AMD's Tonga 384-bit bus available but not enabled

[Sharkyx1]

Well Tonga is essentially just a revision of Tahiti. It takes the same basics (Which were excellent, and are still good to this day), improves efficiency and adds in some of the latest GPU technology. The original Tonga release came out of nowhere, and was kind of useless, because it had the 2GB VRAM, and the 280/280x already both existed. The 285 had no proper place in the market.

 

But the 380 and 380x make much more sense, since the original Tahiti core has been officially retired, and replaced the the improved upon Tonga core.

 

Yeah i'm happy to see that its now supporting freesync, TA, and all the other current features accross the board except that bloody Pitcairn GPU in the middle of the lineup.

 

Just from what i've seen performance is even slightly down in some titles. mainly due to the 970mhz core clock. it seems weird to release an identical core config.

[zMeul]

Do you think it's likely they will use rebadges when it's supposed to be a shrink down to 14/16nm?

who said anything about making a Tonga on 14/16nm?! they will use the same process as is now

[dalekphalm]

Yeah i'm happy to see that its now supporting freesync, TA, and all the other current features accross the board except that bloody Pitcairn GPU in the middle of the lineup.

 

Just from what i've seen performance is even slightly down in some titles. mainly due to the 970mhz core clock. it seems weird to release an identical core config.

The identical core config likely just means less engineering time. No need to design it from the ground up.

[dalekphalm]

who said anything about making a Tonga on 14/16nm?! they will use the same process as is now

Well all the leaks so far say that AMD has 3 new chips for Arctic Islands (Greenland, Baffin, and Ellesmere, I believe they were called?), all three of which will be on 14/16nm FinFet. Each of these new chips is likely to have a "Full fledged" core, plus a "cut down" core, similar to Fury vs Fury X with Fiji.

 

I don't really see a place for a Tonga chip in that lineup that continues to use 28nm. Where would it even fit in? If the leaks are to be believed, then it AMD will have six new GPU's at 14nm. Will some of them be older chips with a die-shrink and whatever necessary tuning? Possibly. We don't know. But in any case, I see Tonga getting a die shrink more likely than it being rebranded whole-sale.

 

Either way, we'll see in a few months.

[zMeul]

Well all the leaks so far say that AMD has 3 new chips for Arctic Islands (Greenland, Baffin, and Ellesmere, I believe they were called?), all three of which will be on 14/16nm FinFet. Each of these new chips is likely to have a "Full fledged" core, plus a "cut down" core, similar to Fury vs Fury X with Fiji.

 

I don't really see a place for a Tonga chip in that lineup that continues to use 28nm. Where would it even fit in? If the leaks are to be believed, then it AMD will have six new GPU's at 14nm. Will some of them be older chips with a die-shrink and whatever necessary tuning? Possibly. We don't know. But in any case, I see Tonga getting a die shrink more likely than it being rebranded whole-sale.

 

Either way, we'll see in a few months.

I doubt AMD will release these GPUs at the same time, maybe 2 at the most

AMD could also "plant" the new Tonga in laptops/similar like it was used in Apple products - the R9 M295X, wasn't Tonga based?

 

they could use it alongside GDDR5X and have a product cheaper than a "Arctic Island" GPU with HBM/HBM2

[Sharkyx1]

who said anything about making a Tonga on 14/16nm?! they will use the same process as is now

My point is that i don't expect AMD to keep 28nm alongside 14nm products in the 200$ range. To me it looks as if AMD was delaying new product releases to the next process shrink and TSMC kept floundering on those so we ended up in our current shitty situation.

 

I doubt AMD will release these GPUs at the same time, maybe 2 at the most

AMD could also "plant" the new Tonga in laptops/similar like it was used in Apple products - the R9 M295X, wasn't Tonga based?

 

they could use it alongside GDDR5X and have a product cheaper than a "Arctic Island" GPU with HBM/HBM2

Well look at the GCN 1.0 launch, 7970 in Jan, 7770 in feb, 7870 in march.

[zMeul]

Well look at the GCN 1.0 launch, 7970 in Jan, 7770 in feb, 7870 in march.

and then you look at Fury lineup

[dalekphalm]

and then you look at Fury lineup

I think that his point was that Fury was an anomaly. Fiji was probably supposed to be 20nm originally. But when TMSC couldn't deliver in any meaningful capacity, AMD had to hurry and retool Fiji for 28nm. Every other potential "Fiji era" chip was then likely scrapped or pushed back and retooled for 14/16nm FinFet.

 

That's as reasonable a guess as anyone has right now.

[zMeul]

I think that his point was that Fury was an anomaly. Fiji was probably supposed to be 20nm originally. But when TMSC couldn't deliver in any meaningful capacity, AMD had to hurry and retool Fiji for 28nm. Every other potential "Fiji era" chip was then likely scrapped or pushed back and retooled for 14/16nm FinFet.

 

That's as reasonable a guess as anyone has right now.

I seriously doubt it, Fiji is nothing more than a beefed up Tonga with specific MCs for HMB

Fury lineup shortage was due to HBM shortages

 

going from 28 to 20nm is not as simple as resizing "something" - I would have to guess

[RagnarokDel]

if it's already there, why not enable it? So what if you cant make more doe from this, it literally costs nothing more, it's already there...

