AMD Radeon RX 6000 Series Specs Leak: RX 6900 XT, RX 6800 XT, RX 6700 Series

[Punkenjoy]

It's wrong to compare CPU cache to GPU cache. Yes both are catch, but the workload is so different that you can't make comparaison on performance.

Also here, it's not an increase or reduction in cache, it's the ability to lookup others L1 caches of others gpu cores to see if the information they need is there.

GPU do many iteration on a image, on texture and etc and they work on the same stuff most of the time. This is why there is a real potential of gain and bandwidth saving (Yes cache save bandwidth... if you have a cache hit, you don't need to read it...).

CPU on the other hands do all kind of work and handle many different process. They process all kind of data and getting cache hit is harder because of that.

 

[bug]

It's wrong to compare CPU cache to GPU cache. Yes both are catch, but the workload is so different that you can't make comparaison on performance.

Also here, it's not an increase or reduction in cache, it's the ability to lookup others L1 caches of others gpu cores to see if the information they need is there.

GPU do many iteration on a image, on texture and etc and they work on the same stuff most of the time. This is why there is a real potential of gain and bandwidth saving (Yes cache save bandwidth... if you have a cache hit, you don't need to read it...).

CPU on the other hands do all kind of work and handle many different process. They process all kind of data and getting cache hit is harder because of that.

I see. So basically everyone was stupid till now adding bandwidth, because with a little bit of cache added they could have solved the problem. It all makes sense now.

Edit: Do we even know this is a data cache and not an instruction cache?

 

[Vya Domus]

It's wrong to compare CPU cache to GPU cache. Yes both are catch, but the workload is so different that you can't make comparaison on performance.

Also here, it's not an increase or reduction in cache, it's the ability to lookup others L1 caches of others gpu cores to see if the information they need is there.

GPU do many iteration on a image, on texture and etc and they work on the same stuff most of the time. This is why there is a real potential of gain and bandwidth saving (Yes cache save bandwidth... if you have a cache hit, you don't need to read it...).

CPU on the other hands do all kind of work and handle many different process. They process all kind of data and getting cache hit is harder because of that.

In summary graphics workloads provide both spatial and temporal locality for data and instructions almost all of the time. Anyway it's not just L1 cache lookup, the chips themselves will probably have a lot of L2 cache as well.

 

[bug]

In summary graphics workloads provide both spatial and temporal locality for data and instructions almost all of the time. Anyway it's not just L1 cache lookup, the chips themselves will probably have a lot of L2 cache as well.

The post quoting from the patent applications says it's 0.09 sq mm / core. I wouldn't bet on "a lot of L2 cache" fitting in there.

I don't doubt AMD has done something smart here, but you guys are simply expecting too much from a L1 cache. The cache my be more effective than it is on a CPU, but bridging hundreds of GB/s of bandwidth difference? I have to see it to believe it.

 

[Punkenjoy]

The post quoting from the patent applications says it's 0.09 sq mm / core. I wouldn't bet on "a lot of L2 cache" fitting in there.

I don't doubt AMD has done something smart here, but you guys are simply expecting too much from a L1 cache. The cache my be more effective than it is on a CPU, but bridging hundreds of GB/s of bandwidth difference? I have to see it to believe it.

Indeed, but it's only a portion of the final solution. Both GPU have all kind of memory tricks like compression to reduce memory bandwidth. People focus on that because it sound cool and it's what they know. Also we don't necessarily know if Ampere is really memory bandwidth bound.

NAVI10 currently have 512 kb of L1 (128 kb per 10 CU block). if AMD keep the same size per block of 10 CU, the available L1 cache will go from 128 kb to 1 MB for full NAVI21 witch is quite a big increase. If they match the cache size Per CU, they could have around 8 MB of L2 cache.

how much that will help, i don't know. No matter what people say, only the benchmark will really reveal the final information. We may be discussing about stuff that have marginal impact on the real performance.

 

[medi01]

Had Nvidia been able to use TSMC I think the supply issues would be considerably less.

Yeah.
Or had NV priced card that is 20-30% faster than $1200 card at around $999.
I wonder why didn't they do it.
Oh wait, this can't be the reason, can it:

https://www.reddit.com/r/Amd/comments/jg21ai

Can someone explain the difference between the 8nm and 7nm? Does that actually make a difference in performance or anything important?

Neither honestly describes what the process really is (e.g. 7nm TSMC means 22x22nm transistor, while 14nm Intel means 24x24nm), but Samsung's 8nm is even further from truth, so there is that.

