TSMC Beats Samsung For NVIDIA's Pascal GPU - Confirmed to Feature 16GB of HBM2

[Clyne]

The entire R9 3** is the old one rebadged. Its been done to death and no matter how much AMD fans would like to think otherwise-AMD hasn't done anything to the 300 series if using held back drivers doesn't count.

Only thing I mad is that Pitcairn is still there and there's no full Tonga card.

[Mahigan]

The entire R9 3** is the old one rebadged. Its been done to death and no matter how much AMD fans would like to think otherwise-AMD hasn't done anything to the 300 series if using held back drivers doesn't count.

 

True, R300 parts do have more VRAM, higher VRAM and core clocks as well. Other than that... same as the 290 series.

 

And guess what? AMD is regaining market share because of those cards. In fact, all over Reddit and in several forums... folks tend to recommend an R9 390 over a GTX 970. For a rebadged part, not bad at all.

[Dabombinable]

Only thing I mad is that Pitcairn is still there and there's no full Tonga card.

Your in denial. Grenada for example is Hawaii.

 

True,

 

And guess what? AMD is regaining market share because of those cards. In fact, all over Reddit and in several forums... folks tend to recommend an R9 390 over a GTX 970. For a rebadged part, not bad at all.

 

R300 parts do have more VRAM, higher VRAM and core clocks as well. Other than that... same as the 290 series.

That's for people upgrading from something with low performance, such as a GTX 750ti or 760. Otherwise upgrading from the older cards is pointless.

[Clyne]

Your in denial. Grenada for example is Hawaii.

 

 

That's for people upgrading from something with low performance, such as a GTX 750ti or 760. Otherwise upgrading from the older cards is pointless.

I know the 390/X were Hawaii/Grenada, they were refreshes.

I'm just mad that Tonga is still 380 and not a full core, and Pitcarin is rebranded/refreshed too many times that you should just get Tahiti instead. 

[patrickjp93]

nVidia's naming scheme is more retarded... AMD's is easier.

x60 series: mid-low

x70 series: mid

x80 series: mid-high

x90 series: high

With nvidia it changes...

750 series: mid low

760-70: mid high

780/titan series: high-enthusiast...

While for 900 series is:

950: mid

960: mid high

970-80: high

980 ti / titan: enthusiast.

You forgot the 780TI and Titan Black, leaving the 780 and Titan at high.

[Briggsy]

True, R300 parts do have more VRAM, higher VRAM and core clocks as well. Other than that... same as the 290 series.

 

And guess what? AMD is regaining market share because of those cards. In fact, all over Reddit and in several forums... folks tend to recommend an R9 390 over a GTX 970. For a rebadged part, not bad at all.

 

Considering the technology in Hawaii is still technically at least a generation ahead of Maxwell 2.0, people buying 390s are getting cards that will last quite a few years for fairly cheap.

[Mahigan]

Your in denial. Grenada for example is Hawaii.

 

 

That's for people upgrading from something with low performance, such as a GTX 750ti or 760. Otherwise upgrading from the older cards is pointless.

 

True, but Hawaii is a great GPU. Look, we're so focused on present-day performance because reviewers, see tech journalists, tend to lack the architectural knowledge required to fully understand the GPUs we're dealing with. They tend to run a few gaming benchmarks, without talking to developers or API creators, and formulate conclusions based on present-day performance.

 

Case in point. At launch, the R9 290x was around 7%, on average over many games available then, faster than a GTX 780. Now run today's games... and the R9 290x is even faster than that compared to the GTX 780. Heck during the 780 Ti launch, the former was 10% faster, across games available then, than the R9 290x. Now? 5-6% faster. With neither the GTX 780 and 780 Ti being parallel monsters, relying mostly on tweaked drivers whose function it is to optimize the software-side scheduling, how do you think DX12 gaming performance will fare on these cards relative to an R9 290x? With less room for driver interventions, replacing shaders at the driver level for optimized shaders being impossible under DX12, how will these cards fare?