[AlTech]

if it's already there, why not enable it? So what if you cant make more doe from this, it literally costs nothing more, it's already there...

Nah. On many of them. It may be damaged and unusable after being manufactured. And by doing this, they'll have higher yields. The bigger a GPU the more defects it is likely to have. Also, why would it be 384 bit? unnecessary power penalty.

[patrickjp93]

Nah. On many of them. It may be damaged and unusable after being manufactured. And by doing this, they'll have higher yields. The bigger a GPU the more defects it is likely to have. Also, why would it be 384 bit? unnecessary power penalty.

Because on its face it's spec-equivalent to the GTX 980 but loses to it by a fair margin.

[AlTech]

Because on its face it's spec-equivalent to the GTX 980 but loses to it by a fair margin.

Not quite spec equivalent. The 380x has half the amount of ROPs/ (32 as opposed to 64 on the GTX 980)

[JoshB2084]

Whatever happening to 512-640 bits? Used be on old ATi and nVidia GPUs, but now don't make it anymore. Or did I missed something?

[dalekphalm]

Whatever happening to 512-640 bits? Used be on old ATi and nVidia GPUs, but now don't make it anymore. Or did I missed something?

More ram chips at faster speeds also increases the memory bandwidth. And yes, some cards still use/used 512-bit -> Hawaii/Grenada. Both the 290 series and the 390 series still have 512-bit buses.

 

I've never seen a GPU with 640-bit bus. Can you link a source? As far as I'm aware, 512-bit was the highest anyone ever went (I could certainly be wrong though).

[JoshB2084]

More ram chips at faster speeds also increases the memory bandwidth. And yes, some cards still use/used 512-bit -> Hawaii/Grenada. Both the 290 series and the 390 series still have 512-bit buses.

 

I've never seen a GPU with 640-bit bus. Can you link a source? As far as I'm aware, 512-bit was the highest anyone ever went (I could certainly be wrong though).

 

Oh, I'm sorry about  640-bit bus. I am slow at catch up on new technology... Technology developing is too fast for me. *sighs*

[zMeul]

More ram chips at faster speeds also increases the memory bandwidth.

someone correct me if I'm wrong, but it doesn't work that way

 

when you say (bus) width you are referring on how many bits can be accessed at the same time per clock cycle

[TheRandomness]

So Tonga has the potential to be released on 3 different generations at least?

 

Or are we getting a a 380xx or a 390-x?

Or what Nvidia has done and started to phase out 2GB GPUs and bring in 3GB GPUs at the same (or a slightly higher) cost?

[dalekphalm]

someone correct me if I'm wrong, but it doesn't work that way

 

when you say (bus) width you are referring on how many bits can be accessed at the same time per clock cycle

Yes, that is correct.

 

The bus width, is only one part of the equation that equals bandwidth though.

 

A wide bus (eg: 512-bit) + slow ram chips can produce the same effective bandwidth as a low bus (Eg: 384-bit) + fast ram chips.

 

Although I'm not certain how the number of physical chips can affect the bus width (It's possible that they do not).

 

My point was simply that GPU manufacturers could get away with a smaller bus by having higher clocked RAM, because they can achieve the same effective bandwidth.

[zMeul]

Yes, that is correct.

 

The bus width, is only one part of the equation that equals bandwidth though.

 

A wide bus (eg: 512-bit) + slow ram chips can produce the same effective bandwidth as a low bus (Eg: 384-bit) + fast ram chips.

 

Although I'm not certain how the number of physical chips can affect the bus width (It's possible that they do not).

 

My point was simply that GPU manufacturers could get away with a smaller bus by having higher clocked RAM, because they can achieve the same effective bandwidth.

not entirely correct, and I need someone with deeper knowledge of this - in theory yes, in practice the data is read in batches

the less number of cycles you need to use for retrieval is better

[Exhilarare]

You don't follow what's said about him too much do you. A lot of people accuse him of making AMD look bad on purpose. Hence my post.

.

I do follow what's said about him, and he hasn't done any outright badmouthing for a few months now. The old beating of the "AMD and zMeuls" horse should come to an end.

[dalekphalm]

not entirely correct, and I need someone with deeper knowledge of this - in theory yes, in practice the data is read in batches

the less number of cycles you need to use for retrieval is better

Perhaps. Until I see a source that says otherwise though, what you're saying is purely speculation.

[patrickjp93]

not entirely correct, and I need someone with deeper knowledge of this - in theory yes, in practice the data is read in batches

the less number of cycles you need to use for retrieval is better

And that's why having both a wide bus and a high memory clock is the best combination (along with low latency and power and heat and...)

[don_svetlio]

And that's why having both a wide bus and a high memory clock is the best combination (along with low latency and power and heat and...)

We have a saying in my country

Не може хем вълка сит, хем агнето цяло.

It means - you can't have it in a way that allows for all the positives to be there without any of the negatives

[Curufinwe_wins]

We have a saying in my country

Не може хем вълка сит, хем агнето цяло.

It means - you can't have it in a way that allows for all the positives to be there without any of the negatives

But.... GM200 has both high memory clocks and wide bus...

 

 

I mean "you can't have your cake and eat it?"