Samsung GDDR6 @ CES

[Krteq]

Look on many forums dude,people are hating on HBM2 for some reason....

I don't get it. Why would someone hate a standard which brings more throughput? Yes, it's been ATi/AMD (namely Joe Macri) who brings all these GDDR3/4/5, HBM(2) standards to us, but those are still INDUSTRY STANDARDS, so everyone can use them without any fees.

Those haters are simply ... because it will brings more performance to us, customers.

 

[Cloudfire777]

I don't get it. Why would someone hate a standard which brings more throughput? Yes, it's been ATi/AMD (namely Joe Macri) who brings all these GDDR3/4/5, HBM(2) standards to us, but those are still INDUSTRY STANDARDS, so everyone can use them without any fees.

Those haters are simply ... because it will brings more performance to us, customers.

We are not haters. Its called reading through the marketing brainwashing AMD have been doing since they launched Fiji and being painfully late with Vega and thought they could use HBM as an argument to why they suck at keeping up with the competition at launching cards at the correct time.

Again,

HBM2 use MORE power than GDDR6. One of the reasons HBM was made and was hyped through marketing was that it used less power than GDDR5. Which it does. But HBM2 use more power than GDDR5. And GDDR6 use less power than GDDR5.

HBM use less space than GDDR. Thats absolutely true. But we have already seen tiny GDDR videocards. Just as small as R9 Nano.
With GDDR6 we get 2x the density as GDDR5. Which means they can build even smaller videocards with that.
In the end GPUs need cooling no matter if it got HBM or GDDR anyway which makes cards big and beefy.

Then there is this bandwidth argument which was another cornerstone of HBM.
Fiji had a bandwidth of 512GB/s while GDDR5 980Ti had 336GB/s.
Fiji wasnt magically faster although the bandwidth was so much higher.
Just like the «HBM takes less space» argument above, its another moot point for videocards. For professional users, absolutely nescessary. For gamers, moot.
Volta V100 with HBM2 have a bandwidth of 900GB/s. A GDDR6 384bit card will have 750GB/s bandwidth.
Its Fiji and 980Ti all over again only this time they both offer much more bandwidth than whats required for games.

HBM makes manufacturing more complicated. Its more expensive.
Wake me up when HBM3 arrives, ie 2020. Because HBM1 and HBM2 is pretty much a huge MEH.

 

[Yotsugi]

HBM2 use MORE power than GDDR6.

redacteddude, hands off the crack. Power savings on memory controllers alone will make the difference.

Fiji had a bandwidth of 512GB/s while GDDR5 980Ti had 336GB/s.

And? 980ti had ~1.2k ALUs less.

Fiji wasnt magically faster although the bandwidth was so much higher.

No redacted, there's more to GPU than that.

A GDDR6 384bit card will have 750GB/s bandwidth.

Assuming you find 16Gbps chips in the wild first.

Its Fiji and 980Ti all over again.

How are different memory standards even related to that?

Because HBM1 and HBM2 is pretty much a huge MEH.

That's why anything HPC or any product that gives zero redactedabout margins uses it.
What a HUUUUEG meh!

No swearing is allowed. Member callouts are not allowed.
nathanddrews
AnandTech Moderator

 

[Cloudfire777]

redacteddude, hands off the crack. Power savings on memory controllers alone will make the difference.

And? 980ti had ~1.2k ALUs less.

No shit, there's more to GPU than that.

Assuming you find 16Gbps chips in the wild first.

How are different memory standards even related to that?

That's why anything HPC or any product that gives redactedabout margins uses it.
What a HUUUUEG meh!

The only argument you had here was HPC, the rest was a bunch of quotes with zero content. As for HPC, you are preaching to the wrong crowd when most users here are gamers and the users that need HBM are the same niche users in the list of HBM products you listed in your previous reply to me.

