Vega refresh - Expected? How might it look?

[swilli89]

Raven Ridge (2400G) 11 CU parts, are not competitive with Kaby-G 24 CU parts.

I expect AMD will deliver a 24+ CU mobile part, either paired with Ryzen, or on it's own that can be used with Ryzen.

It makes no sense at all for AMD to force anyone who wants a better AMD Vega mobile GPU, to buy their competitors CPU in order to get it.

A lot of what is happening at RTG/AMD doesn't make sense. Given the fact that they've basically abandoned the road maps that we received not even a year ago + the fact they are allowing Intel to have the fastest mobile VEGA SOC. Especially weird given Koduri's sudden departure and the nature of it went down..

Something tells me AMD may be looking at the ever increasing amount of R&D required for graphics and might be winding it down to go back to their core x86 focus. Could RTG be spun-off?

 

[PeterScott]

A lot of what is happening at RTG/AMD doesn't make sense. Given the fact that they've basically abandoned the road maps that we received not even a year ago + the fact they are allowing Intel to have the fastest mobile VEGA SOC. Especially weird given Koduri's sudden departure and the nature of it went down..

Roadmaps change, and people leave. That is just normal business.

The only thing you have left is the point I mentioned, giving their best Vega Mobile to Intel as an exclusive. That is the only thing that really doesn't make sense, and I expect it will be remedied soon.

Something tells me AMD may be looking at the ever increasing amount of R&D required for graphics and might be winding it down to go back to their core x86 focus. Could RTG be spun-off?

Where would AMD CPUs be without integrated graphics? Today most people are using integrated graphics on their computers. Spinning off RTG would be a significant disadvantage for AMD. What about all the console APUs they sell? Where would those be without RTG?

AMD CPU and RTG are pretty much inextricably linked at this point.

 

[iBoMbY]

The only thing you have left is the point I mentioned, giving their best Vega Mobile to Intel as an exclusive.

So far I haven't seen that Intel's "Vega M" is equal to "Vega Mobile"? The device IDs are clearly different, and the "Vega M" do have a GFX804 OpenCL string, and also the frequencies are more like Polaris than Vega. The only thing really in common is one stack of HBM2 so far, the size, and number of CUs, may, or may not, be the same, and the package for sure isn't equal.

Edit: At least it was still gfx804 yesterday: http://browser.geekbench.com/v4/compute/2190746

 

[swilli89]

Roadmaps change, and people leave. That is just normal business.

Where would AMD CPUs be without integrated graphics? Today most people are using integrated graphics on their computers. Spinning off RTG would be a significant disadvantage for AMD. What about all the console APUs they sell? Where would those be without RTG?

AMD CPU and RTG are pretty much inextricably linked at this point.

You're over generalizing. They have failed to execute on major roadmap achievements and lost the head of the graphics business to their competitor.

Also.. "Where would AMD CPUs be without integrated graphics?" Well.. in terms of Ryzen.. they'd be exactly where they are now.

Roadmaps may change but theirs hasn't.. its just fallen silent with major milestones missed. If they don't spin off RTG with some sort of legacy licensing deal then I see it as limping along to basically be the iGPU component of the increasingly competitive Ryzen products. Console APUs helped them.. but even with 100% of the mainstream console market it barely kept them afloat while total revenues were down big time.

Once Ryzen/Epyc pick up in terms of market share the console APUs and semi-custom in general will start to pale in comparison in terms of overall revenue. Console APUs are already very low margin and this works against their goal of raising gross corporate margins.

Vega arrived super late, extremely inefficient for its large die space, and it won't see any successor to either next summer's 7nm Vega or a late 2019/early 2020 7nm Navi. The trend is downward facing in terms of any type of discrete graphics adapter market for AMD, unfortunately. I support the company and would like to see them reverse this trend, but a trend it is!

 

[PeterScott]

You're over generalizing. They have failed to execute on major roadmap achievements and lost the head of the graphics business to their competitor.

Did you react this way when they lost Jim Keller? Keller was arguably better at his job.

The reality is people in these positions move on a regular basis.

Also.. "Where would AMD CPUs be without integrated graphics?" [Well.. in terms of Ryzen.. they'd be exactly where they are now.

Again, ignoring that most of the CPU market is actually using integrated graphics parts, and this trend is increasing over time.

Abandoning the portion of the CPU market that is growing, to concentrate on the shrinking part, is not a brilliant business move, it's a company destroying move.

 

[SirDinadan]

According to many here, AMD should have spun off / cash in on RTG when the CPU business was losing big-time - pre-Ryzen times. Back then AMD Radeon solutions were rather competitive against NV products.

Now, the tide has turned. But AMD should spin-off RTG nevertheless. Funny ...

 

[zinfamous]

Did you react this way when they lost Jim Keller? Keller was arguably better at his job.

The reality is people in these positions move on a regular basis.



Again, ignoring that most of the CPU market is actually using integrated graphics parts, and this trend is increasing over time.

Abandoning the portion of the CPU market that is growing, to concentrate on the shrinking part, is not a brilliant business move, it's a company destroying move.

