AMD Raven Ridge 'Zen APU' Thread

[lobz]

I dont like to be a sourpuss here, but if you've visited the gpu forums a bit over the last few weeks, whatever information we're getting out of the vega architecture, matching or exceeding Polaris IPC doesn't seem to be it, disregarding that Pascal seems to be able to run higher clocks than Polaris/Vega at the same power consumption to begin with..

don't bet too much money on that being the final state of IPC

 

[AtenRa]

GT 1030 has a 64bit GDDR5 at 6008MHz memory with 48,06GB/s memory bandwidth

RavenRidge with 128bit DDR-4 3200MHz memory will have 51,2GB/s memory bandwidth

A desktop RavenRidge 65W TDP top SKU should be close to GT 1030 performance, much like Kaveri was close to GT730 GDDR5 64bit.

 

[The Stilt]

RavenRidge with 128bit DDR-4 3200MHz memory will have 51,2GB/s memory bandwidth

In before we start to see Raven laptops equipped with a single 2400MHz SO-DIMM from the ODMs

 

[.vodka]

In before we start to see Raven laptops equipped with a single 2400MHz SO-DIMM from the ODMs

That would be the same mistake for the seventh or eighth time in a row... AMD would be like:

 

[AtenRa]

In before we start to see Raven laptops equipped with a single 2400MHz SO-DIMM from the ODMs

I was talking about desktop 65W TDP RavenRidge not laptops.

 

[ao_ika_red]

In before we start to see Raven laptops equipped with a single 2400MHz SO-DIMM from the ODMs

Is that true? AMD really didn't learn from their previous mistake. It's okay to share socket platform from low-end up to high-end APU, but it's extremely unwise to share its memory configuration as well.

 

[itsmydamnation]

Is that true? AMD really didn't learn from their previous mistake. It's okay to share socket platform from low-end up to high-end APU, but it's extremely unwise to share its memory configuration as well.

Its not up to AMD

 

[ao_ika_red]

Its not up to AMD

at least they would prevent it by a proper guideline.

 

[DisEnchantment]

From what I can see in the code is that it looks to me as if RR will not have HBM.

static int gmc_v9_0_sw_init(void *handle)
{
    int r;
    int dma_bits;
    struct amdgpu_device *adev = (struct amdgpu_device *)handle;

    gfxhub_v1_0_init(adev);
    mmhub_v1_0_init(adev);

    spin_lock_init(&adev->mc.invalidate_lock);

    switch (adev->asic_type) {
    case CHIP_RAVEN:
        adev->mc.vram_type = AMDGPU_VRAM_TYPE_UNKNOWN;
        if (adev->rev_id == 0x0 || adev->rev_id == 0x1) {
            adev->vm_manager.vm_size = 1U << 18;
            adev->vm_manager.block_size = 9;
            adev->vm_manager.num_level = 3;
        } else {
            /* vm_size is 64GB for legacy 2-level page support*/
            amdgpu_vm_adjust_size(adev, 64);
            adev->vm_manager.num_level = 1;
        }
        break;
    case CHIP_VEGA10:
        /* XXX Don't know how to get VRAM type yet. */
        adev->mc.vram_type = AMDGPU_VRAM_TYPE_HBM;

Another thing, if the GFX V9 and the SoC 15 family is using Infinity Fabric what are the chances that the GPU is fed from the same Memory controller?

 

[PeterScott]

GT 1030 has a 64bit GDDR5 at 6008MHz memory with 48,06GB/s memory bandwidth

RavenRidge with 128bit DDR-4 3200MHz memory will have 51,2GB/s memory bandwidth

A desktop RavenRidge 65W TDP top SKU should be close to GT 1030 performance, much like Kaveri was close to GT730 GDDR5 64bit.

1st issue: RR is spec'd for DDR-4 2400MHz max, you are essentially invoking running overclocked memory.
2nd issue: It's shared bandwidth for the APU, and dedicated for the dGPU.
3rd issue: NVidia appears to be more efficient and using memory bandwidth than AMD.

I bet GT 1030 kicks RR ass in gaming.

But hey it might beat a Radeon 430/530 rebrands running old 28nm, GCN 1, on 64 bit bus, if you are so keen on having it better a standalone card.

But no one is going to build a new dGPU inferior to an APU, that would be utterly pointless. You can really only expect it to challenge old GPUs.

 

[AtenRa]

1st issue: RR is spec'd for DDR-4 2400MHz max, you are essentially invoking running overclocked memory.
2nd issue: It's shared bandwidth for the APU, and dedicated for the dGPU.
3rd issue: NVidia appears to be more efficient and using memory bandwidth than AMD.

I bet GT 1030 kicks RR ass in gaming.

But hey it might beat a Radeon 430/530 rebrands running old 28nm, GCN 1, on 64 bit bus, if you are so keen on having it better a standalone card.

But no one is going to build a new dGPU inferior to an APU, that would be utterly pointless. You can really only expect it to challenge old GPUs.

1. Desktop SKU will use more than 2400MHz
2. Yes i know, same as Kaveri but look bellow.
3. Kaveri Desktop SKUs are already as fast as GT730 GDDR-5 64bit

Im talking about desktop here, Desktop Kaveri is already as fast as R5 250 (not 250X) so please get your facts straight first
Even without any memory bandwidth improvements from VEGA mArch, just having 50% more bandwidth from the memory it will be enough to reach close to GT1030 performance at 1080p.

