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]

No

You are not contradicting me with the text afterward though. It may have been a static frequency before, now it's a dynamic one.

 

[Karnak]

No, Sony did that because they realized that the XSX was far more powerful and they wanted to close the gap as far as they could.

I think it was clear from the beginning for both how high clocks RDNA2 is able to achieve. But regardless you need to make sure that every single console will be able to run at those (up to) 2.23GHz.

And that's why I doubt it. You don't crank up the clocks to >2.2GHz while originally it was planned to hover somewhere around 1.8-2.0GHz.

 

[jpiniero]

And that's why I doubt it. You don't crank up the clocks to >2.2GHz while originally it was planned to hover somewhere around 1.8-2.0GHz.

It can do it, but it's not advisable since it blows the power consumption out. Sony probably was forced to put a better, more expensive cooler in and even then it's not a guaranteed clock speed.

 

[Kenmitch]

-It think you're misinterpreting the on going conversation for the individual post. I'll bring it back around and duck out, don't want to take up too much of the thread on bickering.

All I am saying is the longer AMD goes without releasing additional info, the more sales they lose to people that would normally upgrade to them this cycle (like me).

You did imply that you were ready to pull the trigger on what looks like a good deal to you on the 2080Ti.

AMD might like to have your business and the others in your shoes, but they aren't going to deviate from their intended plans just to make a couple extra bucks.

Option 1....Wait and see
Option 2....Find a deal on a 2080Ti that you can live with
Option 3....If option 1 doesn't work out....Option 2 at a lower price point?
Option 4....Complaining in forums about no news while cyber threatening to spend your money elsewhere.

3/4 of the options seem reasonable to me. 1 of them get's you nowhere in the end....I'll let others decide which is which.

 

[Glo.]

It can do it, but it's not advisable since it blows the power consumption out. Sony probably was forced to put a better, more expensive cooler in and even then it's not a guaranteed clock speed.

I don't expect that 2.23 GHz 36CU GPU to consume more than 140W of power.

P.S.2. I don't expect that PS5 will consume more power than Xbox Series X.

 

[Kenmitch]

It can do it, but it's not advisable since it blows the power consumption out. Sony probably was forced to put a better, more expensive cooler in and even then it's not a guaranteed clock speed.

Only Sony's and AMD's engineers know what the PS5 started as and ended up being in the end. It is what it is when it's launched so it doesn't really matter in the end.

I'm pretty sure Sony isn't gonna just crank up the clocks, throw a beefy cooler on it, and prey it'll last thru the warranty period.

 

[jpiniero]

I'm pretty sure Sony isn't gonna just crank up the clocks, throw a beefy cooler on it, and prey it'll last thru the warranty period.

Well, if the chip does degrade, it will just throttle more. That's certainly a 'benefit' of having variable frequency.

 

[Kenmitch]

Well, if the chip does degrade, it will just throttle more. That's certainly a 'benefit' of having variable frequency.

I always thought variable frequency was for power consumption. Why burn more power when you don't actually need to was the reasoning behind it.

Thanks for pointing out that it's actually more or less a secondary parachute. /s

 

[jpiniero]

I always thought variable frequency was for power consumption. Why burn more power when you don't actually need to was the reasoning behind it.

Thanks for pointing out that it's actually more or less a secondary parachute. /s

It's a console, the GPU is going to be going at max all of the time during gameplay, no?

 

[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;
    }
}

 

[Kenmitch]

It's a console, the GPU is going to be going at max all of the time during gameplay, no?

I have no idea? They do tend to fluctuate in the desktop arena depending on the load. Maybe somebody with a APU could look and see how theirs reacts. I'm leaning towards they fluctuate depending on cpu/gpu load. Kind of like a teeter totter effect within the available power envelope.

Besides....What does that have to do with your Sony pushed it to the limit on the cpu and thru a big cooler on it theory?

 

[jpiniero]

What does that have to do with your Sony pushed it to the limit on the cpu and thru a big cooler on it theory?

It doesn't. Ultimately we will just have to wait and see how often it throttles.

 

[Kenmitch]

It doesn't. Ultimately we will just have to wait and see how often it throttles.

How would you even know if it throttles? Did they add a overlay like you can enable on the pc so you can see what up?

 

[jpiniero]

How would you even know if it throttles? Did they add a overlay like you can enable on the pc so you can see what up?

That was one of the questions I had, could the throttling be severe enough that the average gamer would notice.

 

[Kenmitch]

That was one of the questions I had, could the throttling be severe enough that the average gamer would notice.

I doubt it, unless you throw it in a entertainment centers closed off compartment that lacks airflow. I've always treated consoles like a PC and left them out in the open.

