Question Speculation: RDNA2 + CDNA Architectures thread

[uzzi38]

All die sizes are within 5mm^2. The poster here has been right on some things in the past afaik, and to his credit was the first to saying 505mm^2 for Navi21, which other people have backed up. Even still though, take the following with a pich of salt.

Navi21 - 505mm^2

Navi22 - 340mm^2

Navi23 - 240mm^2

Source is the following post: https://www.ptt.cc/bbs/PC_Shopping/M.1588075782.A.C1E.html

 

[moinmoin]

Of course they won't show everything, but if they are going to bring a tease, then I am sure they brought out the best numbers they can bring out. I am very sure if they had better than 3080 numbers to show, they would have done it.

And turn the upcoming event into a lame boring showcase of everything that isn't as good as the best numbers they already handpicked as a tease for it? Please.

 

[HurleyBird]

Official specs.

https://www.amd.com/en/products/specifications/compare/graphics/10516%2C10521%2C10526

 

[Mopetar]

What AMD need is DLSS alternative.

I'm one of those types that's generally against the way this type of technology is being used to sell gamers on false numbers to cover for a lack of RT capabilities, so I really don't get the argument.

Part of me thinks that if Nvidia developed a technology that would stab you in the eyes while running one of their cards there would still be people lining up demanding that AMD also implement their own eye-stabbing solution in their cards.

Get better RT hardware so that upscaling isn't necessary. Or just play at a lower resolution. If you show an eagerness to purchase deceit, don't act surprised when you get lied to a lot more in the future.

 

[.vodka]

2080S?

With all the perf/clk improvements RDNA2 has (+21%), plus that ridiculous frequency (~2850MHz gfx clk) the 6700XT should probably be at ~2080Ti/3070 level at 1080p, but probably not quite make it at 1440p. Still closer to them than the 2080S I guess.

It all comes down to how much IC AMD is going to use on Navi22. AMD themselves provided the cache hit rates for 128, 96, 64, 48, 32MB etc of cache in one of their slides. 96MB should make for a quite solid 1440p card, 64MB would be a bit limited, maybe, for that resolution considering there's a 192 bit bus behind that cache.

Considering there's a few less memory controllers apart from having half the resources of Navi21 (~510mm²), I'd guess they didn't skimp on the ~30mm² it'd require to implement the extra 32MB to do 96MB. Navi10 is ~250mm², Navi22 should be around that die size after all the cutting down.

It'd be ridiculous to see a little 250mm² die mostly match up against a 754mm² x02 monster and the typical ~400mm² X04 die, lol

If they go for 96MB cache, it should be an excellent high refresh rate 1080p card that can comfortably do not quite as high refresh rate 1440p gaming.

64MB, at 1440p it'd probably behave like Navi21 does at 4k, it loses some of its punch, but it can still do its thing.

AMD Radeon RX 6700 XT ray tracing performance has been leaked - VideoCardz.com

AMD Radeon RX 6700 XT faster than GeForce RTX 3070 without raytracing It looks like Wccftech got their hands on RX 6700 XT performance number, including ray tracing. New performance figures have been shared by Wccftech day. The site has provided information on RX 6700 XT framerates in 1440p...

videocardz.com

That was a good prediction! So, it actually turned out to be a middle sized 335mm² chip mostly matching the 3070/ 754mm² 2080Ti in rasterization, lol. Somewhat bigger than expected.

Should be a complete match or faster at 1080p.

96MB IC too. Great stuff. All recent AMD architectures' sweet spot keeps being ~40CU.

 

[Stuka87]

Everything is possible.
19 Tflops? That would be for example 72CU at 2.05GHz, that is doable.
I have to wonder, If RDNA2 is really so much better than RDNA1 and because of that AMD didn't bother to make a bigger RDNA1 chip to combat 2080 Ti or It was because of some limitation in RDNA1.

I think AMD knew from the outset that RDNA 1 was just a stepping stone (which they have mentioned) and with limited 7nm capacity (at that time) they went with the mainstream market, which significantly outsells the high end market.

 

[maddie]

If your main limitation is fab capacity, you would be better off prioritizing high margin products over high volume products.

Design costs amortization. Total returns not just margins.

