Possible Radeon 390X / 390 and 380X Spec / Benchmark (do not hotlink images!!!!!!)

[drunkenmaster]

Firstly, cpus/gpus being fundamentally limited by memory performance is a fact, read up any book on cpu architecture to have it confirmed. Presuming that a single architecture doesn't gain a massive benefit from more memory speed is daft, the architecture itself was built around a memory bandwidth and latency limit. Intel know exactly what memory will be used and what platform the chip will be used in when it's being designed.

If Intel could provide 200GB/s at same or lower latency within the same power limit current memory uses then their architecture would be monumentally different to currently, it would be 100% unrecognisable.

So taking a current chip and saying IT doesn't need more bandwidth thus another chip with twice the bandwidth won't gain an advantage is, honestly, ignorant. If you paired quad channel DDR4 with a P4 3.8Ghz chip you would see no performance boost over ddr, yet if you paired ddr with a current i7 you'd see a truly dramatic drop in performance.

When you get told you can use 40GB/s bandwidth you design one chip, when you get told you can use 100GB/s you design a different chip, when you get told you can use 640GB/s you design a completely different chip again.

Comparing the bandwidth of a 390x to an existing card and saying it won't make any difference is stupid, because you could say the same about any current cards bandwidth from the view point of a card from a previous generation with half the bandwidth. Lets take for example the 4870, first gpu to use gddr4, why did it get such a performance boost over a 4850, because AMD knew that memory would be available and designed the architecture to use the higher bandwidth, as such the gddr3 version was a little bandwidth starved and the gddr5 version had enough to feed the beast. Lets not forget that the 4870 with significantly more memory bandwidth with significantly higher efficiency competed with Nvidia's best gpu, double the size, better than almost any other generation.... coincidence... not even slightly.

Again read any architecture book, memory performance is the fundamental limit to which every chip is designed around, particularly latency but bandwidth is also important.


Then we move on to as mupsmebeauty is saying, regardless of what you want to believe Kaap, HBM is drastically lower power than gddr5, and most of the power difference is in signalling rather than the chips themselves. Within the same power budget even 10W saved is 5% higher gpu core power you can use, if it's more like 40W then it's more like 20% more power the gpu core can use instead.


AS for bandwidth at a given resolution, the same frame rate at a given resolution will require more bandwidth.... but double the frame rate at 1080p and the gpu will still require more bandwidth to achieve it. More FPS at ANY resolution requires more bandwidth. The gpu may be only capable of feeding 30fps at 4k in a game or 120fps at 1080p in the same game, but bandwidth usage could be identical, it's accessing the gpus memory for every frame being produced, more frames = more memory access = more bandwidth required.

For higher performance regardless of resolution more bandwidth is required.

 

[GFX-Kid]

I am mainly a lurker and very rarely do I post, but I just wanted to say:

you have just stumbled onto one of the most misunderstood things with Ram Frequency and timings, even though the frequency is increasing, the timing increases with the frequency as the timings are a measure of how many clock cycles the memory controller has to wait before it receives a signal, the overall latency from the memory cell to the memory controller barely changes as the frequency changes, just the overall bandwidth increases. But between different DDR generations, latency has slowly decreased.

when it comes to the performance, 1600mhz DDR3 has plenty of bandwidth for games.

this does give a good explanation, if you would like to read about it.
http://www.chipestimate.com/techtalk.php?d=2011-11-22

+

Firstly, cpus/gpus being fundamentally limited by memory performance is a fact, read up any book on cpu architecture to have it confirmed. Presuming that a single architecture doesn't gain a massive benefit from more memory speed is daft, the architecture itself was built around a memory bandwidth and latency limit. Intel know exactly what memory will be used and what platform the chip will be used in when it's being designed.

If Intel could provide 200GB/s at same or lower latency within the same power limit current memory uses then their architecture would be monumentally different to currently, it would be 100% unrecognisable.

So taking a current chip and saying IT doesn't need more bandwidth thus another chip with twice the bandwidth won't gain an advantage is, honestly, ignorant. If you paired quad channel DDR4 with a P4 3.8Ghz chip you would see no performance boost over ddr, yet if you paired ddr with a current i7 you'd see a truly dramatic drop in performance.

When you get told you can use 40GB/s bandwidth you design one chip, when you get told you can use 100GB/s you design a different chip, when you get told you can use 640GB/s you design a completely different chip again.

Comparing the bandwidth of a 390x to an existing card and saying it won't make any difference is stupid, because you could say the same about any current cards bandwidth from the view point of a card from a previous generation with half the bandwidth. Lets take for example the 4870, first gpu to use gddr4, why did it get such a performance boost over a 4850, because AMD knew that memory would be available and designed the architecture to use the higher bandwidth, as such the gddr3 version was a little bandwidth starved and the gddr5 version had enough to feed the beast. Lets not forget that the 4870 with significantly more memory bandwidth with significantly higher efficiency competed with Nvidia's best gpu, double the size, better than almost any other generation.... coincidence... not even slightly.