 

[itsmydamnation]

You people actually have to go and listen to what Mark Cerny said. Your only throttling in the extreme case not the average case, and even then its only a few percent of clock and if the "rumors" are true adjusted at a far faster speed then even Renoir Can.

To me its hedging for 256bit SIMD on the CPU, because who knows how much adoption and how dense that code will be , if it doesnt see heavy use you are leaving power/performance on the table.

 

[Olikan]

Scaling..... Why isn't the XSX higher clocked?
52 CU Xbox Series X GPU @ 1.825Ghz
36 CU PS5 GPU can reach 2.23GHz

It's the problem of fixed clocks, XSX have to keep at 1.82Ghz even if gpu is running furmark and the CPU some AVX2 loop code...
The real question is, how hard does PS5 downclock?

 

[Helis4life]

We know from DF's look at the XBSX the 52CUs running at 1.8ghz, on a zen2 at 3.8gz was benchmarked the same as a 2080. Could it be possible the
3.8ghz chip was bottlenecking the gpu? Does gears5 support a low level api such as DX12 on XBSX? would that negate the low clocked cpu?

 

[raghu78]

We know from DF's look at the XBSX the 52CUs running at 1.8ghz, on a zen2 at 3.8gz was benchmarked the same as a 2080. Could it be possible the
3.8ghz chip was bottlenecking the gpu? Does gears5 support a low level api such as DX12 on XBSX? would that negate the low clocked cpu?

The 2080 compare was on a Gears of War demo which was built in 2 weeks on Xbox Series X. It was not optimized for Series X. Anything which is developed for Series X can be optimized much better than DX12 on PC. Consoles have close to metal optimization because their OS and driver stack is hyper efficient and meant to provide low level access to the hardware. Even a 40CU desktop RDNA2 GPU is going to beat the RTX 2080 and even RTX 2080 Super given higher perf/clock than RDNA and 25% higher clocks.

 

[Glo.]

Navi 23 - RTX 2080 Super performance levels +10%, @150W TBP.
Navi 22 - 10-20% above RTX 2080 Ti performance levels @225 TBP.
Navi 21 - 40-50% above RTX 2080 Ti performance @ 275W TBP.

Those were performance targets for RDNA2, that I got hint few weeks ago.

I think its worth quoting this, with a little summary that can be gathered from rumors, to get a little context.

N23 - 40 CUs/192 bit bus.
N22 - 60 CUs/256 bit bus.
N21 - 80 CUs/384 bit bus.

 

[Gideon]

I think its worth quoting this, with a little summary that can be gathered from rumors, to get a little context.

N23 - 40 CUs/192 bit bus.
N22 - 60 CUs/256 bit bus.
N21 - 80 CUs/384 bit bus.

Wow, so a 40CUs / 192bit bus GPU beats a 2080 Super (48 SMs and 495.9 GB/s BW)?

That's 20% less CUs and 50% less bandwidth (336.0 GB/s vs 495.9 GB/s as it almost certainly has 14 Gbps GDDR6 modules, rather than the more expensive 15.5 Gbps ones only 2080 Super uses this-far).
It would need a clock speed of about 2.2Ghz to compensate the shader-unit deficiency so I guess with 2.3+ Ghz, it could be +10% faster in some cases, provided it's very bandwidth-efficent.

Awesome if true, can't wait for any kind of leaks even broadly supporting these claims

 

[Glo.]

I will not be surprised if AMD will use 16 Gbps GDDR6, but this is just my specualtion. I have no data on this front.

 

[Gideon]

I will not be surprised if AMD will use 16 Gbps GDDR6, but this is just my specualtion. I have no data on this front.

It's quite possible (along with ~2.2 GHz game-clocks), if the stated performance gains are true.

 

[Shivansps]

Last time AMD had the chance to release a really good price/perf GPU lineup because Nvidia went crazy on prices and low on perf improvements, they went ahead and only bothered with increasing the prices to match Nvidia gpus in order to offer just a bit better performance for the same money, while lacking all the extras like RT and DLSS.

So im a little exceptical, specially because it has been A WEEK, it they had something they would have show it by now, why let RTX3000 gain momentum like this?

 

[Panino Manino]

It's the problem of fixed clocks, XSX have to keep at 1.82Ghz even if gpu is running furmark and the CPU some AVX2 loop code...
The real question is, how hard does PS5 downclock?

But is that really "a question"?
Cerny's goal with the variable clock was to reduce noise. The explained that it annoyed him greatly that the CPU was always at max forcing the cooler to spin and make a ruckus even when the game shouldn't be pushing the hardware, even on menus, even on simpler indie games.
I think it's not about how far can the GPU go, it's about how low, to reduce noise. How far can the GPU go it's irrelevant, thought because the same principle applies partly to the CPU developers can push the GPU a bit higher than the standard.