(Debunked) AMD Fury X production limited to 30,000 units in 2015.

[Sakkura]

Yes they are. Sparc systems use it specifically for high-bandwidth memory for the integrated accelerators used (usually Xilinx FPGAs), though moreover HMC was designed with more functionality in mind, such as being useable as a cache, as volatile transactional memory, or as a little of both. HMC is built for what Nvidia wants and more. It's more than ready, as made clear by Intel's massive order of it for specific use in a GPU-like compute accelerator (with additional functionality and performance in workloads which pure GPUs will never achieve). Nvidia's only problem was being at the back of the line.

 

Seriously people, it's not that hard.

 

The only problem is it's just not true.

 

 

Nvidia’s Volta GPU module was scheduled for 2015 release and was supposed to use the Micron Hybrid Memory Cube (HMC). Rumors are that after HMC’s development fell behind the proposed roadmap timing, Nvidia moved the Volta introduction out to > 2016 after the Pascal module, which will be based on Hynix HBM 2 and is scheduled to be commercial in 2016.

[Bouzoo]

The only problem is it's just not true.

 

Nvidia’s Volta GPU module was scheduled for 2015 release and was supposed to use the Micron Hybrid Memory Cube (HMC). Rumors are that after HMC’s development fell behind the proposed roadmap timing, Nvidia moved the Volta introduction out to > 2016 after the Pascal module, which will be based on Hynix HBM 2 and is scheduled to be commercial in 2016.

Not saying it might not be true, but your source is based on rumors, no official confirmation. If you're gonna post a source behind your reasoning, post a trustworthy one.

[Opcode]

Four pages of people buying into fabricated rumors for baiting traffic for it to be then "debunked". Come on guys really, has anyone learned to not buy into social media until the product is actually launched or the actual company themselves make such a statement?  :rolleyes:

[patrickjp93]

The only problem is it's just not true.

Unsubstantiated rumors, not to mention Nvidia fast-tracked Maxwell since it had development breakthroughs and it could read market more quickly on 28nm. And, you can look for yourself in Oracle Sparc system sales. HMC has been in use for nearly two years already.

[patrickjp93]

Four pages of people buying into fabricated rumors for baiting traffic for it to be then "debunked". Come on guys really, has anyone learned to not buy into social media until the product is actually launched or the actual company themselves make such a statement? :rolleyes:

If you have to ask, the answer is no.

[Sakkura]

Not saying it might not be true, but your source is based on rumors, no official confirmation. If you're gonna post a source behind your reasoning, post a trustworthy one.

 

As opposed to all the reliable sources @patrickjp93 hasn't provided.

 

It's a highly trustworthy source.

[Bouzoo]

As opposed to all the reliable sources @patrickjp93 hasn't provided.

 

It's a highly trustworthy source.

They may have had a great history in news reporting, but as they said, "Rumors are that", means they heard it somewhere and it was newer officially stated. Then it has a high chance of being wrong. If The New York Times came and said AMD is gonna bankrupt based on a trustworthy source I'd take it with a grain of salt. If WCCF had an official statement from AMD that they are gonna bankrupt I'd believe them. There are differences yes, but you get where I'm getting at.

I never said anything about patrick being right, if he did the same thing I'd have said the same thing to him. You said it's just not true like it was 100% confirmed which it never was. In the end, a rumour is just a rumour until proven otherwise.

[patrickjp93]

As opposed to all the reliable sources @patrickjp93 hasn't provided.

It's a highly trustworthy source.

That only says it's a RUMOR as to why Volta was moved back.

[patrickjp93]

As opposed to all the reliable sources @patrickjp93 hasn't provided.

 

It's a highly trustworthy source.