They didn't lose Jim Keller. He signed a short term contract that only involved the general design behind the Zen core, to establish a ~5 year+ roadmap for AMD. Completing this, his contract was over. He's worked for what, AMD twice, Apple once, ...now he's at Tesla after AMD? I forget. He's a known and expected journeyman in the field.

 

[sandorski]

AMD won't spin off RTG at this time.

 

[ZGR]

AMD needs an efficient option out of the box. I hope the next RX 56 equivalent's voltage and clocks are a lot lower than the next 64. Maybe a Vega Nano.

 

[Guru]

Lisa Su herself said not to expect any new graphic cards before late 2018, more likely 2019. I think they are working on it, but they are likely waiting for the 12nm process to mature even further as they use it in processors for now and for HBM2 to become more available and cheaper.

They are also likely to use GDDR6 memory and that is very low in supply as well, so even some of they key components for gpu's are in super low supply.

If we look at the reality of the situation, there is not much room for improvement. Going from 14nm/16nm to 12nm is not going to yield big gains, the only solution is to increase die size, but that would also increase cost and make mid tier and entry tier graphics not worth it.

 

[jpiniero]

Lisa Su herself said not to expect any new graphic cards before late 2018, more likely 2019. I think they are working on it, but they are likely waiting for the 12nm process to mature even further as they use it in processors for now and for HBM2 to become more available and cheaper.

That's Vega 20, which is TSMC 7 nm and primarily targeting HPC/DL market and is not a gaming chip.

 

[raghu78]

AMD won't spin off RTG at this time.

AMD will never spin RTG off. The GPU ip is critical to their product differentiation and they have a clear leadership over Intel in graphics performance. Moreover semi custom business depends on SoC with strong CPU and GPU. The CPU division has turned around and will payback the GPU division for the pain it had to endure when Zen was prioritized over everything else. RTG will get the resources it needs to build a competitive next gen GPU arch. I think we will see a RTG turn around by 2020.

 

[swilli89]

Did you react this way when they lost Jim Keller? Keller was arguably better at his job.

The reality is people in these positions move on a regular basis.

Well Keller left a successful product behind that will power AMD for years to come. Raja? Well RTG is in a very bad space, especially once crypto stuff normalizes.

I agree they 95% likely WON'T spin-off RTG unless there was some licensing clause that basically allowed them to integrate iGPU's into their designs sort of how Apple has traditionally integrated a 3rd party GPU directly on die to their in-house mobile CPUs.

IF AMD decides the discrete graphics card market isn't worth the investment and IF they could secure a licensing deal to continue using GCN or whatever replaces it on their Zen cores then I could see this happening. They raise a massive amount of capital, end investments in GPU development, and maintain iGPUs. It just depends on who would be a potential buyer.

 

[krumme]

Who else than Intel is there?

 

[Headfoot]

Vega refresh talk seems premature when people can't even buy the first version

 

[nismotigerwvu]

Did you react this way when they lost Jim Keller? Keller was arguably better at his job.

I don't really think there's any rational argument you could construct to paint Raja as better at his job than Keller. You can randomly draw any of Keller's work out of a hat and it would have had a greater long positive term impact on the industry than even a cherry picked contribution Raja has made. The dude implemented the basis for what we call infinity fabric today back in 96. 20 years is EONS in tech and yet there it is, still out there on the cutting edge. Oh and he authored x64, worked on the DEC Alpha line, K7 and was lead architect on K8. Those are just highlights from what he's done on the desktop/workstation field. Raja seems like a pretty cool guy and is clearly talented, but let's be realistic about their places on the totem pole.

 

[DisEnchantment]

https://support.apple.com/en-us/HT208544

All AMD eGPUs for macOS High Sierra, Apple had it enough I think.
Unannounced AMD 9A90 (Vega 12) card support is present in High Sierra 10.13.4

For Linux Vega 12 is supported in Kernel version 4.17
Interestingly Vega 12 has some differences from Vega 10 that warrants a a change in userland Mesa rendering Libraries beside usual adding another case for the ASIC enum.

https://cgit.freedesktop.org/mesa/mesa/commit/?id=5425d32fcf400bac42a2667d230fb3823c3a3f0c

Updates both for RADV and AMDVLK

Kernel 4.17 will also bring Wattman to Linux.

 

[DisEnchantment]

And Vega 20 has now shown up in Kernel side drivers

https://cgit.freedesktop.org/~agd5f/linux/commit/?h=vg20-dkms-4.13

+ /* Vega 20 */
+ {0x1002, 0x66A0, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA20},
+ {0x1002, 0x66A1, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA20},
+ {0x1002, 0x66A2, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA20},
+ {0x1002, 0x66A3, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA20},
+ {0x1002, 0x66A7, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA20},
+ {0x1002, 0x66AF, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA20},

Some interesting changes in the memory configuration.

df_v3_6 changed DF_CS_AON0_DramBaseAddress0_IntLvNumChan_SHIFT/ MASK. And vega20 has 32 bit channels of HBM2.

Reviewing the calculation, Vega 20 will have a maximum of 4096 bits wide HBM compared to 2048 of Vega 10.
With newer 2.4 Gbps 1.2v parts from both Samsung and Hynix 1.2 TB/s would be guaranteed with Q1 availability and additional volumes being available from Samsung shortly.