 

[PeterScott]

3. Kaveri Desktop SKUs are already as fast as GT730 GDDR-5 64bit

1030 is twice as fast as 730 in just about every real world gaming benchmark:
https://www.youtube.com/watch?v=fXCSG6zdxdQ

AMD is only claiming it's new GPU is up to 40% faster.

1030 will easily beat Raven Ridge for gaming.

 

[AtenRa]

1030 is twice as fast as 730 in just about every real world gaming benchmark:
https://www.youtube.com/watch?v=fXCSG6zdxdQ

AMD is only claiming it's new GPU is up to 40% faster.

1030 will easily beat Raven Ridge for gaming.

Take a better look at what AMD says, its 40% at half the power (35W)
That means more than 40% at same power of 65W TDP (BristolRidge)

edit: the above slide is for Mobile, where they will not use more than 2400MHz memory.

 

[PeterScott]

Take a better look at what AMD says, its 40% at half the power (35W)
That means more than 40% at same power of 65W TDP (BristolRidge)

I think you are just setting yourself up for disappointment.

It's a marketing slide, it will have a combination of numbers that are unlikely to be achieved in the real world.

I doubt there will be a high power desktop model of RR. AMD will be pushing Ryzen for that role.

RR is for mobile and low power SFF.

I am interested in a RR SFF (possibly with passive cooling in Streacom case), but I am tempering my expectations, to be less than marketing claims (which are usually exaggerations), not in excess of marketing claims.

 

[NTMBK]

Take a better look at what AMD says, its 40% at half the power (35W)
That means more than 40% at same power of 65W TDP (BristolRidge)

edit: the above slide is for Mobile, where they will not use more than 2400MHz memory.

They're still limited to dual channel DDR4, don't expect any performance miracles.

 

[Phynaz]

at least they would prevent it by a proper guideline.

AMD isn't in a position to dictate configurations to OEMs. AMD needs to move product.

 

[Phynaz]

Take a better look at what AMD says, its 40% at half the power (35W)
That means more than 40% at same power of 65W TDP (BristolRidge)

edit: the above slide is for Mobile, where they will not use more than 2400MHz memory.

You really haven't learned about marketing slides yet

 

[AtenRa]

I think you are just setting yourself up for disappointment.

It's a marketing slide, it will have a combination of numbers that are unlikely to be achieved in the real world.

I doubt there will be a high power desktop model of RR. AMD will be pushing Ryzen for that role.

RR is for mobile and low power SFF.

I am interested in a RR SFF (possibly with passive cooling in Streacom case), but I am tempering my expectations, to be less than marketing claims (which are usually exaggerations), not in excess of marketing claims.

You probable have mixed up the names, RR = RavenRidge = ZEN CPU Cores + VEGA iGPU. It will come both in Mobile and up to 65W TDP Desktop in 2018 for AM4 platform.

 

[AtenRa]

They're still limited to dual channel DDR4, don't expect any performance miracles.

My estimation is for a dual channel DDR-4 at 3200MHz with 50GB/s bandwidth, which is 50% more than what kaveri has.

 

[AtenRa]

You really haven't learned about marketing slides yet

This one is very possible, new more efficient CPU mArch, new more efficient GPU mArch and on top off those two a new 14nm FF process vs Bulldozer based 28nm Planar SOC.

 

[SPBHM]

Take a better look at what AMD says, its 40% at half the power (35W)
That means more than 40% at same power of 65W TDP (BristolRidge)

edit: the above slide is for Mobile, where they will not use more than 2400MHz memory.

seems very unrealistic for anything "Vega" or "Polaris" based as an IGP to beat the 1030;
1030 uses around 25-30W, and 48GB/s memory, we know memory bandwidth efficiency is one huge win for nvidia,
RX 550 is around the same performance as a 1030 but uses over 50W and 112GB/s memory, AMD IGP as usual is going to be bandwidth and power limited, even with DDR4 3200.

 

[NTMBK]

My estimation is for a dual channel DDR-4 at 3200MHz with 50GB/s bandwidth, which is 50% more than what kaveri has.

But don't forget that there's also a quad core Zen CPU sharing that, which is significantly more than 50% faster than a two module Steamroller CPU.

 

[ao_ika_red]

AMD isn't in a position to dictate configurations to OEMs. AMD needs to move product.

Thank you for taking me back into reality.

 

[AtenRa]

seems very unrealistic for anything "Vega" or "Polaris" based as an IGP to beat the 1030;
1030 uses around 25-30W, and 48GB/s memory, we know memory bandwidth efficiency is one huge win for nvidia,
RX 550 is around the same performance as a 1030 but uses over 50W and 112GB/s memory, AMD IGP as usual is going to be bandwidth and power limited, even with DDR4 3200.

Kaveri with dual channel DDR-3 2133MHz memory (34GB/s Bandwidth) is equal to GT730 Kepler GDDR-5 64bit 40GB/s Bandwidth.
A new Vega iGPU with 50GB/s bandwidth should be very close to GT1030 (48GB/s bandwidth)

 

[AtenRa]

But don't forget that there's also a quad core Zen CPU sharing that, which is significantly more than 50% faster than a two module Steamroller CPU.

Yes but in games the APUs are always GPU limited, thus no need for a lot of memory Bandwidth for the CPU.