 

[dalekdukesboy]

I really like the "feel" of this launch. They didn't do anything too crazy or exotic with the cards and 256 bit wide bus obviously easier to manufacture and get better yields working properly than the Nvidia approach. The really high clock speeds make me wonder if they had to do it to compete with Nvidia or they just have an exceptional product that they squeezed to the max and maybe even outperforms some of the equivalent Nvidia stack...however I will say as someone mentioned the launch CAN'T be worst than Nvidia's 3080 launch...if they get 10 cards out in week 1 they probably did better .

 

[yeeeeman]

Nvidia will have a bit of trouble with these new cards.

 

[Punkenjoy]

if you look at OC Navi10 cards, they already do 2 GHz game clock. So the "official" numbers for all NAVI21 cards aren't that far away from the current NAVI10 clock.

We could say AMD want to push it to compete with Nvidia, but on the other end, why would they limit the frequency if people are able to OC these cards with custom bios and stuff? They better just sell them with clock as fast as they can do.

 

[Chrispy_]

if you look at OC Navi10 cards, they already do 2 GHz game clock. So the "official" numbers for all NAVI21 cards aren't that far away from the current NAVI10 clock.

Navi21 is on N7P that's 7% faster (according to TSMC) at the same power level.

So Navi10's 2GHz game clock becomes 2.15GHz without even changing the TDP or accounting for any architectural improvements that always get made to improve clockspeeds between generations.

 

[MxPhenom 216]

Yeah.
Or had NV priced card that is 20-30% faster than $1200 card at around $999.
I wonder why didn't they do it.
Oh wait, this can't be the reason, can it:

https://www.reddit.com/r/Amd/comments/jg21ai

Neither honestly describes what the process really is (e.g. 7nm TSMC means 22x22nm transistor, while 14nm Intel means 24x24nm), but Samsung's 8nm is even further from truth, so there is that.

So then we are assuming that either the 6800XT is slower by a tad and Nvidia knows this, but will be a lot cheaper or what?

If its faster, best believe AMD will price it accordingly.

 

[medi01]

So then we are assuming that either the 6800XT is slower by a tad and Nvidia knows this, but will be a lot cheaper or what?

AMD has a very competitive lineup inbound.
Pricing on NV cards is BS, street price on 3070 is around 700 Euros, 3080 is missing in action on top of that.

NV won't be able to recover with Ampere on Samsung 8nm, it an un-redeemable clusterf*ck.

 

[MxPhenom 216]

AMD has a very competitive lineup inbound.
Pricing on NV cards is BS, street price on 3070 is around 700 Euros, 3080 is missing in action on top of that.

NV won't be able to recover with Ampere on Samsung 8nm, it an un-redeemable clusterf*ck.

They definitely do. Been a long time coming really. AMD hasnt been very competitive in the high end GPU space for years now. probably since the R9-290x.

How do you know that when the 3070 isnt even out yet?

Nvidia will probably counter next year with Super and Ti variants using TSMC 7nm.

 

[N3M3515]

Nvidia will probably counter next year with Super and Ti variants using TSMC 7nm.

I don't know about the 3070 super, but 3080 super? when 3090 is at spiting distance(12% more less)....
Maybe a 3080 super with 20GB and 6% better than regular 3080

 

[MxPhenom 216]

I don't know about the 3070 super, but 3080 super? when 3090 is at spiting distance(12% more less)....
Maybe a 3080 super with 20GB and 6% better than regular 3080

Id suspect itll mostly just be a process change, higher clocks if they can and more memory. No cores unlocked, etc.

That 3090 is a fucking joke for gamers.

 

[Max(IT)]

But they're stable in the long run. I've got an RX 470 and 580 in the house in lesser machines and they run totally smooth, i keep forgetting they're not Nvidia and try to open the overlay with the NV shortcut of alt-Z to stream things.

unfortunately 5700XT are still not stable today, unlike the RX580 series...

 

[MxPhenom 216]

unfortunately 5700XT are still not stable today, unlike the RX580 series...

What do you mean? Drivers still a nightmare?

 

[Max(IT)]

AMD has a very competitive lineup inbound.
Pricing on NV cards is BS, street price on 3070 is around 700 Euros, 3080 is missing in action on top of that.

NV won't be able to recover with Ampere on Samsung 8nm, it an un-redeemable clusterf*ck.

There is no street price for the 3070, since the 3070 isn’t yet “on the street “.

What do you mean? Drivers still a nightmare?

Yep.
I wouldn’t call it a nightmare (it was until April) but I still have some freezes and black screens, from time to time.

E

Nvidia will have a bit of trouble with these new cards.

especially if these new cards will actually be available (and I’m not entirely sure about that).

 

[MxPhenom 216]

There is no street price for the 3070, since the 3070 isn’t yet “on the street “.