 

[DisEnchantment]

So I loaded the amdgpu as an eclipse C++ project to check some stuffs and discovered they actually support reading the umc via ioctl and not only from SMI. Didn't know this is exposed via ioctl

static void gmc_v10_0_set_umc_funcs(struct amdgpu_device *adev)
{
    switch (adev->asic_type) {
    case CHIP_SIENNA_CICHLID:
        adev->umc.max_ras_err_cnt_per_query = UMC_V8_7_TOTAL_CHANNEL_NUM;  // --> 16 ( 2 * 8 )
        adev->umc.channel_inst_num = UMC_V8_7_CHANNEL_INSTANCE_NUM; // 2
        adev->umc.umc_inst_num = UMC_V8_7_UMC_INSTANCE_NUM; // 8
        adev->umc.channel_offs = UMC_V8_7_PER_CHANNEL_OFFSET_SIENNA;
        adev->umc.channel_idx_tbl = &umc_v8_7_channel_idx_tbl[0][0];
        adev->umc.funcs = &umc_v8_7_funcs;
        break;
    default:
        break;
    }
}


gmc_v10_0_early_init() --> gmc_v10_0_set_umc_funcs()

You can actually perform ioctl to query the HBM for errors.
When you perform ioctl to get the memory error count this call chain gets invoked which ends up in a register read.

amdgpu_ctx_ioctl() 
    --> amdgpu_ctx_query2() 
        --> amdgpu_ras_query_error_count() 
            --> query_ras_error_count() 
               --> umc_v8_7_query_ras_error_count()  // from gmc_v10_0_set_umc_funcs

Same thing with Vega20 and Arcturus

static void gmc_v9_0_set_umc_funcs(struct amdgpu_device *adev)
{
    switch (adev->asic_type) {
    case CHIP_VEGA10:
        adev->umc.funcs = &umc_v6_0_funcs;
        break;
    case CHIP_VEGA20:
        adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
        adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
        adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
        adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_VG20;
        adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
        adev->umc.funcs = &umc_v6_1_funcs;
        break;
    case CHIP_ARCTURUS:
        adev->umc.max_ras_err_cnt_per_query = UMC_V6_1_TOTAL_CHANNEL_NUM;
        adev->umc.channel_inst_num = UMC_V6_1_CHANNEL_INSTANCE_NUM;
        adev->umc.umc_inst_num = UMC_V6_1_UMC_INSTANCE_NUM;
        adev->umc.channel_offs = UMC_V6_1_PER_CHANNEL_OFFSET_ARCT;
        adev->umc.channel_idx_tbl = &umc_v6_1_channel_idx_tbl[0][0];
        adev->umc.funcs = &umc_v6_1_funcs;
        break;
    default:
        break;
    }
}

 

[Helis4life]

Wow, so it is possible to have a Youtuber that can understand how hardware works.

*Goes into shock*.

Anyway, here's some recommended watching for those of you wanting to put 20 minutes in. Audio's not the best, but it's still audible and gets the job done.

What I thought was interesting was his comment about AMDs RT implementation being faster/more efficient than nvidias, if the developer implemented it appropriately. Coupled with the fact that both consoles will be using rdna2s implementation of RT, this might mean we see wider developer adoption for AMDs pathway.

I'm curious how a Dev like CDProjektRed will then handle the differing RT pathways, one for the pc and one for the consoles and whether for instance both pathways could be implemented and the appropriate one chosen by the engine at runtime

The nvidia skinworks comment made me chuckle a bit. Can definitely see something like that in the future

 

[CakeMonster]

I played Control in 4K with DLSS (changed monitors around) and going back to 1600p and native was a huge relief.

I'm very sensitive to sharpening and it really bothered me. If sharpening can't be disabled with DLSS in the future, I'm out, there's no way I'll be using it for as long as I can hold off.

Also, DLSS messed up small text (signs etc) where the resolution would fool you into thinking it would be discernable but its just a sharpened mess. Its minor but not pretty.

Thirdly, on high contrast edges DLSS approached MSAA effect both with regards to resolution and lack of staircase effect. However with low contrast its still not very aliased but its a low resolution blur and does not look 4K. The huge grey column in the Foundation DLC against a gray background has very blurred edges compared to native.

The film grain effect (for those who like that) is pretty much ruined with DLSS too.