Again read any architecture book, memory performance is the fundamental limit to which every chip is designed around, particularly latency but bandwidth is also important.


Then we move on to as mupsmebeauty is saying, regardless of what you want to believe Kaap, HBM is drastically lower power than gddr5, and most of the power difference is in signalling rather than the chips themselves. Within the same power budget even 10W saved is 5% higher gpu core power you can use, if it's more like 40W then it's more like 20% more power the gpu core can use instead.


AS for bandwidth at a given resolution, the same frame rate at a given resolution will require more bandwidth.... but double the frame rate at 1080p and the gpu will still require more bandwidth to achieve it. More FPS at ANY resolution requires more bandwidth. The gpu may be only capable of feeding 30fps at 4k in a game or 120fps at 1080p in the same game, but bandwidth usage could be identical, it's accessing the gpus memory for every frame being produced, more frames = more memory access = more bandwidth required.

For higher performance regardless of resolution more bandwidth is required.

= Kaapstad pwned!

 

[Kaapstad]

Firstly, cpus/gpus being fundamentally limited by memory performance is a fact, read up any book on cpu architecture to have it confirmed. Presuming that a single architecture doesn't gain a massive benefit from more memory speed is daft, the architecture itself was built around a memory bandwidth and latency limit. Intel know exactly what memory will be used and what platform the chip will be used in when it's being designed.

If Intel could provide 200GB/s at same or lower latency within the same power limit current memory uses then their architecture would be monumentally different to currently, it would be 100% unrecognisable.

So taking a current chip and saying IT doesn't need more bandwidth thus another chip with twice the bandwidth won't gain an advantage is, honestly, ignorant. If you paired quad channel DDR4 with a P4 3.8Ghz chip you would see no performance boost over ddr, yet if you paired ddr with a current i7 you'd see a truly dramatic drop in performance.

When you get told you can use 40GB/s bandwidth you design one chip, when you get told you can use 100GB/s you design a different chip, when you get told you can use 640GB/s you design a completely different chip again.

Comparing the bandwidth of a 390x to an existing card and saying it won't make any difference is stupid, because you could say the same about any current cards bandwidth from the view point of a card from a previous generation with half the bandwidth. Lets take for example the 4870, first gpu to use gddr4, why did it get such a performance boost over a 4850, because AMD knew that memory would be available and designed the architecture to use the higher bandwidth, as such the gddr3 version was a little bandwidth starved and the gddr5 version had enough to feed the beast. Lets not forget that the 4870 with significantly more memory bandwidth with significantly higher efficiency competed with Nvidia's best gpu, double the size, better than almost any other generation.... coincidence... not even slightly.

Again read any architecture book, memory performance is the fundamental limit to which every chip is designed around, particularly latency but bandwidth is also important.


Then we move on to as mupsmebeauty is saying, regardless of what you want to believe Kaap, HBM is drastically lower power than gddr5, and most of the power difference is in signalling rather than the chips themselves. Within the same power budget even 10W saved is 5% higher gpu core power you can use, if it's more like 40W then it's more like 20% more power the gpu core can use instead.


AS for bandwidth at a given resolution, the same frame rate at a given resolution will require more bandwidth.... but double the frame rate at 1080p and the gpu will still require more bandwidth to achieve it. More FPS at ANY resolution requires more bandwidth. The gpu may be only capable of feeding 30fps at 4k in a game or 120fps at 1080p in the same game, but bandwidth usage could be identical, it's accessing the gpus memory for every frame being produced, more frames = more memory access = more bandwidth required.

For higher performance regardless of resolution more bandwidth is required.

It is going to be very embarrassing for you if the 980Ti or whatever NVidia want to call it turns out to be faster than the 390X.

I think the NVidia card will be faster and strange as it may sound it could be cheaper than the 390X too.

 

[SiDeards73]

a lot of internet hero's and Archmair professionals in this forum but i have to agree with Mauller and DM, unfortunately Kaap i think your talking out of your bottom on this one mate

Why else would Nvidia be moving to HBM if GGDR5 is not holding current architechture back?

 

[Ayahuasca]

Why else would Nvidia be moving to HBM if GGDR5 is not holding current architechture back?

For the future when more people are using 4k and beyond.

The majority of the market is still at 1080p.

 

[Kaapstad]

a lot of internet hero's and Archmair professionals in this forum but i have to agree with Mauller and DM, unfortunately Kaap i think your talking out of your bottom on this one mate

Why else would Nvidia be moving to HBM if GGDR5 is not holding current architechture back?