 

With higher compute workloads, in DX12 titles, compared to DX11 titles and a trend which is looking to rely less on Graphics performance and more on compute... what can we discern? Well a Hawaii based GPU has nearly the same theoretical compute performance as a GTX 980 Ti (both stock clocked). Hawaii does shine when it is thrown parallel workloads and this is what Asynchronous Compute does.

 

Basically an R9 290x will be faster than both the GTX 780 and GTX 780 Ti under DX12 titles. I'd be surprised to see a GTX 780 Ti being faster. 

 

So why upgrade if you don't need too? In order to make the R300 series more attractive, AMD added twice the VRAM in the 390/390x relative to the 290/290x. They also bumped up the clock speed of the core and VRAM. This makes these cards attractive because nVIDIAs GTX 7xx series, as well as lower end GCN parts, are starting to show their age. AMD didn't need to make big changes to Hawaii, in Grenada, because the architecture is strong to begin with. Obviously Hawaii will be retired soon but it should be quite the performer come the first batch of DX12 titles and should continue to shine for some time to come.

[BuckGup]

NVLink is for Datacentres. It allows many GPUs (numbering in the 100s and thousands) to be connected with one another, over a super fast medium, so that they may sync up there workloads. It won't be useful for Desktop PCs because we're only talking about up to 4 GPUs being connected together.

Thank god as I was questioning the whole proprietary Nvidia Motherboard thing. Yeah I don't have to upgrade my Haswell!

[ivan134]

Seems to me like arguing over a naming scheme is petty no? I find neither the nVIDIA or AMD naming schemes confusing.

I know that R7 and R9 indicate the GPU architectural level and market and that the following numbering scheme indicates the performance, of each part, relative to one another on that architectural level. 

 

For nVIDIA, I know that a GTX 9xx vs a GTX 7xx indicates the architectural level and the following numbers GTX x60, x70 x80 indicates the performance, of each part, relative to one another on that architectural level.

 

Not confusing to me in the slightest.

I agree for the naming scheme of both companies, but for some reason this is a popular straw grasping subject on this forum. The only problem with the R9 200 series was when AMD released Tonga out of nowhere and you had to explain to a lay person why a 280x was better than a 285. Other than that, complaining about the naming scheme is making a mountain of a hole that goes from here to China.

[Dabombinable]

True, but Hawaii is a great GPU. Look, we're so focused on present-day performance because reviewers, see tech journalists, tend to lack the architectural knowledge required to fully understand the GPUs we're dealing with. They tend to run a few gaming benchmarks, without talking to developers or API creators, and formulate conclusions based on present-day performance.

 

Case in point. At launch, the R9 290x was around 7%, on average over many games available then, faster than a GTX 780. Now run today's games... and the R9 290x is even faster than that compared to the GTX 780. Heck during the 780 Ti launch, the former was 10% faster, across games available then, than the R9 290x. Now? 5-6% faster. With neither the GTX 780 and 780 Ti being parallel monsters, relying mostly on tweaked drivers whose function it is to optimize the software-side scheduling, how do you think DX12 gaming performance will fare on these cards relative to an R9 290x? With less room for driver interventions, replacing shaders at the driver level for optimized shaders being impossible under DX12, how will these cards fare?

 

With higher compute workloads, in DX12 titles, compared to DX11 titles and a trend which is looking to rely less on Graphics performance and more on compute... what can we discern? Well a Hawaii based GPU has nearly the same theoretical compute performance as a GTX 980 Ti (both stock clocked). Hawaii does shine when it is thrown parallel workloads and this is what Asynchronous Compute does.

 

Basically an R9 290x will be faster than both the GTX 780 and GTX 780 Ti under DX12 titles. I'd be surprised to see a GTX 780 Ti being faster. 

 

So why upgrade if you don't need too? In order to make the R300 series more attractive, AMD added twice the VRAM in the 390/390x relative to the 290/290x. They also bumped up the clock speed of the core and VRAM. This makes these cards attractive because nVIDIAs GTX 7xx series, as well as lower end GCN parts, are starting to show their age. AMD didn't need to make big changes to Hawaii, in Grenada, because the architecture is strong to begin with. Obviously Hawaii will be retired soon but it should be quite the performer come the first batch of DX12 titles and should continue to shine for some time to come.