Edit: A good experiment that I recommend everyone do (while standing still in place) is turn DLSS resolution way down to study what it does, then turn it gradually back up toward native and compare the effects, then lastly turn it off on native.

 

[Hail The Brain Slug]

Just checked techspot (hardware unboxed) and in gears 5 4k ultra they have the 3080 at 72 fps avg and the 5700xt at 41fps. The 6000 is a 78% uplift over the 5700xt in this review. Techspot also tested with the 3950x.

Guru 3d also show a lot of variation in 3080 and 5700xt numbers so I wouldn't say the numbers AMD showed were conclusive as it seems to swing from 90% of 3080 to on par with 3080.

If that is 80CUs @ 2.2Ghz with IPC improvements it seems really poor to be honest and looks like RDNA2 did not fix the scaling issues GCN had. If it is a cut down 72CU @ 2Ghz part then it looks much better.

NV are going to sell every 3080 and 3090 they can put out between now and October 28th (and beyond probably) so if AMD have not shown the top tier card to allow them to show something surprising on October 28th then its not like it matters.

That's exactly my thought. They may be sandbagging as part of their misinformation campaign knowing so few Ampere cards are even going to be available to purchase before their big announcement in 20 days.

 

[TESKATLIPOKA]

He didn't say RTX3070 will be 20% faster than RTX 2080Ti. It was actually meant that 16Ghz 256bit GDDR6 could provide enough bandwidth to feed a RDNA 2 GPU 20% more powerful than RTX 2080ti.

 

[uzzi38]

This has been known for a while. Microsoft mentioned they would be equivalent. That is full system power draw mind you. Mind blowing...

Yup, just posting this because a lot of people didn't believe them when they said it.

 

[kurosaki]

Yes, overhyped, as everything AMD always does.

Not to be mean or anything. But aren't you overhyping a bit now? I mean, it won't perform better than a five years old handdrawing, and still you proclaim the the 6000-series will perform almost like a videocard?
Isn't there any shame in your body? Do you think you can fool anyone by this. No, the 6900xt will render as bad as a HTC Hero, cut in half, by an axe. Everything else is biased hype.

Heard it from a very trustworthy source on YouTube. He talked about it for like 15 mins, so it must be true.

 

[TBytemaster]

This just in, Navi 21 display engine has been confirmed to only support S-Video output. Comes in an SECC cartridge too, all the PCIe stuff was a red herring.

 

[Glo.]

During AMD Zen 3 keynote, they demoed Big Navi in Gears 5 getting average of 73 FPS, at 4K, ultra preset.

This is 20% better performance, than you get at 52 CUs clocked at 1825 MHz.

So as we can see, this is confirming, that during the Zen 3 keynote, they demoed the largest, and fastest version of the GPU. Thats all there is.

80 CUs, clocked at 2.4 GHz roughly 20% faster than 52 CUs clocked at 1825 MHz.

Scaling. AMD is so incompetent, they thought they can run away with 256 bit bus! Its going to be crap, from top to bottom.

 

[Glo.]

I think this is redacted .No way big navi eat more than 3x more power than consoles.I think igor lost it.Well 8 more days.

AMD Radeon RX 6000 based on Navi 21 XT to feature 320W TBP, 16 Gbps GDDR6 memory - VideoCardz.com

A fresh report from Igor’sLAB puts a new claim on Navi 21 power consumption. AMD Navi 21 XT to feature 320W TGP/TBP Igor Wallossek published his first ‘big leak’ on Big Navi. According to his data, AMD Navi 21 XT will feature TGP at around 320W, not 255W as reported earlier. The difference comes...

videocardz.com

Btw TGP is for entire card.Dont know what the fuck igor doing?

Graphics Cards: TDP and TGP (and don't forget TBP, GCP and MPC...) | Geeks3D

Graphics Cards: TDP and TGP (and don't forget TBP, GCP and MPC...)

www.geeks3d.com

Don't call Igor that "he lost it". That is shooting the messenger just because we don't like his messages.

And he may just be bringing what AIBs are feeding him with.

320W of power he says?

I can tell you guys that this might be an indication that AMD has decided to go all out, and not leave anything on the table.

When I wrote the performance targets for Navi dies, or rather CU counts, I had information that initial performance targets were for 250W boards. As in 250W in total power drawn by the boards. It was at the time when 2.1 GHz max boost speeds were on the table.