HBM is the future and does have advantages.

Whether I am talking out of my bottom or not at least I have just stuck my neck out and said that I think the 980Ti will be faster than the 390X.

This is a whole lot better than typing a wall of text and still sitting on the fence.

 

[Tonester0011]

980ti (if it's called that) will be faster than the 390X.
HBM has advantages, obviously. However we're currently not at the point where those advantages will show in games. Give it another generation, or two

 

[TNA]

HBM is the future and does have advantages.

Whether I am talking out of my bottom or not at least I have just stuck my neck out and said that I think the 980Ti will be faster than the 390X.

This is a whole lot better than typing a wall of text and still sitting on the fence.

980ti (if it's called that) will be faster than the 390X.
HBM has advantages, obviously. However we're currently not at the point where those advantages will show in games. Give it another generation, or two

You guys could be right, but apart from gut feeling and maybe looking at historical data, what are you basing your conclusions on exactly?

 

[drunkenmaster]

It is going to be very embarrassing for you if the 980Ti or whatever NVidia want to call it turns out to be faster than the 390X.

I think the NVidia card will be faster and strange as it may sound it could be cheaper than the 390X too.

DId you see me say the 390x would be faster than a 980ti? Thus why would it be embarrassing? Making claims off specious reasoning doesn't make someone right if the claim becomes true. History is full of people making somewhat normal claims but with incorrect reasoning, then when the thing happens they take it as truth that their incorrect reasoning must be correct.

Saying chips aren't fundamentally limited by memory based off incorrect reasons does make you wrong, it doesn't mean a chip will gain performance with more memory performance, I haven't claimed it will, I said it could and explained why.

Anyone can make a random claim, when you back up your claim with sensible reasoning it means a lot more. Your claim could turn out right but because you would post reasoning it could be seen your reasoning was correct, or completely wrong even if the claim is still true.

You're saying that more bandwidth doesn't increase performance, it will be useless, etc, this is categorically incorrect.... that doesn't mean AMD made a chip that WILL use more bandwidth effectively, it means they COULD, there is fundamentally more performance available when you design a chip around 512GB/s of lower latency bandwidth than 320GB/s of higher latency bandwidth.

AS it happens the 390x is rather like the 6970, a chip designed for a different but late process. The chip they would want, a 14nm 350-400mm^2 chip with HBM, more transistors and lots of design might not be what they can make. I could easily see the 390x ending up being a slight improvement on existing 28nm architectures and they've matched it with HBM to get some experience with it which they can use to potentially make improvements over their current 14nm designs. It could be their full on next architecture ported back to 28nm... as with the 6970 with die size limits it's likely that anything planned for 14nm will have things missing when back ported, maybe key features, maybe a change in architecture, who knows.

I don't know what the 390x is, but your reasoning for why it won't be much faster is incorrect, it's really that simple.

For the future when more people are using 4k and beyond.

The majority of the market is still at 1080p.

Bandwidth is important at all resolutions, 1080p can use exactly as much bandwidth as 4k. Your gpu accesses memory for every single frame it produces, producing 120fps at 1080p uses drastically more bandwidth than producing 30fps at 1080p. For more performance at ANY resolution you need more bandwidth.

 

[bru]

Processors are always bottlenecked by memory, especially memory that is off die, there are two things that determine the bottleneck but they are both independent of eachother; Latency and Bandwidth.

Total rubbish

Kaap is quite correct here, as explained by DM here.

If you paired quad channel DDR4 with a P4 3.8Ghz chip you would see no performance boost over ddr, yet if you paired ddr with a current i7 you'd see a truly dramatic drop in performance.

Processors are not 'ALWAYS' bottlenecked by memory.

= Kaapstad pwned!

So no not really.

But then DM is quite right that a chip is normally designed to utilise the amount of bandwidth that it will have access too. give the chip a bit more and it might perform better, but only up until the memory bandwidth outstretches the maximum that chip can handle, extra after that would be wasted.


As for how well HBM will improve things, none of us know, we are all guessing.

Theoretically the improvement could be huge, but how well that will translate into actual usable performance we will have to wait and see.
Even then with the 390x it will be difficult to judge, for example if the 390 is the same speed as the TitanX how much of that performance is down to the new chip and how much is down to the new memory architecture, we will never know.
Because there wont be (as far as I can see it isn't possible) a HBM and GDDR5 variant of the same chip, we will never know which part is giving us how much performance.

 

[nightrider1470]

HBM and GDDR5 variant will have to be ensured by selecting the north face bridge connection

 

[Ayahuasca]

Bandwidth is important at all resolutions, 1080p can use exactly as much bandwidth as 4k. Your gpu accesses memory for every single frame it produces, producing 120fps at 1080p uses drastically more bandwidth than producing 30fps at 1080p. For more performance at ANY resolution you need more bandwidth.