Its a reason why Nvidia stayed with GDDR and will continue to do that with Volta (minus the niche products not for gamers).
But keep on preaching AMDs smoke and mirrors. The only one smoking crack here must be you guys
Its pretty much telling that its mostly gamers too that believe HBM is the holy grail of gaming. So easily deceived.

«Why are you waiting on Vega dude? Its delayed again after several delays.»
«It got HBM man!»

A guy hyped by marketing.

 

[Yotsugi]

users that need HBM are the same niche users in the list of HBM products you listed in your previous reply to me

Users don't *need* HBM. Manufacturer decides whether they want to sacrifice the margins.

reason why Nvidia stayed with GDDR

Called margins.
You gotta milk the userbase.

 

[Qwertilot]

Users don't *need* HBM. Manufacturer decides whether they want to sacrifice the margins.

Called margins.
You gotta milk the userbase.

Doubt if its margins - the (in)ability to produce enough chips must have been a huge deterrent for NV in the consumer space.

 

[Yotsugi]

Doubt if its margins - the (in)ability to produce enough chips must have been a huge deterrent for NV in the consumer space.

It's totally margins, because maybe not memory (volume discounts are a thing, besides the cartel is pricing even G5/G5X outta their minds) but 2.5D packaging.
Only Intel currently has a cost-effective way for 2.5D integration, being EMIB, that is.
Amkor promised S-SLIM but it's still nowhere to be found.
TSMC and SPIL are still pushing 2.5D interposers.

 

[raghu78]

It's totally margins, because maybe not memory (volume discounts are a thing, besides the cartel is pricing even G5/G5X outta their minds) but 2.5D packaging.
Only Intel currently has a cost-effective way for 2.5D integration, being EMIB, that is.
Amkor promised S-SLIM but it's still nowhere to be found.
TSMC and SPIL are still pushing 2.5D interposers.

Exactly. The problem with HBM2 is packaging technology. 2.5D is not cost and yield effective for high volume mainstream usage. I am sure the OSATs like Amkor and foundries like TSMC/GF/Samsung are working on cost effective and simpler packaging for 7nm products. I think Amkor is in development on SLIM which eliminates the need for TSVs. I think we will see better packaging technologies in high volume production by 2020. Long term once packaging issues are solved HBM will become the dominant memory standard for high performance compute (which it is already) and graphics.

Sent from my SM-G935V using Tapatalk

 

[CatMerc]

Sigh, as always, people aren't aware that HBM's main purpose for existence is to stave off the proportionally increasing amount of energy required just for memory as compared to GPU. We're increasing bandwidths significantly, but the power/bit barely gets reduced. HBM was born out of trying to delay this longer. Had AMD's GPU's been as efficient as NVIDIA's, their cards would be significantly more efficient in practice because the memory subsystem would take a fraction the energy than that of NVIDIA's.

Spoiler: Large Image

https://www.anandtech.com/show/9266/amd-hbm-deep-dive

The benchmark for HBM failing would be the energy/bit of GDDR to somehow match or be below HBM's, which it's nowhere near doing.

 

[Bier667]

Isn't HBM main purpose based on neural networking redacted?

And people with a Zen CPU still arguing about whats best for the their fellowers?


Sorry but GDDR6 gives enough headroom for density and GB/s for your yearly AC and COD/Battlefield game galore

No profanity in tech forums.
Usandthem

 

[Qwertilot]

HBM is, I'd think quite safely, self evidently failing as a tech for mass market consumer cards - its just too expensive/tricky to produce for the required cost/volume.

Doesn't mean that its failing entirely of course

 

[tential]

Well that is more than questionable, dude.
The cartel will milky-milky nVidia to death.

Which cartel is milking Nvidia to death? I want to get milked like that.....

 

[CatMerc]

Isn't HBM main purpose based on neural networking redacted?

And people with a Zen CPU still arguing about whats best for the their fellowers?


Sorry but GDDR6 gives enough headroom for density and GB/s for your yearly AC and COD/Battlefield game galore

No profanity in tech forums.
Usandthem

wat