AMD Readies Larger 7nm "Navi 12" Silicon to Power Radeon RX 5800 Series?

[AnarchoPrimitiv]

300w incoming tho, I suspect, wich scares me

Why? You live in the remotest area of a third world country? The only place on earth where electricity is still expensive? In America the average is 12.5 cents per kwhr, meaning the difference between 250w and 300 is less that $50 for a year of constant use

 

[kapone32]

For all the people complaining about new releases think how owners of Vega 64s felt when Navi was released. In (unless mining takes off again) the next 6 months you should be able to get a 5700XT for around $300. Or used ones on EBay when the AIB cards are released in a couple weeks.

 

[ppn]

Navi 12 is RX 5600, Navi 14 is RX 5500, there is no big GPU.

 

[Shatun_Bear]

An overclocked reference 5700 XT is already not that far away from a 2080 so a version with higher CUs/more stream processors with a 150Mhz+ bump to core clock and possibly with addition of HBM2 would blow past that and land in between a 2080 and 2080 Ti.

 

[ratirt]

For all the people complaining about new releases think how owners of Vega 64s felt when Navi was released. In (unless mining takes off again) the next 6 months you should be able to get a 5700XT for around $300. Or used ones on EBay when the AIB cards are released in a couple weeks.

Do you know what it means a new GPU release? What is the difference between new GPU and boost in core number? How can you compare Vega to Navi when these are two different architectures? It's not what it is between non-super and super RTX's. These are the same chips. Turing.
BTW. I'm Vega 64 owner and I don't feel bad with Navi release. It is not Vega chip with a boost. Although, I'm sure there are people here that will disagree. For them GCN and RDNA are the same.

 

[ToxicTaZ]

An overclocked reference 5700 XT is already not that far away from a 2080 so a version with higher CUs/more stream processors with a 150Mhz+ bump to core clock and possibly with addition of HBM2 would blow past that and land in between a 2080 and 2080 Ti.

RTX 2080 woops Radeon 7 across the board then there's the RTX 2070 Super with the same performance as Radeon 7 but at a lower price point thus Killing the Radeon 7.

Radeon 7 woops the 5700 XT across the board.

RTX 2070 Super woops 5700 XT now means 5800/5800 XT are going after RTX 2080 and 2080 Super performance levels.

Leaving RTX 2080Ti for Navi 20

 

[ratirt]

Radeon 7 woops the 5700 XT across the board.

Well I wouldn't say it whoops 5700XT but is faster and not across the board

RTX 2070 Super woops 5700 XT

No it doesn't. In some games 5700XT is faster. So how can you say 2070S whoops 5700XT?

 

[Totally]

Historically, AMD's biggest dies have had codenames in multiples of 10; Polaris 10, 20, 30, Vega 10, 20 are all the largest dies of those particular architectures/generations.

Polaris 11, 12, and 21 were smaller dies, cut-down from 2304SP of the full Polaris chip to 1024 and 640SP

I don't think Navi 12 will be a larger chip - The halo market is vanishingly small and expensive to operate in. AMD are far more likely to be chasing the mass-market where 90% of sales are with a lower-TDP, cheaper-to-make chip that serves the $200 and lower market, as well as being laptop-friendly. I'm only guessing, but I think that *if* AMD make a larger Navi die, it'll be called Navi 20, or given an new codename altogether.

Historically? There was never a big Polaris was a small die that was cut down smaller with successive versions. Vega was a similar story with Vega it was the big chip. The codenames themselves demarcate big/little dies. Navi 12 is highly likely the 5400, 5500, 5600.

 

[Shatun_Bear]

RTX 2080 woops Radeon 7 across the board then there's the RTX 2070 Super with the same performance as Radeon 7 but at a lower price point thus Killing the Radeon 7.

Radeon 7 woops the 5700 XT across the board.

RTX 2070 Super woops 5700 XT now means 5800/5800 XT are going after RTX 2080 and 2080 Super performance levels.

Leaving RTX 2080Ti for Navi 20

I meant 'heavily overclocked'. A 5700XT OC to 2000Mhz is faster than Radeon 7 and the great thing is the XT only has 36 CUs. Unless it scales up disastrously, a 64 CU Navi with a high enough boost is going to be faster than a 2080 and 2080 Super (which is just 5% faster than a standard 2080).

