Read more.Could it put the brakes on Nvidia's Pascal plans in 2016?
Read more.Could it put the brakes on Nvidia's Pascal plans in 2016?
This could be quite a coup for AMD. Combined with the experience their engineers will already have with working on HBM1 it could give them quite an advantage over Nvidia in the next gen of cards.
One can hope at least.
Sound like a double edged sword, does AMD hoover up as much HBM2 as they can in the hope that they can shift plenty of GPUs and hopefully limit Nvidia's supply, running the risk of having unsold stock.
Or is something like memory production not done on a pre-order basis?
They probably only have to take enough stock that nvidia couldn't launch a new gpu line using it, since they would never go to the expense of having 1 specific card using it and all others on the old ram system.
given that HBM is in limited supply - holding up the Fury range - contracting to buy everything from1 supplier until they dont need anymore would be the Apple way of doing it.
Could be AMDs final bad move if they try and leverage that to stifle competition. Do they have the funds to buy it all up and let it sit in a warehouse gathering dust?
After some of the comments about nVidias business practices recently, not one of them come close to trying to prevent a competitor from making a new product like this seems to imply.
I guess this could get interesting and may force nVidia to throw it's weight around a bit more.
Main PC: Asus Rampage IV Extreme / 3960X@4.5GHz / Antec H1200 Pro / 32GB DDR3-1866 Quad Channel / Sapphire Fury X / Areca 1680 / 850W EVGA SuperNOVA Gold 2 / Corsair 600T / 2x Dell 3007 / 4 x 250GB SSD + 2 x 80GB SSD / 4 x 1TB HDD (RAID 10) / Windows 10 Pro, Yosemite & Ubuntu
HTPC: AsRock Z77 Pro 4 / 3770K@4.2GHz / 24GB / GTX 1080 / SST-LC20 / Antec TP-550 / Hisense 65k5510 4K TV / HTC Vive / 2 x 240GB SSD + 12TB HDD Space / Race Seat / Logitech G29 / Win 10 Pro
HTPC2: Asus AM1I-A / 5150 / 4GB / Corsair Force 3 240GB / Silverstone SST-ML05B + ST30SF / Samsung UE60H6200 TV / Windows 10 Pro
Spare/Loaner: Gigabyte EX58-UD5 / i950 / 12GB / HD7870 / Corsair 300R / Silverpower 700W modular
NAS 1: HP N40L / 12GB ECC RAM / 2 x 3TB Arrays || NAS 2: Dell PowerEdge T110 II / 24GB ECC RAM / 2 x 3TB Hybrid arrays || Network:Buffalo WZR-1166DHP w/DD-WRT + HP ProCurve 1800-24G
Laptop: Dell Precision 5510 Printer: HP CP1515n || Phone: Huawei P30 || Other: Samsung Galaxy Tab 4 Pro 10.1 CM14 / Playstation 4 + G29 + 2TB Hybrid drive
If you want 10 chips for a prototype run then that is from stock, but 40000 of them to make a 10000 graphics cards productions run?
Most chip production is done on a "just in time" basis as no-one wants to hold large stocks. Factories order the chips something like 3 to 9 months in advance. If you demand chips right now, then that is someone else's allocation they are taking them from to try and get you supply.
If supply is constrained, then someone is going to miss out. If AMD are currently helping Hynix to develop and sell HBM1, then a bit of mutual back scratching seems reasonable when it comes to HBM2.
Or to invert it, I don't see why Nvidia should think it can turn up late to the party and help itself to the drinks cabinet.
It is in the interest of SK Hynix to make enough of these devices for anyone that wants them, so I don't see the situation lasting too long.
azrael- (14-07-2015)
Yeah a lot of this seems to be speculation on top of rumour with a bit of misunderstanding thrown in.
It's possible that, since HBM is indeed co-developed by AMD and Hynix that AMD would have priority access in one way or another. It doesn't make sense that a party would put years of R&D into development only to have a competitor benefit equally or greater from it at only purchase cost.
It also wouldn't make business sense to just buy up capacity and not use it, but that's not what's being said by the source article anyway. As DanceswthUnix said, semiconductors are very long lead-time parts, you don't come along last minute and expect the manufacturer to have a few million in stock waiting for you. The manufacturer will need to know well in advance what you need so they can allocate capacity.
