AMD RX 5700 Navi New arch

[Trik'Stari]

At least they finally woke up and stopped trying to make HBM a thing.

 

Now they need to just stop aiming only at the cheap seats and make a high end card people actually want to buy,

[TheDankKoosh]

9 hours ago, Firewrath9 said:

yeah, but it is more than 10% cheaper (1080s were 480$ IIRC, 1070 Ti was ~420$)

You can always OC the mem, I got a +620mhz before FPS got worse (10.5ghz mem) on my 1070 SC2.

My 1070 ti can only do +500 on the mem, so I got the metaphorical shaft in that manor, if I could do 10ghz like your 1070 then I wouldn't even be complaining, but only samsung memory can do that as far as I can tell, mine has micron.

[RejZoR]

1 hour ago, Trik'Stari said:

At least they finally woke up and stopped trying to make HBM a thing.

 

Now they need to just stop aiming only at the cheap seats and make a high end card people actually want to buy,

HBM is great actually, but AMD was just ahead of time so much it just wasn't feasible from a supply and price point. You can be assured it'll be a standard sometime in the future like GDDR5/5X/6 is now. Because it's really the only way to go forward as PCB simply cannot carry enough traces to all the memory modules when it comes to large capacity beyond 16GB.

[Firewrath9]

On 5/31/2019 at 7:45 AM, RejZoR said:

HBM is great actually, but AMD was just ahead of time so much it just wasn't feasible from a supply and price point. You can be assured it'll be a standard sometime in the future like GDDR5/5X/6 is now. Because it's really the only way to go forward as PCB simply cannot carry enough traces to all the memory modules when it comes to large capacity beyond 16GB.

Hmm. RTX Quadro 8000 has 48GB Vram, HBM2 is good, but too expensive. Also makes cooling a bitch as the hbm stacks might be taller or shorter than the die

[Firewrath9]

On 5/31/2019 at 7:28 AM, TheDankKoosh said:

My 1070 ti can only do +500 on the mem, so I got the metaphorical shaft in that manor, if I could do 10ghz like your 1070 then I wouldn't even be complaining, but only samsung memory can do that as far as I can tell, mine has micron.

+500 on mem is 10ghz actual clock for GDDR5...

 

[TheDankKoosh]

10 minutes ago, Firewrath9 said:

+500 on mem is 10ghz actual clock for GDDR5...

 

No it isn't, most pascal cards have 8ghz memory stock, not 9ghz.

[RejZoR]

1 hour ago, Firewrath9 said:

Hmm. RTX Quadro 8000 has 48GB Vram, HBM2 is good, but too expensive. Also makes cooling a bitch as the hbm stacks might be taller or shorter than the die

It's funny though, we make things at 7nm scale, but have problems aligning two big ass chips measured in millimeters to the same height on the PCB... XD

[Trik'Stari]

6 hours ago, RejZoR said:

HBM is great actually, but AMD was just ahead of time so much it just wasn't feasible from a supply and price point. You can be assured it'll be a standard sometime in the future like GDDR5/5X/6 is now. Because it's really the only way to go forward as PCB simply cannot carry enough traces to all the memory modules when it comes to large capacity beyond 16GB.

My issue was primarily the price, availability, and more importantly, the performance just wasn't there to justify the first two. IIRC none of their HBM cards were capable of beating out a 1080ti overclocked.

[Trixanity]

46 minutes ago, Trik'Stari said:

My issue was primarily the price, availability, and more importantly, the performance just wasn't there to justify the first two. IIRC none of their HBM cards were capable of beating out a 1080ti overclocked.

The memory was, the GPU itself not so much. It's like strapping 8 channels of DDR4 memory to Bulldozer and wondering why Intel could be faster with dual channel DDR3. If you remove the memory bottleneck but you're bottlenecked in multiple other areas then you won't accomplish much. Besides, HBM will be for ultra high performance going forward until HBM gets cheaper and/or GDDR is no longer viable.