 

[Chrispy_]

Historically? There was never a big Polaris was a small die that was cut down smaller with successive versions. Vega was a similar story with Vega it was the big chip. The codenames themselves demarcate big/little dies. Navi 12 is highly likely the 5400, 5500, 5600.

So you're agreeing with another of my posts that the codename would need to be different (not Navi) for it to be a bigger chip and that Navi 12 is just chopped-down (smaller) Navi 10 for lower-end cards?

Small GPUs in laptops are dead, the iGPU killed them. The only laptop GPUs that sell well are pro GPUs or high end gaming GPUs, both of which are dominated by nvidia. AMD needs a big laptop chip, not tiny ones.

I dunno, GTX 1050, 1050Ti laptops are STILL crazy popular and the reason Max-Q exists is because high-TDP chips make laptops bulky, noisy, or both. 1650 laptops are selling like hotcakes too, despite being relatively new to market.

Not everyone wants to spend well over $2000 on a non-upgradeable RTX gaming laptop that can't even match a $500 desktop GPU.

 

[ValenOne]

I meant 'heavily overclocked'. A 5700XT OC to 2000Mhz is faster than Radeon 7 and the great thing is the XT only has 36 CUs. Unless it scales up disastrously, a 64 CU Navi with a high enough boost is going to be faster than a 2080 and 2080 Super (which is just 5% faster than a standard 2080).

RX 5700 XT has 40 CU.

CU scaling is not enough without scaling towards 6 prim units and 96 ROPS. Hint; Titan RTX has six GPCs, 96 ROPS and 6 MB L2 cache.

RX 5700 XT has 40 CU, four prim units, 64 ROPS and 4 MB L2 cache. TU106 has four GPCs, 64 ROPS and 4 MB L2 cache, hence similar to NAVI 10 XT.

Historically? There was never a big Polaris was a small die that was cut down smaller with successive versions. Vega was a similar story with Vega it was the big chip. The codenames themselves demarcate big/little dies. Navi 12 is highly likely the 5400, 5500, 5600.

Largest Polaris baseline IP is Xbox One X's 44 CU scaling with 384 bit bus.

 

[Totally]

So you're agreeing with another of my posts that the codename would need to be different (not Navi) for it to be a bigger chip and that Navi 12 is just chopped-down (smaller) Navi 10 for lower-end cards?

Yes, I was pointing out that you were correct but your reasoning wasn't. That said I don't understand how such a conclusion could be reached by the author when AMD has yet to release a complete lineup, and they usually work their way down the stack after a release, then plug the holes if/when practical.

 

[Juankato1987]

I guess Manu is scared about 300W TDP, on non overcloked cards.
That could leave to 350W on thrid party cards with more headroom for overclocking.
RTX 2080 Ti is a 250W TDP card and its performance is 34% higher tha RX 5700XT
RX 5700XT is a 225W TDP card, and if we put some maths (134%) to get a higher TDP
we end up with more than 300W, even if perfomance could be linear, wich isn't,
RX 5800 has to become a 300W TDP card to compete with RTX 2080 Ti.


The table below shows the consumption peak, which is rarely given in real life,
and is only for academic purposes.

Why Manu, why are you selectively scared ?

 

[Dbiggs9]

Lisa Su just confirmed high end Navi is Ramping

 

[Juankato1987]

I thinking about a possibility of 2x Navi 10 chips conected with InfiniyFabric.
Would that be possible?

 

[Chrispy_]

Yes, I was pointing out that you were correct but your reasoning wasn't.

You give me far too much credit if you think I was reasoning. Spitballing, maybe

 

[medi01]

Navi 12 is RX 5600, Navi 14 is RX 5500, there is no big GPU.

250mm2 chip is surely max AMD could do, as else it would threaten 2080+.

Lisa Su just confirmed high end Navi is Ramping

Where?

 

[Alpha_Lyrae]

That sounds very optimistic. On technical side of things resource allocation for RX5700 is identical to RTX2070 (36 CU/GPC) and RX5700XT is identical to RTX2070S (40 CU/GPC). These are not favourable comparisons to Navi, the strength for Navi is basically lower price and going agaist less-equipped competition. RTX2080Ti is a 68 GPC GPU. 64 CU Navi will clearly have trouble competing. 46/48 GPC RTX2080/2080S are more realistic targets.

For a realistic estimate look at RTX cards - their efficiency is largely identical at the same performance level. Either way, high-end card will inevitably take the 250W TDP.

RTX2080Ti MSRP is 999$ and current street prices at least in Europe are already below that.

