Radeon VII

[melmac]

I swear these sites do my nut, they take a story add their own spin to it like its their info and then publish it..apparently without reading through it. May as well have said " we have it on good authority this card will be faster than a geforce 2".

Some of those sites can be pretty annoying. I just don't click links anymore unless I know the site is decent enough.

 

[Davedree]

Considering Fury and Vega were losses, i'd consider that V7 follows the same trajectory with the additional fact Amd cannot ride the mining wave anymore, and their v7 only competes with the mid range chip from the opposition.

You do forget is a chopped chip not the full Vega 7nm....

same applies to the rtx 2080 though, it's not the full chip

 

[nashathedog]

FuryX benchmarks against the 980ti ring any bells.

And as I said, AMD's own marketing suggest it falls short of the 2080.
https://www.eurogamer.net/articles/...-first-radeon-7-benchmark-results-in-25-games

At lower resolutions Radeon 7 will fall further behind, as it is only the huge bandwidth helping at 4K maxed out.

So even if we trust AMD's figures are 100% accurate, then by their own admission it is slower than a 2080.

The only question is how much slower.

I remember & I put that down to the slippery snake known as Raj Koduri, If there's one thing I learnt from watching how he dealt with AMD's consumers it's that he couldn't be trusted to tell the truth or give a straight answer, He isn't an honest person, Look at how the Vega 64 & Vega VII pre-release info compares. For the Vega 64 we got given an out the box speed claim of 1630mhz but if you can reach 1630 once you do all the tweaking & overclocking your doing well, we got quoted numbers that made it look close to Pascal on paper but the difference was that with Pascals claimed box speeds were at the low end of what it can do while the quoted speeds for Vega were the opposite.

This time we're being told in advance that it's up to 1800 mhz. To me that distinction being made upfront is a good sign, Hopefully they've come to terms with the fact that they need to earn back the trust Koduri lost for them & a good start would be being straight about the VII & what it can do, We'll see though.

 

[Gerard]

Considering Fury and Vega were losses, i'd consider that V7 follows the same trajectory with the additional fact Amd cannot ride the mining wave anymore, and their v7 only competes with the mid range chip from the opposition.

same applies to the rtx 2080 though, it's not the full chip

I doubt very much Vega was a loss considering the miners went mental for it. Plus the FE card probably sold decently as well.

 

[Illuminist]

The performance will be between 1080ti and 2080ti....could they cast their net of speculation a wee bit wider? Talk about hedging your best ffs. And they will feature upto 16 gigs of ram...umm they're already confirmed to be 16 gig cards, what is this "upto" ****?

Maybe there will be 8 and 16gb versions despite what we all think

 

[Gerard]

​

Maybe there will be 8 and 16gb versions despite what we all think

8 gig would half the bandwidth and hurt performance quite a bit. Not to mention the full 16gig is on the package already and paid for, so they would have to disable half of that and sell the card for less even though the ram is physically there. It's happened with the 480 but that was using much cheaper ram at the time.

 

[TheRealDeal]

​

8 gig would half the bandwidth and hurt performance quite a bit. Not to mention the full 16gig is on the package already and paid for, so they would have to disable half of that and sell the card for less even though the ram is physically there. It's happened with the 480 but that was using much cheaper ram at the time.

Couldn't they just change the 4gb stacks to 2gb stacks. That would leave bandwidth the same and 2gb stacks would be cheaper. They would need to manufacture them like this though instead of using the failed chips which would all have 16gb.

 

[Illuminist]

​

8 gig would half the bandwidth and hurt performance quite a bit. Not to mention the full 16gig is on the package already and paid for, so they would have to disable half of that and sell the card for less even though the ram is physically there. It's happened with the 480 but that was using much cheaper ram at the time.

So the bandwidth is due to how much there is ? It's not the speed of the hbm to the system? It's 2 8gb block hitting the system twice so doubles the speed right?

 

[Illuminist]

Couldn't they just change the 4gb stacks to 2gb stacks. That would leave bandwidth the same and 2gb stacks would be cheaper. They would need to manufacture them like this though instead of using the failed chips which would all have 16gb.