Not as trustworthy as Oracle itself. This is a currently deployed system and has been deployed at multiple locations for 8 months. The HMC modules were in niche use before that. https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=en&ie=UTF-8&u=http%3A%2F%2Fwww.heise.de%2Fnewsticker%2Fmeldung%2FSparc-Prozessor-fuer-100-Petaflop-Rechner-2290438.html&edit-text=&act=url

 

Now, I suggest you take your misinformation and stick it where the sun doesn't shine. HMC is more than ready and is more than capable of going toe to toe with HBM. Nvidia just couldn't wait in line.

[Sakkura]

Not as trustworthy as Oracle itself. This is a currently deployed system and has been deployed at multiple locations for 8 months. The HMC modules were in niche use before that. https://translate.google.com/translate?sl=auto&tl=en&js=y&prev=_t&hl=en&ie=UTF-8&u=http%3A%2F%2Fwww.heise.de%2Fnewsticker%2Fmeldung%2FSparc-Prozessor-fuer-100-Petaflop-Rechner-2290438.html&edit-text=&act=url

 

Now, I suggest you take your misinformation and stick it where the sun doesn't shine. HMC is more than ready and is more than capable of going toe to toe with HBM. Nvidia just couldn't wait in line.

 

Oracle, or that article, says nothing about what happened in the development of HMC.

 

Note that the HMC in that article is delivering an inadequate 240 GB/s.

[patrickjp93]

Oracle, or that article, says nothing about what happened in the development of HMC.

Note that the HMC in that article is delivering an inadequate 240 GB/s.

You do know that's per module, right? That's better than HBM per module, given it takes 4 separate modules on 4 memory controllers to reach 512 or 640GB/s. Multiply 240 by 4 and you get 960GB/s.

At least Opcode knows his stuff more than half the time. You're an even more blind AMD/Hynix fanchild than he is.

Also, it does discuss the development. For HMC to be deployed now, it had to have been completed and proven. Causality's a bitch isn't she?

[Sakkura]

You do know that's per module, right? That's better than HBM per module, given it takes 4 separate modules on 4 memory controllers to reach 512 or 640GB/s. Multiply 240 by 4 and you get 960GB/s.

 

 

No, I don't know that, and your source disagrees with you.

 

 

At each CMG also hang four Hybrid Memory Cubes of each in each direction 30 GByte / s shoveling.

[patrickjp93]

No, I don't know that, and your source disagrees with you.

Do you even know what shoveling means?

[Sakkura]

Do you even know what shoveling means?

 

Schaufeln. Or shoveling shit, I guess, but you might have more expertise there.

[patrickjp93]

Schaufeln. Or shoveling shit, I guess, but you might have more expertise there.

In either case you're entirely incorrect and you can read the technical specs of HMC yourself directly from Micron. Each module at this point on the high end is up to 240GB/s each way. Sure Oracle took the early products which still blew DDR3 and DDR4 out of the water in raw bandwidth, performance efficiency, and raw power usage.

[Sakkura]

In either case you're entirely incorrect and you can read the technical specs of HMC yourself directly from Micron. Each module at this point on the high end is up to 240GB/s each way. Sure Oracle took the early products which still blew DDR3 and DDR4 out of the water in raw bandwidth, performance efficiency, and raw power usage.

 

They do blow DDR4 out of the water, but HBM and arguably even GDDR5 blow HMC out of the water (unless you're building a supercomputer).

 

Here are Micron's specs. 120-160 GB/s from a 31x31mm die is just... not great.

[patrickjp93]

They do blow DDR4 out of the water, but HBM and arguably even GDDR5 blow HMC out of the water (unless you're building a supercomputer).

Here are Micron's specs. 120-160 GB/s from a 31x31mm die is just... not great.

On 28nm lithography, that's not bad. Hynix's HBM is on 20nm currently and is a less complex and versatile memory structure. It's not transactional (great both for databases for speed and for security in hypervisors).

[Sakkura]

It's bad enough to make HMC unviable for graphics cards, forcing Nvidia to adopt HBM instead.

[MrRavens]

AMD plz (Q_Q ) Unless they have this crazy genius plan, like the smart guys in an anime that when every thought they lost, they had a plan all along that's all complex and shit.