I think what you'll see from AMD is a shift to more shader engines or even a Graphics Compute Complex where groups of 4 shader engines are paired together and selectively cut down from there. With the new primitive unit (and primitive shaders), rasterizer, and ROPs tightly integrated and a single geometry processor delegating and load balancing work to said primitive units, I think RDNA does not have the 4 shader engine limit we've seen in GCN.

The fixed function geometry hardware is all gone. Vertex, geometry, and tessellation stages are all combined into the primitive shader and those stages are emulated in older games for compatibility (which AMD says is still faster than previous Vega's dedicated hardware ... oof!). I believe the geometry processor can support any number of primitive units through virtualization of work queues, a bit like the HWS does for ACEs. Not much has been divulged though, so take that with a boulder of salt.

So, with a theoretical maximum of 8 shader engines with 10 CUs each, that's a total of 80 CUs/5120SPs or double Navi 10's 40/2560SPs.

That's obviously overkill, so that would probably only replace Vega 20 (if paired with HBM3) for compute and compete with Nvidia's top end Quadros, maybe in 2020 or later.

But, I think it's very possible we could see 72CU and lower configurations with only 6 shader engines, much like TU102 and TU104 have 6 GPCs with varying SMs per GPC. TU102 has 12 SMs per GPC (72 max), while TU104 has 8 (48 max), both in full die configuration. TU102-300 in RTX 2080Ti has 4 SMs, one 32-bit MC with 512KB L2, and 8 ROPs disabled. So, it's 68 SMs rather than GPCs. GPCs are the complete processing cluster units.

 

[ToxicTaZ]

I think what you'll see from AMD is a shift to more shader engines or even a Graphics Compute Complex where groups of 4 shader engines are paired together and selectively cut down from there. With the new primitive unit (and primitive shaders), rasterizer, and ROPs tightly integrated and a single geometry processor delegating and load balancing work to said primitive units, I think RDNA does not have the 4 shader engine limit we've seen in GCN.

The fixed function geometry hardware is all gone. Vertex, geometry, and tessellation stages are all combined into the primitive shader and those stages are emulated in older games for compatibility (which AMD says is still faster than previous Vega's dedicated hardware ... oof!). I believe the geometry processor can support any number of primitive units through virtualization of work queues, a bit like the HWS does for ACEs. Not much has been divulged though, so take that with a boulder of salt.

So, with a theoretical maximum of 8 shader engines with 10 CUs each, that's a total of 80 CUs/5120SPs or double Navi 10's 40/2560SPs.

That's obviously overkill, so that would probably only replace Vega 20 (if paired with HBM3) for compute and compete with Nvidia's top end Quadros, maybe in 2020 or later.

But, I think it's very possible we could see 72CU and lower configurations with only 6 shader engines, much like TU102 and TU104 have 6 GPCs with varying SMs per GPC. TU102 has 12 SMs per GPC (72 max), while TU104 has 8 (48 max), both in full die configuration. TU102-300 in RTX 2080Ti has 4 SMs, one 32-bit MC with 512KB L2, and 8 ROPs disabled. So, it's 68 SMs rather than GPCs. GPCs are the complete processing cluster units.

Is that why the TU104 is faster than Navi 10 hmm interesting. Also TU104 is more complex with 13+ Billion Transistors vs Navi is around 10+ Billion Transistors.

 

[ValenOne]

I think what you'll see from AMD is a shift to more shader engines or even a Graphics Compute Complex where groups of 4 shader engines are paired together and selectively cut down from there. With the new primitive unit (and primitive shaders), rasterizer, and ROPs tightly integrated and a single geometry processor delegating and load balancing work to said primitive units, I think RDNA does not have the 4 shader engine limit we've seen in GCN.

The fixed function geometry hardware is all gone. Vertex, geometry, and tessellation stages are all combined into the primitive shader and those stages are emulated in older games for compatibility (which AMD says is still faster than previous Vega's dedicated hardware ... oof!). I believe the geometry processor can support any number of primitive units through virtualization of work queues, a bit like the HWS does for ACEs. Not much has been divulged though, so take that with a boulder of salt.

So, with a theoretical maximum of 8 shader engines with 10 CUs each, that's a total of 80 CUs/5120SPs or double Navi 10's 40/2560SPs.

That's obviously overkill, so that would probably only replace Vega 20 (if paired with HBM3) for compute and compete with Nvidia's top end Quadros, maybe in 2020 or later.