AMD wouldn't in any way be 'preventing' Nvidia from making anything. If HBM had not been developed by AMD/Hynix then I doubt Nvidia would have just given up producing GPUs. In fact, the pictures of 'Pascal' prototypes I've seen around the web are not HBM (note no interposer). Going back a few years Nvidia were planning to use HMC if memory serves correctly but decided to switch to HBM more recently - it would be pretty bad business to switch to a competitor's product and simply assume there would be capacity if they hadn't discussed it with the involved parties well in advance.
If anything the 'priority' is probably something like AMD is guaranteed capacity since they invested in it, similar to the WSA with GloFo.
Last edited by watercooled; 14-07-2015 at 02:17 PM.
Yeah, bandwidth is nothing without a nice GPU. We saw that with the Fiji. No gain whatsoever. Maybe next time.
Main PC: Asus Rampage IV Extreme / 3960X@4.5GHz / Antec H1200 Pro / 32GB DDR3-1866 Quad Channel / Sapphire Fury X / Areca 1680 / 850W EVGA SuperNOVA Gold 2 / Corsair 600T / 2x Dell 3007 / 4 x 250GB SSD + 2 x 80GB SSD / 4 x 1TB HDD (RAID 10) / Windows 10 Pro, Yosemite & Ubuntu
HTPC: AsRock Z77 Pro 4 / 3770K@4.2GHz / 24GB / GTX 1080 / SST-LC20 / Antec TP-550 / Hisense 65k5510 4K TV / HTC Vive / 2 x 240GB SSD + 12TB HDD Space / Race Seat / Logitech G29 / Win 10 Pro
HTPC2: Asus AM1I-A / 5150 / 4GB / Corsair Force 3 240GB / Silverstone SST-ML05B + ST30SF / Samsung UE60H6200 TV / Windows 10 Pro
Spare/Loaner: Gigabyte EX58-UD5 / i950 / 12GB / HD7870 / Corsair 300R / Silverpower 700W modular
NAS 1: HP N40L / 12GB ECC RAM / 2 x 3TB Arrays || NAS 2: Dell PowerEdge T110 II / 24GB ECC RAM / 2 x 3TB Hybrid arrays || Network:Buffalo WZR-1166DHP w/DD-WRT + HP ProCurve 1800-24G
Laptop: Dell Precision 5510 Printer: HP CP1515n || Phone: Huawei P30 || Other: Samsung Galaxy Tab 4 Pro 10.1 CM14 / Playstation 4 + G29 + 2TB Hybrid drive
Talking things up is a skill of theirs. I expect to see Pascal boards held out for all to see which turn out to have wooden HBM chips screwed into place
Seriously though, I think Fury proves that the current generation isn't generally bandwidth constrained. Perhaps the next high end (will that be Titan Y?) will need HBM to feed double the shaders of a Titan X, but Pascal mid range can probably stay on DDR5 without performance problems.
OTOH, if AMD are betting the farm by putting their whole range on HBM then things could get interesting.
Most likely. I find it extremely hard to believe Nvidia would be caught off-guard by HBM volume available to them. Committing to use something without any knowledge of volume would be a catastrophically bad move. Moving away from HMC in the first place was likely to do with it not being ready in time (and/or at necessary volumes/costs) after all.
As I say I imagine any sort of priority is simply a reserved fab capacity like AMD also have with GloFo, or Samsung have with their own fabs.
And before that they were talking about Pascal with HMC, and also generically about "3D" memory. And they're currently talking about Pascal in 2016, so that's at least 6 months away. So it's reasonable to think that supply contracts may not have been finalised yet.
AMD getting priority doesn't mean nvidia aren't allowed through the door. It may mean nvidia have to delay launch of Pascal, or launch in limited quantities. But nvidia are hardly going to start bad-mouthing SK Hynix's product allocation plans if they're also reliant on SK Hynix for the memory for their next generation of cards, are they...? Just because they aren't whinging doesn't mean they're happy with the deal they're being offered.
I see Intel is using HMC parts for Xeon Phi.
http://www.anandtech.com/show/9436/q...h-100-isc-2015
Edit to add: Although I suppose Larrabee is well known for delays, so shouldn't read much into that.
There are currently 1 users browsing this thread. (0 members and 1 guests)