Now we are seeing 2400 MHz clock speeds at 300W board power. If we take this in the context - everything starts to make sense.

How is that 2.4 GHz clock speed going to affect performance?

I don't believe anybody looked at it, but yesterday I posted comparison of performance between what AMD demoed during Zen Keynote, assuming it was full specs for Navi 21 die, with Xbox Series X, wondering if anybody will see something.

52 CUs clocked at 1825 MHz = 12.1 TFLOPs.
80 CUs clocked at 2410 MHz = 24.5 TFLOPs.

2X performance over Xbox Series X, at least - the theoretical maximum performance. 2.4 GHz at 300W board power means that AMD might be going all out. Straight up for win.

Will they win? We'll see...

 

[DJinPrime]

RTG recently posted a few videos where he stressed that the consoles are using custom RDNA2 and can't directly make an estimate for the desktop parts because AMD haven't given out any info. His reasonings seems pretty logical, things were added to the console parts (Sony and MS spending tons of money on these) and things were taken out (not required by the consoles). So, who knows what's desktop big Navi, other than it will be competitive.

 

[leoneazzurro]

Hardware- und Nachrichten-Links des 22. Oktober 2020 | 3DCenter.org

Twitterer Kopite7kimi hat Anzeichen einer weiteren GA102-Grafikkarte vernommen, welche zwischen GeForce RTX 3070 (GA104-300) und GeForce RTX 3080 (GA102-200) liegen und möglicherweise auf den Codenamen "GA102-150" hören soll. Hiermit könnte

www.3dcenter.org

germans are theorizing that the leaked synthetics is actually a successor of the 5700xt and not a high or top end card based purely on the fact that 5700xt regularly could beat 2080s on synthetics and lose on actual games

They are ignoring the fact that Igor's lab gave an edge to Navi21 in TSE test, too - and frankly they are calling FSU an "AMD biased test". Strange how they had not called TSE to be an "Nvidia biased test" when there was a big controversy years ago about the way it handles async compute. Sigh.

 

[Veradun]

They are ignoring the fact that Igor's lab gave an edge to Navi21 in TSE test, too - and frankly they are calling FSU an "AMD biased test". Strange how they had not called TSE to be an "Nvidia biased test" when there was a big controversy years ago about the way it handles async compute. Sigh.

Everything working well on nvidia HW is the way it's meant to be played after all, everything else is obviously AMD biased.

I look forward to Intel entering the market and see what happens to fanboys in the press.

 

[Zoal]

One thing to note about that 6700XT vs RTX 3070 - I think people should not look at this "battle" this way.

RX 6700 XT may lose 5-10% to RTX 3070, but it will also be massively more efficient.

If raster performance is comparable NV will heavily push the RT difference in 'modern' games

 

[AmericanLocomotive]

You guys realize that TSMC has been given little credit for the success of AMD products in the last few years. I think TSMC should be given much of the credit for their silicon.

What do you mean? Pretty much everyone acknowledges and even rants/raves about TSMC's 7nm process being superior to just about anyone else's right now. But you almost must keep in mind that 7nm isn't a magic problem solver. Look at Radeon Vega VII vs the 5700XT. The Vega VII is only 5% faster despite having 50% more CUs and consuming significantly more power. Both are on the same 7nm process. Now Navi2 is pushing that efficiency even further. That's AMD's design work there, iterating and improving.

 

[ElFenix]

 

[Stuka87]

Keep in mind that AMD did the smart thing and did all the bench marking with their new 5000's unreleased CPU's. This is compared to NVidia using Intel.

Hence all the AMD numbers had a 15% CPU advantage added over Nvidia.

Still very impressive and glad that I held off buying the "10GB gimped" 3080

Per the notes, the nVidia cards were tested with the same CPU's.

 

[Veradun]

Who cares about Upscaling tech in a world in which AMD delivers faster AND more efficient GPUs than Nvidia, both, at the same time?

Why suddenly upscaling tech has become the "make or break" reason for buying a product?

Because people are unable to use the non automated version that's called "moving the graphics settings' sliders"

 

[Glo.]

Don't listen to Wendel, guys, he is BSing you, people.

As we all should know, 6000 series will be as bad as everything AMD always does.