel's own fabs.
Of course one way to increase production is to use larger wafers. 300 mm wafer were expected to be replaced by 450 mm wafers by now (Intel's Fab 42 was expected to be their first 450 mm facility.) Despite the silicon industry sinking billions into the development of 450 mm wafer standards, the main consortium behind it collapsed after numerous delays. There is no technical reason preventing the development of 450 mm wafers appearing in the future but it will be years away in the unlikely event it will ever happen now. However, had 450 mm production started as originally envisioned in the 2015-2016 time frame, there wouldn't be the severe shortage of chips we are currently facing. The arrival of 450 mm wafers would have likely offset the volume bottleneck that multipattern processes created. One interesting aspect of 450 mm wafers is that large monolithic dies like nVidia's >800 mm^2 GV100 and GA100 monsters would be far more economical to produce.
The only thing that will ease the current crunch a bit is EUV which is just now being adopted and effectively a requirement for processes post 7 nm. This reduces the number of steps in the production line as 14 nm class node relied heavily on multi-patterning to achieve their feature size. However the roll out of EUV machinery has been slow. The arrival of additional EUV machinery and the transition to those nodes using it will likely be the most immediate resolution to the current manufacturing crunch.
While not directly affecting production, large designs are quickly moving to the direction of chiplets. This will allow designs to leverage mixed nodes and have higher yields overall due to less wasted silicon.
N4CR
Supreme [H]ardness
- Joined
- Oct 17, 2011
- Messages
- 4,947
Don't forget Hector Ruinez... That guy screwed AMD hard.
Armenius
Extremely [H]
- Joined
- Jan 28, 2014
- Messages
- 42,084
Ironically, TSMC's 7nm is less dense than Intel's 10nm.
Was going to say. Intel has been shipping 10nm for a couple years, now. Just not on the -S high-powered desktop parts people on this forum want.
Is that still the case? I know that was true with how Intel initially planned 10nm, but my understanding was that they relaxed density targets twice due to the defect rate, so is that still the case?
TSMC N7FF 97.5
TSMC N7FF+ 113.9
Intel 10nm 100.8
So it either is or is depending on which TSMC 7nm node you are comparing too. Intel isn’t the only company who can add a + to their process names after all.
Armenius
Extremely [H]
- Joined
- Jan 28, 2014
- Messages
- 42,084
Actual numbers I've seen for Intel's 10nm were between 96 and 102 MTr/mm². TSMC's original 7nm was 91 MTr/mm², but I know they have made some improvements since then. They are probably similar to each other in density today.
Problem is Intel has admitted to backing up 10nm density but they have never released exact specs on what they have relaxed to get it to work. But the higher up fab folks at Intel have all suggested they had to back up their initial target densities to get yields that where not complete shit. Of course all those people have been fired since. The truth is I'm not anyone sure outside of Intel really knows at this point how dense their current 10nm stuff. It probably doesn't matter much anyway.... lets be honest Intel is basically sticking to 14 till they have their 7 ready to go. Which is probably very close to TMSCs process. The Intel marketing propaganda of their 10nm being > then 7 has really caught hold. Their plans was aggressive for sure... but they never really got it to work and they backtracked a lot of those design decisions to get it half working.
Intel’s 10nm has been weighed in at 100.8, falling ahead of TSMC’s N7FF but behind their N7FF+ process. Intel ended up changing the shape of the transistors and their layout. They were originally trying to deploy 10nm with their lateral nanowire transistors but never made that work at 10nm but it will resurface later as gate all around probably in the 3-5nm range. But instead of this they found a way to reduce the number of dummy gates between the transistors and remove remove the air gap between them.
That's fair... I would suggest the ways they found don't really work however. I mean their yields at 10nm are still too crap to build their best parts at 10. Perhaps those air gaps where a good way to ensure you didn't have a prohibitively high defect rate... which is the main issue Intel is having. They can make 10nm work but it seems the defect rates are higher a lot higher. Perhaps its a simple probability issue they should have considered.... high density sounds great and all but it drastically increased the number of defects that hit actual important bits rather then air gaps.
I don’t know much about their yields, but all their 10’th and 11’th gen mobile parts have been at 10nm so they are getting them out in quantity. From my understanding it’s been more of an issue of power scaling, their parts have been having issues when subjected to lighter power targets beyond the 45-65w found in laptops, which could be a yield issue, unknown. But this year we do get the first 10nm Desktop, HEDT, and Server parts so we’ll know more in a few months.That's fair... I would suggest the ways they found don't really work however. I mean their yields at 10nm are still too crap to build their best parts at 10. Perhaps those air gaps where a good way to ensure you didn't have a prohibitively high defect rate... which is the main issue Intel is having. They can make 10nm work but it seems the defect rates are higher a lot higher. Perhaps its a simple probability issue they should have considered.... high density sounds great and all but it drastically increased the number of defects that hit actual important bits rather then air gaps.
That is a good point... I assume yield. Your point is valid though it could be more about power budgets and leakage with the removal of the gaps as you say. In any event its not the wonder node Intel expected it to be... I do really hope they get their own 7 up and running soon. Intel needs to be able to produce at 7 for Intel GPUs to really help with supply issues in the market. And if there going to actually compete with AMD in the high end down the road. It looks like their last 14nm chips will perhaps win some single thread crowns... but that doesn't mean much if there getting lapped (not even an exaggeration) in multi core by AMD chips with twice the cores.
Perhaps clearing out all the FAB leadership last year was for the best... I really hope so as much as I rag on Intel. It does worry me though that the new CEO as much as I was hoping for an actual engineer to head things... was there previous CTO who actually oversaw all those perhaps in retrospect insane aggressive fab plans. And his first few public statements on the matter sound more like him doubling down then admitting they Fd up. Intels Geniuses will overcome... rather then we shot to high and we have it dialed in now. (Dr. Murthy Renduchintala did admit they Fd up and needed to aim at realistic targets... and I don't know but I bet that was a contributing factor to his canning)
From their internals 7nm is “on-track” and the DoE dropped their lawsuit, or at least back it off so that must mean Intel is back on making the required progress in their supercomputer. The yields are probably absolutely garbage though.
though it could be both yield and process for 10nm. The design just may not work on a monolithic chip which is why their higher power chips are all looking to be multi chip designs.
FWIW I don't wish Intel ill-will on any of this. For one thing, the US needs a leading-edge foundry, having all of the world's advanced foundries in Taiwan is bad news. And competition is good. Just Intel hasn't delivered in a while, so I will believe it when they start to ship it.
Someone sawed through a couple of Intel and TSMC chips a while back and put up pictures. Shouldn't be any harder than that was to do it again. Take the cheapest Tiger Lake laptop CPU and a 5600X and do the same, maybe right through the middle of the ALU and right through the L3 cache. 5600Xs are relatively easy to find recently (Micro Center Houston has listed "25+" in stock for weeks, for example.)
If AMD announced that they were building their own fabs, I would sell my stock the same hour. I invested after they got out of the fab business, and it was one of the best investments I’ve made. Go look up the returns AMD after the split and tell me if it was a good move. Then tell me if you would want AMD to start their own fabs again.
Last edited:
griff30
Supreme [H]ardness
- Joined
- Jul 15, 2000
- Messages
- 7,206
I remember the 1999 Taiwan Earthquakes caused a massive problem with memory shortage.
I sold back a bunch of EDO ram for 10x the price I paid for it to a memory sales place I originally bought it from.
Imagine if that happens again, because it will. You don't build a silicone empire on a Taiwanese house of cards.