Yep.
I wouldn’t call it a nightmare (it was until April) but I still have some freezes and black screens, from time to time.

E

especially if these new cards will actually be available (and I’m not entirely sure about that).

Yeah so that is what i've been worried about if the 6800XT is enticing enough. In the past I have had issues with AMD drivers, but the last time i had one of their cards was a 5870.

 

[Max(IT)]

Yeah so that is what i've been worried about if the 6800XT is enticing enough. In the past I have had issues with AMD drivers, but the last time i had one of their cards was a 5870.

I’m considering a 6800XT too, but I know I’m taking a risk.
AMD is great at hardware production, but not in the software development

 

[Umbral]

Why would AMD sell a card basically the same as the 3090 for less than half the price? The delusion is real

slower than 3090, 16 GB vs 24 GB, closer to 3080 than 3090.

Simple.

 

[c1979h4life]

I wonder if the 6700 XT will perform around or a bit faster than the 5700 XT? Logic dictates it will be faster than a 5700 XT but who knows how much faster and if it will actually be faster.

You still are waiting for the 3080? oh boy. The Navi 6000 will definitely come before you get the card so you will have your chance of comparison. I'm also aiming at the more less 3080 type of performance and waiting for new Navi to see what it will offer. Either way I need a new faster GPU.

The RDNA2 GPUs were touted by AMD as being up to 50% more powerful than last gen, so I at least think it will be between 30-40% faster than the 5700XT, which would probably put it in the 500-600 dollar range. They went up on the new 5xxx series CPUS also. I want to see what the 6800 base model does.

 

[bug]

The RDNA2 GPUs were touted by AMD as being up to 50% more powerful than last gen, so I at least think it will be between 30-40% faster than the 5700XT, which would probably put it in the 500-600 dollar range. They went up on the new 5xxx series CPUS also. I want to see what the 6800 base model does.

Remember, 5700 XT features a 256 bit memory bus, 5600(XT) supposedly cuts that back to 192 bit. That means a cheaper PCB and we'll have to wait and see what it means for performance.

 

[Zach_01]

Indeed, but it's only a portion of the final solution. Both GPU have all kind of memory tricks like compression to reduce memory bandwidth. People focus on that because it sound cool and it's what they know. Also we don't necessarily know if Ampere is really memory bandwidth bound.

NAVI10 currently have 512 kb of L1 (128 kb per 10 CU block). if AMD keep the same size per block of 10 CU, the available L1 cache will go from 128 kb to 1 MB for full NAVI21 witch is quite a big increase. If they match the cache size Per CU, they could have around 8 MB of L2 cache.

how much that will help, i don't know. No matter what people say, only the benchmark will really reveal the final information. We may be discussing about stuff that have marginal impact on the real performance.

8MB? No... no...
Leaks are "saying" about 128MB total cache in RNDA2 flagship at least. We cant really confirm it though.

Rough estimation
If we accept the biggest Navi21 die is 536mm² then:

5700XT (40CU + SOC) = 251mm²
From pictures of the die about 150mm² is the 40CUs and the other 100mm² for SOC stuff.
80CUs = 300mm², but with enhanced 7NP node (+10~15% density) this could result a 250~270mm² area for 80CUs + 100~150mm² for SOC = 80CU + SOC = 350~420mm².
This leaving another 115~185mm² up to 536 for cache only.

Remember that Navi21 has the same 256bit bandwidth so it does not need a more complex mem cotroller (no extra space on die from Navi10).

So the 100+MB cache can be a thing on the Navi21 die...

 

[r9]

Mhmm.

From what I understand, GDDR6X is in very short supply. Only Micron is making it right now. Also Samsung 8nm (10nm+) yields are pretty bad.

Or Nvidia is purposefully reducing the supply. If thats the case, itll just hurt Nvidia in the end since most people will just buy these new AMD cards if the performance is there.

They just can't produce enough, low gpu and ddr6x yields. Hopefully they get on top of it and so does AMD with

Fixed that for you, how come Turing was priced the same everywhere (except where taxes are the issue), but Ampere costs up to 40% outside the US.

If they released for $1099 the extra money would have went in their pocket and not in a bunch of vultures pockets. Without AMD for $1099, $100 cheaper you get 30% perf boost so it's not totally outrageous.
I'm not trying to defend nvidia, but the situation they are in it's not intentional just a result of some bad calls like manufacturing and ddr6x. I'm sure if it was up to them they would like to be able to make and sell more. One for obvious reason $$$ and second the more they sell now the less market is left for AMD. But I highly doubt that AMD availability will be much better. And to remind those who didn't know amd and nvidia have been sued for having agreement between and price fixing. So I don't expect any price undercutting or anything, the most we can hope for is buy at regular price.