So then why do insane memory overclocks rarely ever give much more performance?

 

[TheRealDeal]

So then why do insane memory overclocks rarely ever give much more performance?

From what he said it's because the chip is designed around the bandwidth available at the time. So a 290x was designed to use 320 gb/s so increasing this through overclocking won't gain you much in the way of performance. The 390x should in theory be designed to use what HBM is kicking out which is a lot more Bandwidth.

 

[Mauller]

Bru, you completely missed the part where i talked about all of the prediction and caching circuitry which is used to alleviate the issues with the processor being fed with data, but at the end of the day, all processors are memory limited due to the latency between the processor and off die memory.

A lot of research and improvements in processors, goes into the prediction circuitry and caching mechanisms between each generation.

 

[Ayahuasca]

From what he said it's because the chip is designed around the bandwidth available at the time. So a 290x was designed to use 320 gb/s so increasing this through overclocking won't gain you much in the way of performance. The 390x should in theory be designed to use what HBM is kicking out which is a lot more Bandwidth.

Sounds like ******** to me. We'll see when it arrives just how much better it is at the lower resolutions as well.

 

[TonyTurbo78]

the rate its going, the 980ti will be out first anyway lol

 

[Gregster]

I hope HBM does give some fantastic improvements but whilst there is no real strain on the memory for the resolutions we play at already, I cant see it giving anything fantastical in fps gains. Maybe at 8K you will see it stomp ahead but 1080P/1440P will be miniscule gains.

 

[humbug]

You guys could be right, but apart from gut feeling and maybe looking at historical data, what are you basing your conclusions on exactly?

They aren't, they are talking nonsense, why they are doing that i will leave others to decide.

This:

If a GDDR5 setup is not bottlenecked switching to HBM will give zero performance improvement. Graphics cards get their performance from the core not the memory.

Is being ridiculous.

You only need to look at Tahiti LE with 1536 Shaders 190GB/s of bandwidth and compare its performance to Tahiti Pro with 1792 Shader and 288GB/s of bandwidth, its a solid 15% faster (100% scaling), now compare Tahiti Pro with Tahiti XT 1796 Shaders and 288GB/s of bandwidth vs 2048 Shaders and the same bandwidth, there the performance difference is only about 5% (40% Scaling).

Its clear from that Tahiti is Bottlenecked by Bandwidth.

You can see it again moving up from Tahiti to Hawaii. the 290X has 35% more Shaders, its over 30% faster (90% Scaling), Why: 288GB/s vs 320GB/s.
Now take Hawaii vs Hawaii. 2560 Shaders vs 2816 Shader at the same bandwidth, the performance difference isn't 90 - 100% scaling like it is with the step up GPU's with more bandwidth, it is again 50% because its bottlenecked again by its memory bandwidth.

A bit of hard data shoots down baseless blanket statments every time.

 

[Kaapstad]

They aren't, they are talking nonsense, why they are doing that i will leave others to decide.

This:



Is being ridiculous.

You only need to look at Tahiti LE with 1536 Shaders 190GB/s of bandwidth and compare its performance to Tahiti Pro with 1792 Shader and 288GB/s of bandwidth, its a solid 15% faster (100% scaling), now compare Tahiti Pro with Tahiti XT 1796 Shaders and 288GB/s of bandwidth vs 2048 Shaders and the same bandwidth, there the performance difference is only about 5% (40% Scaling).

Its clear from that Tahiti is Bottlenecked by Bandwidth.

You can see it again moving up from Tahiti to Hawaii. the 290X has 35% more Shaders, its over 30% faster (90% Scaling), Why: 288GB/s vs 320GB/s.
Now take Hawaii vs Hawaii. 2560 Shaders vs 2816 Shader at the same bandwidth, the performance difference isn't 90 - 100% scaling like it is with the step up GPU's with more bandwidth, it is again 50% because its bottlenecked again by its memory bandwidth.

A bit of hard data shoots down baseless blanket statments every time.

Your figures don't make sense.

You are also trying to tell me that AMD make badly setup cards, this is something I don't believe.

 

[africanos23]

Ok so here is something for all of you

Why does the GTX980 256-bit with 224 GB/s dominate (except for a few hand picked games) the R9-290X 512-bit 320 GB/s ?

In a "perfect world" like most people want to make it, isn't the R9-290X supposed to be faster?

Just so you don't tell me its a newer chip (look at the 780Ti aswell)

This is a hypothesis as we know nothing about first gen HBM.

You are going to be beta-testing a new memory architecture for AMD so Nvidia can make it better in Pascal