That was my next question

 

[TheRealDeal]

So the bandwidth is due to how much there is ? It's not the speed of the hbm to the system? It's 2 8gb block hitting the system twice so doubles the speed right?

Hbm 2 gives 256GB/s of bandwidth per stack. Vega 7 has 4 x 4gb stacks.

 

[Illuminist]

Hbm 2 gives 256GB/s of bandwidth per stack. Vega 7 has 4 x 4gb stacks.

So what's stopping them replacing them with stacks of 2gb and having 4x2gb stacks on radeon VII?

 

[LePhuronn]

Isn't the entire and central point here the Radeon 7 is not a new product? It's Vega 20 + HBM packages that didn't make the cut for an Instinct card and would otherwise have just been thrown into the bin. But as it turns out, AMD could smash some more power through it, take the clocks up and lo and behold it can supposedly fight with the RTX 2080. easy win.

So all this talk about "but why couldn't they do this" or "couldn't they do that" is totally moot. No, they couldn't do this because AMD have not made anything new.

However, given that these packages were destined for the trash anyway, there's no real excuse to pitch the card at $699. Could've done $599 easily because that's still $599 they wouldn't otherwise have had.

 

[TheRealDeal]

So what's stopping them replacing them with stacks of 2gb and having 4x2gb stacks on radeon VII?

The rumour is that the actual chips being used are from failed MI50 Instinct chips. If that's the case they are already manufacted with the 4 x 4gb stacks. If they are making them from scratch then nothing is stopping them from making them with 4 x 2gb stacks.

 

[LePhuronn]

So what's stopping them replacing them with stacks of 2gb and having 4x2gb stacks on radeon VII?

Because that would explicitly mean building Radeon 7 GPU packages with dedicated RAM purchases. This card is not a new product, it's a package already made to be an Instinct card that ultimately didn't make the cut and would be otherwise discarded.

 

[Gerard]

So what's stopping them replacing them with stacks of 2gb and having 4x2gb stacks on radeon VII?

The fact that the hbm is already on the package, if these are failed instinct cards (as seems to be the consensus) then its basically impossible to remove the hbm without destroying it due to how its attached.

 

[Illuminist]

Roger, hear you loud and clear, repackaged instinct cards that woulda gone in the bin.
I don't know if i completely believe that, as mentioned they would have gone for a $600 price point and sold twice as many

 

[ltron]

So what's stopping them replacing them with stacks of 2gb and having 4x2gb stacks on radeon VII?

Buildzoid said that none of the memory manufacturers make 2GB stacks, 4GB is the lowest available size.

 

[4K8KW10]

Buildzoid said that none of the memory manufacturers make 2GB stacks, 4GB is the lowest available size.

Guess what, ask AMD to begin ordering such, then. What's wrong?
Oh, and put lower frequency stacks, too.. lol

 

[Panos]

Considering Fury and Vega were losses, i'd consider that V7 follows the same trajectory with the additional fact Amd cannot ride the mining wave anymore, and their v7 only competes with the mid range chip from the opposition.

same applies to the rtx 2080 though, it's not the full chip

Vega sold very well, both due to mining craze and especially on the business and entertainment industries. Also the GPU is pretty fine, and faster than its direct competitor, the GTX1080, as it holds better on high resolutions, especially when things like HDR are activated.

Also wrote few times here, there are companies using scores of V64s on servers for AI & Machine Learning for their businesses (there are two in York alone).
Because is dirty cheap and very powerful compared to "dedicated" products like the NV TV100. (1/8 the price for 90% the performance is bargain).

That's the benefits of using AMD GPUOpen, nothing is hidden behind licences and hardware restrictions.

 

[Panos]

Guess what, ask AMD to begin ordering such, then. What's wrong?
Oh, and put lower frequency stacks, too.. lol

Companies prefer to simplify production to the most limited production lines. Producing only 4GB stacks is cheaper on the long run than having 2GB and 4GB products. And any "custom" 2GB won't save AMD money.
In addition AMD would prefer to sell MI50/MI60, than RVII, which in effect is a salvaged part.