A certain card called the R9 290X 8GB already existed. And to get an 8GB version of the 290, all that needed to be done was the flawed GPU used and a different bios. The vRAM that is used on the 300 series is simply of a higher quality and could have been used on the 200 series. As for the clock speed. Its quite easy to do that yourself.

[Mikjul]

Oh fuck my ass sideways and call me bonny...........

 

 

 

 

 

 

 

16 chuggabits of vee rum?

[Mahigan]

A certain card called the R9 290X 8GB already existed. And to get an 8GB version of the 290, all that needed to be done was the flawed GPU used and a different bios. The vRAM that is used on the 300 series is simply of a higher quality and could have been used on the 200 series. As for the clock speed. Its quite easy to do that yourself.

 

Yes, an 8GB version of the 290x existed. AMD simply relaunched the 300 series to be 8GB versions all around. When the 290 series launched, AMD relied on the 512-bit bus in order to obtain a large amount of memory bandwidth. Back then, considering they were competing with a GTX 780, it didn't make sense to include fast GDDR5 memory. AMD used slightly slower memory instead (to keep costs down). Nowadays, GDDR5 memory is relatively cheap... therefore you can include much faster memory for the same price.

 

While it is true that you can overclock a card yourself, as I do, not every user overclocks their cards as it may void your warranty.

 

http://wccftech.com/amd-regaining-gpu-market-share-nvidia/

 

 

 

Barron’s analysis is based on tracking GPU pricing data of different discrete graphics products at various price points / segments over a period of time. The analysis is claimed to be highly accurate and works on the basis that GPU vendors adjust product pricing – usually by cutting prices – based on demand and sell-through to gain market share, maintain market share or minimize market share loss.  The analysis found that prices for AMD’s $200 and $400 GPUs either matched or very slightly exceeded MSRPs.

 

AMDs marketshare is up because of the 300 series. Seems to me that they're selling quite well.

[Dabombinable]

Yes, an 8GB version of the 290x existed. AMD simply relaunched the 300 series to be 8GB versions all around. When the 290 series launched, AMD relied on the 512-bit bus in order to obtain a large amount of memory bandwidth. Back then, considering they were competing with a GTX 780, it didn't make sense to include fast GDDR5 memory. AMD used slightly slower memory instead (to keep costs down). Nowadays, GDDR5 memory is relatively cheap... therefore you can include much faster memory for the same price.

 

While it is true that you can overclock a card yourself, as I do, not every user overclocks their cards as it may void your warranty.

 

http://wccftech.com/amd-regaining-gpu-market-share-nvidia/

 

 

AMDs marketshare is up because of the 300 series. Seems to me that they're selling quite well.

That's OK as long as people aren't fooled into upgrading for example from an R9 290X to an R9 390X. (295X2 to 395X2 is a different story, AMD went full retard with the 295X2's power connectors)

[Mikjul]

Also that chart says that it'll have FP64 support, which is something that nVidia doesn't put on the consumer cards because it uses a lot of power and generates a lot of heat.

GTX 480 was the last true FP64 GPU from Nvidia with full support, damn thing was a compute monster back then.

[Zah]

True, but Hawaii is a great GPU. Look, we're so focused on present-day performance because reviewers, see tech journalists, tend to lack the architectural knowledge required to fully understand the GPUs we're dealing with. They tend to run a few gaming benchmarks, without talking to developers or API creators, and formulate conclusions based on present-day performance.

 

Case in point. At launch, the R9 290x was around 7%, on average over many games available then, faster than a GTX 780. Now run today's games... and the R9 290x is even faster than that compared to the GTX 780. Heck during the 780 Ti launch, the former was 10% faster, across games available then, than the R9 290x. Now? 5-6% faster. With neither the GTX 780 and 780 Ti being parallel monsters, relying mostly on tweaked drivers whose function it is to optimize the software-side scheduling, how do you think DX12 gaming performance will fare on these cards relative to an R9 290x? With less room for driver interventions, replacing shaders at the driver level for optimized shaders being impossible under DX12, how will these cards fare?