But, I think it's very possible we could see 72CU and lower configurations with only 6 shader engines, much like TU102 and TU104 have 6 GPCs with varying SMs per GPC. TU102 has 12 SMs per GPC (72 max), while TU104 has 8 (48 max), both in full die configuration. TU102-300 in RTX 2080Ti has 4 SMs, one 32-bit MC with 512KB L2, and 8 ROPs disabled. So, it's 68 SMs rather than GPCs. GPCs are the complete processing cluster units.

NAVI 10 has two shader engines with four prim (triangle) units, 40 CU, 64 ROPS, 4MB L2 cache coupled with 256 bit bus.

7870 has two shader engines with two geometry units, 20 CU, 32 ROPS, 512KB L2 cache coupled with 256 bit bus.
R9-290X has four shader engines with four geometry units, 44 CU, 64 ROPS, 1MB L2 cache coupled with 512 bit bus. R9-290X is 2X scaled 7870.

Your argument for " theoretical maximum of 8 shader engines with 10 CUs each" didn't factor in memory controller relationship!

Each NAVI shader engine has two prim units, 20 CU, 32 ROPS, 2MB L2 cache and 128 bit bus.

 

[londiste]

Is that why the TU104 is faster than Navi 10 hmm interesting. Also TU104 is more complex with 13+ Billion Transistors vs Navi is around 10+ Billion Transistors.

TU104 is 48 SM chip, Navi10 is 40 CU chip. TU104 has 20% more resources (everything except memory bus width).

 

[ValenOne]

TU104 is 48 SM chip, Navi10 is 40 CU chip. TU104 has 20% more resources (everything except memory bus width).

TU104 has six GPCs from TU102 with TU104's 64 ROPS and 4MB L2 cache, hence it's half way house between TU102 and TU106.

TU104 has similar geometry power like TU102 with TU106's read/write ROPS and 4MB L2 cache.

 

[Alpha_Lyrae]

View attachment 128132View attachment 128132


Is that why the TU104 is faster than Navi 10 hmm interesting. Also TU104 is more complex with 13+ Billion Transistors vs Navi is around 10+ Billion Transistors.

Don't forget that TU104 has extra hardware too. Tensor cores and BVH intersection/traversal accelerators (RT cores). The 2 extra rasterizers can help in certain situations, but there is a point of diminishing returns. It's best shown off in RTX 2080Ti or even Titan where you have enough shader cores to shade polygons faster and fully utilize the extra rasterization capabilities. I don't know how much of 2070 Super's performance is down to extra raster engines or simply Turing's architectural efficiency (compressing any data that is compressible to save cache/bandwidth and keeping data on-chip with direct read/write of assets without decompression).

AMD's only chance against Nvidia is to move to more shader engines. That way they don't have to push the clocks outside of 7nm's efficiency range (2.1GHz OC destroyed efficiency of RDNA) and they'll gain the same rasterization capabilities as TU102/104. It'll help when AMD themselves move to RT capable hardware, which is expected in RDNA 2.

I'm curious to see what Nvidia does with Ampere. I'm thinking they'll try to at least double ray tracing performance or more. Not sure how much die space that'll take on Samsung's 7nm EUV. All that means is that these higher GPU prices are here to stay. =\

 

[Ran]

Noup - that is a false rumour. Navi 10 high end, navi 12 mid range navi 14 low range. Navi 20 (in 2020) will be the one fighting 2080 series cards.

For all the people complaining about new releases think how owners of Vega 64s felt when Navi was released. In (unless mining takes off again) the next 6 months you should be able to get a 5700XT for around $300. Or used ones on EBay when the AIB cards are released in a couple weeks.

There is always the "wait and you get better", but if you do that you never get anything. You need to decide when you need a new GPU and then buy it and be happy with what you have. The meaning of life is not to always have the latest tech. I run AMD 580 and it runs the games I play just fine, no need of a VEGA or NAVI atm. There is some hystery to compare framerates, if WOT runs with RX 580 at 170-210 fps now, and it would do 240 with VEGA.... so what???
Mining still runs fine, the value for ethereum went down, but has recovered a little, on the long term it might give some profit. The uncertainty lies in the possibility that governments might want to control crypto currencies, and when someone suggests that the value fall, at least temporarily.

 

[crotach]

I'm sure nvidia will come out with a new SUPER DUPER series to compete with AMD.