[CarlBar]

5 hours ago, Firewrath9 said:

Hmm. RTX Quadro 8000 has 48GB Vram, HBM2 is good, but too expensive. Also makes cooling a bitch as the hbm stacks might be taller or shorter than the die

 

 There's a reason thats a 10k card, (beyond the usual reasons Quadro's are expensive). 4GB GDDR6 chips are really expensive, i don't know anything like exact numbers but i wouldn't be shocked to find that it's on par with HBM pricing. It's also the current upper limit. Whilst the cost is going to come down and the capacities likely will increase some more it's really not a good situation for GDDR6 as the top end cards keep pushing the limits of what it can do whilst HBM based cards are nowhere near the limits, (thats at least 128Gb of capacity and 3.2TB of bandwidth).

[_Hustler_One_]

Seeing this RX 5700 GPU is possibly at RTX 2070 level, with possibly lower price than 2070, this might be the GPU which we (or just me) really need for over 6Gb VRAM 1080p 144 fps standards without being overkill.

 

Do you guys think so?

 

I'm still looking for a GPU for my 1080p 144 fps gameplay with more than 6Gb VRAM without cost me too much money or bad value for my need, and the one which power efficient.

[Humbug]

John Bridgman, one of AMD's Linux driver developers made a couple of comments on RDNA at the phoronix forums, I cannot really interpret these so will just post them here.
 

Quote

we used to talk about GCN as an ISA, but it seems that most people outside AMD think of GCN as a micro-architecture instead (ie an implementation of the ISA). RDNA is GCN ISA but not what you think of as GCN architecture.

 

With ISA you mean Instruction Set Architecture?

 

Right... programming model would be another term.

R3xx-R5xx was separate PS/VS with vector+scalar instructions, R6xx was unified shader VLIW SIMD, GCN is unified shader SIMD+scalar.

There seemed to be a lot of confusion between ISA and architecture, leading to unintentional forum comments that in the CPU world would translate to something like "Ryzen is never going to be much better than an 8086 because they are both x86".

 

Can you delve into this a bit deeper? There's very little actual information about Navi.

 

Sorry, but not yet. I think the next level of detail will come out around E3.

 

https://www.phoronix.com/forums/forum/phoronix/latest-phoronix-articles/1103202-amd-is-aiming-for-radeon-rx-5700-navi-support-in-linux-5-3-mesa-19-2

[RejZoR]

He's basically mentioning what I've complained about when people always started with "but muh GCN" when it came to modern Radeon cards...

[Trixanity]

14 minutes ago, RejZoR said:

He's basically mentioning what I've complained about when people always started with "but muh GCN" when it came to modern Radeon cards...

Both yes and no. We know the ISA hasn't changed since Tahiti but we also know the uarch couldn't keep up since 2014 and it being largely iterative from the start with a few novel ideas. So it's definitely a valid complaint. People use the words AMD use in their marketing and they've referred to it as GCN + generation for years. They've only really stopped using it less than a week ago.  There's probably a very good reason for that and it isn't to get people to stop talking about "muh GCN". 

[RejZoR]

I think it is the exact reason AMD is dropping it. Because GCN now has negativity attached even if there is nothing wrong with it. Card could be a stellar performer and if GCN was mentioned anywhere, people would have doubts literally because of "muh GCN" syndrome.

[Trixanity]

3 minutes ago, RejZoR said:

I think it is the exact reason AMD is dropping it. Because GCN now has negativity attached even if there is nothing wrong with it. Card could be a stellar performer and if GCN was mentioned anywhere, people would have doubts literally because of "muh GCN" syndrome.

Very doubtful. It's more likely that they've actually reworked the uarch considerably. They used to update 2-3 blocks per iteration and now they've finally found the time and money to do something different to get out from under the crushing weight of a uarch that reached its limits half a decade ago.

 

It remains to be seen if this RDNA thing can actually do something. Both short term and long term.

[Humbug]

13 minutes ago, Trixanity said:

Very doubtful. It's more likely that they've actually reworked the uarch considerably. They used to update 2-3 blocks per iteration and now they've finally found the time and money to do something different to get out from under the crushing weight of a uarch that reached its limits half a decade ago.