 

With higher compute workloads, in DX12 titles, compared to DX11 titles and a trend which is looking to rely less on Graphics performance and more on compute... what can we discern? Well a Hawaii based GPU has nearly the same theoretical compute performance as a GTX 980 Ti (both stock clocked). Hawaii does shine when it is thrown parallel workloads and this is what Asynchronous Compute does.

 

Basically an R9 290x will be faster than both the GTX 780 and GTX 780 Ti under DX12 titles. I'd be surprised to see a GTX 780 Ti being faster. 

 

So why upgrade if you don't need too? In order to make the R300 series more attractive, AMD added twice the VRAM in the 390/390x relative to the 290/290x. They also bumped up the clock speed of the core and VRAM. This makes these cards attractive because nVIDIAs GTX 7xx series, as well as lower end GCN parts, are starting to show their age. AMD didn't need to make big changes to Hawaii, in Grenada, because the architecture is strong to begin with. Obviously Hawaii will be retired soon but it should be quite the performer come the first batch of DX12 titles and should continue to shine for some time to come.

It's good to see though the 290x (meant to compete with the Titan and 780ti) actually starting to be able to to beat them in games and benchmarks.

[bogus]

That's OK as long as people aren't fooled into upgrading for example from an R9 290X to an R9 390X. (295X2 to 395X2 is a different story, AMD went full retard with the 295X2's power connectors)

 

Actually the "full retardness" allowed AMD to still have the fastest graphics card to this date - either cooled with an AIO or with air.

Full retardness is a subpar card @ twice the price ($3000).

[Dabombinable]

Actually the "full retardness" allowed AMD to still have the fastest graphics card to this date - either cooled with an AIO or with air.

Full retardness is a subpar card @ twice the price ($3000).

Let me see, there is a reason as to why there is the recommended power draw for PCIe power connectors-safety. My point was AMD went full retard and used far less on the 295X2 than they should have. The 395X2 actually makes far more sense.

[bogus]

Let me see, there is a reason as to why there is the recommended power draw for PCIe power connectors-safety. 

 

So you are claiming the R9 295x2 ... fried pcie connectors? I would ask you for a trustworthy source there mate.

[TroubleKlef]

Did they mean 15.5gb? 

[Dabombinable]

So you are claiming the R9 295x2 ... fried pcie connectors? I would ask you for a trustworthy source there mate.

Lol, I've seen the fried connectors and melted wires. While some PSU can handle graphics cards that run out of spec-it doesn't mean that its a good idea to do so.

[Casperian]

And here I was, wanting to skip over Pascal and pick up a Volta chip. The large amount of VRam is also great, the size and complexity of the neural network you could run on that thing sounds great to me. 16GB may be overkill for gaming but at least I'll get to have some fun with it.

[patrickjp93]

@patrickjp93

what's your prediction now?

The headline of this article is incorrect on its face. TSMC is only making the GPU die. We have 0 confirmations on who is providing the HBM. If you're going to swipe at me, it's best you don't miss.

[VagabondWraith]

The headline of this article is incorrect on its face. TSMC is only making the GPU die. We have 0 confirmations on who is providing the HBM. If you're going to swipe at me, it's best you don't miss.

I hope its Samsung. Overclocking potential of course. AMD can have the shitty Hynix memory.

[spartaman64]

Awesome, maybe we don't need 16GB of VRAM right now, but without this kind of cards technology won't improve. AMD should jump on VRAM bandwagon too.

remember back then when nvidia only had 2GB vram max and amd had 3 and 4GB of vram and everyone was shitting on amd saying we dont need that much

[patrickjp93]

remember back then when nvidia only had 2GB vram max and amd had 3 and 4GB of vram and everyone was shitting on amd saying we dont need that much

Back in 720p and 1080p, we didn't need more. Frankly we shouldn't need more than 2 for 1080p at all. There's no excuse for how badly memory is managed in games.