 

It remains to be seen if this RDNA thing can actually do something. Both short term and long term.

Ya, that's where I stand on it atm. I think they have put significant engineering resources into reworking it.

 

Now whether it's any good both in the form of Navi and other future RDNA iterations remains to be seen. 

[Stefan Payne]

37 minutes ago, Trixanity said:

Both yes and no. We know the ISA hasn't changed since Tahiti but we also know the uarch couldn't keep up since 2014 and it being largely iterative from the start with a few novel ideas. So it's definitely a valid complaint. 

So you're saying that Radeon HD5870, 6870 or maybe even the 6970 are the same as Radeon HD2900, 3850 because both are Terrascale, based on the VLIW Architecture??

[Trixanity]

32 minutes ago, Stefan Payne said:

So you're saying that Radeon HD5870, 6870 or maybe even the 6970 are the same as Radeon HD2900, 3850 because both are Terrascale, based on the VLIW Architecture??

[Stefan Payne]

Just showing what AMD means with GCN, as the models I listed are all what AMD calles "TeraScale".

HD2000 Series and 6000 series are however reallly different..

AMD even had a VLIW4 Chip instead the usual VLIW5 ones (only 6900 series and Trinity + Richland APUs)...

 

With "GCN" all they mean is the way the CUs work, nothing more...

[Trixanity]

3 minutes ago, Stefan Payne said:

Just showing what AMD means with GCN, as the models I listed are all what AMD calles "TeraScale".

HD2000 Series and 6000 series are however reallly different..

AMD even had a VLIW4 Chip instead the usual VLIW5 ones (only 6900 series and Trinity + Richland APUs)...

 

With "GCN" all they mean is the way the CUs work, nothing more...

You can't retcon this shit. Certainly not that way.

[Firewrath9]

19 hours ago, CarlBar said:

 

 There's a reason thats a 10k card, (beyond the usual reasons Quadro's are expensive). 4GB GDDR6 chips are really expensive, i don't know anything like exact numbers but i wouldn't be shocked to find that it's on par with HBM pricing. It's also the current upper limit. Whilst the cost is going to come down and the capacities likely will increase some more it's really not a good situation for GDDR6 as the top end cards keep pushing the limits of what it can do whilst HBM based cards are nowhere near the limits, (thats at least 128Gb of capacity and 3.2TB of bandwidth).

well, first of all, it uses 16Gb (2GB GDDR6) chips, (just 2 of them for every slot). It's basically a RTX Titan with an extra set of 2GB chips and Quadro Drivers.

RTX Titan is 2500$, so at MOST the price of doubling the VRAM is 2500$ (ripping off the GDDR6 on a different titan and putting it on a Titan + Driver = Titan RTX)

Most likely its ~200$, as 12 GB of 8Gb ddr6 is ~50-80$

 

HBM is fine for datacenter usage, but it does not belong on mainstream cards.

Also, engineering problems due to differing heights, which is why several Vegas had TJunction issues

[Firewrath9]

23 hours ago, RejZoR said:

It's funny though, we make things at 7nm scale, but have problems aligning two big ass chips measured in millimeters to the same height on the PCB... XD

they would have to adjust the substrate, which I don't think is possible.

[Firewrath9]

On 5/31/2019 at 10:26 AM, TheDankKoosh said:

No it isn't, most pascal cards have 8ghz memory stock, not 9ghz.

sigh

 

+500mhz on base memory clock is 2500mhz, which is multiplied by 4 to get the effective clock, 10ghz.

Can you do some research on things before saying them?

[TheDankKoosh]

25 minutes ago, Firewrath9 said:

sigh

 

+500mhz on base memory clock is 2500mhz, which is multiplied by 4 to get the effective clock, 10ghz.

Can you do some research on things before saying them?

ddr means DOUBLE data rate, that doesn't change just because it's gddr. Just because amd calculates it in that way does not mean that it's the true way of calculating it.