Vega 20 interesting bits related to PCIe 4.0 and xGMI are interesting to say the least.
With Navi already taped out I can imagine GFX9 to be largely unchanged and the bulk of that 7 nm gain is for the DPFP, with tiny changes trickling over from Navi.

Would be interesting to know once the 4.17 will be merged and 4.18 is available.
Going forward the strategy is to have full ASIC support before announcement which means if Vega 20 is merging in 4.18 the product could have a chance of being announced by October else would be the version after beginning next year.

Videocardz.com has so far been spot on with many of these leaks related to AMD.

Update:
if git rebasing and upstreaming is not buggy (sometimes is), Vega 20 changes were being authored all the way from 2017 October. Vega 12 from 2017 September.

author Shaoyun Liu <Shaoyun.Liu@amd.com> 2017-10-31 13:32:53 -0400
committer Alex Deucher <alexander.deucher@amd.com> 2018-03-28 10:28:38 -0500

 

[HurleyBird]

I wonder if Vega 20 doubled the ROPs as well as the width of the memory interface...

 

[DisEnchantment]

I wonder if Vega 20 doubled the ROPs as well as the width of the memory interface...

If you check the code, the Memory interface is increased from 2048 to 4096 bits. About the ROPs I dont think so, this is GFX9 for all intents and purposes.
But coupled with higher HBM2 transfer rates and 1/2 DPFP this would be a beast compute card.
but more interesting than anything else is xGMI

 

[Krteq]

Some interesting changes in the memory configuration.

df_v3_6 changed DF_CS_AON0_DramBaseAddress0_IntLvNumChan_SHIFT/ MASK. And vega20 has 32 bit channels of HBM2.

Reviewing the calculation, Vega 20 will have a maximum of 4096 bits wide HBM compared to 2048 of Vega 10.

Can you please elaborate?

//THX a lot

 

[DisEnchantment]

Can you please elaborate?

Vega 20 has 4 stacks of HBM2. Each HBM stack has 8 Memory Channels of 128 Bit wide.
So 4 Stacks of HBM = 4 *8 * 128 = 4096 bit Interface.
Also 4096 * 2.4 Gbps = 9.6 Tbps = 1.2TB/s (or 1 TB/s using slower 2.0 Gbps HBM2)

df_v3_6 changed DF_CS_AON0_DramBaseAddress0_IntLvNumChan_SHIFT/ MASK. And vega20 has 32 bit channels of HBM2.

If you checkout the code, the index in an array of the number of channels is read out from a HW register and it use to calculate the interface width above.
so interface width = Channels count * 128.

Spoiler: drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c

static u32 df_v3_6_channel_number[] = {1, 2, 0, 4, 0, 8, 0, 16, 32, 0, 0, 0, 2, 4, 8};

    chansize = 128;
 
    tmp = RREG32_SOC15(DF, 0, mmDF_CS_AON0_DramBaseAddress0);
 
    switch (adev->asic_type) {
    case CHIP_VEGA20:
        tmp &= 0x3C;
        tmp >>= 0x2;
        if (tmp <= ARRAY_SIZE(df_v3_6_channel_number))
            numchan = df_v3_6_channel_number[tmp];
        else
            numchan = 1;
             break;
   
   
    adev->gmc.vram_width = numchan * chansize;

But interesting thing is that looking at the array of valid values, you can guess the interface width (and therefore guess BW) of other SKUs.

 

[DisEnchantment]

Hmmmm ... not sure if for real. Polaris with GDDR6? Or Vega 12 ? 5% performance uplift

https://www.tweaktown.com/news/61392/amd-rumored-release-radeon-rx-500x-series-q3-2018/index.html

According to industry sources of mine, AMD will be releasing a new Radeon RX 500X series family, which should be around 5-6% faster than the RX 500 series. A new 'X' series makes more sense than a full push into the RX 600 series, as the RX 500X series won't be offering anything new. AMD will reportedly be releasing the Radeon RX 500X series sometime in June-July, which should be able to counter the lower/mid-range graphics card market once NVIDIA stamps down again on the higher-end GTX 1170/1180 markets. There's not much news here as the name of the card and 5-6% additional performance was the only thing I could squeeze out of my source.

 

[amenx]

TPU deleted this article, but fwiw...

http://webcache.googleusercontent.c...techpowerup.com/gpudb/3191/radeon-rx-vega-64x

7nm vega 64x. 4% faster than vega 64 with presumably better efficiency and maybe lower price?

 

[DisEnchantment]

TPU deleted this article, but fwiw...

http://webcache.googleusercontent.c...techpowerup.com/gpudb/3191/radeon-rx-vega-64x

7nm vega 64x. 4% faster than vega 64 with presumably better efficiency and maybe lower price?

Looks like a placeholder, Vega 20 is 4096 Bits wide and around 1 TB/s using slower HBM2. The code in the vega 20 bringup branch should be more trustworthy than TPU.
Also at same performance the 7nm should offer lower TDP than the stated TPU TDP of 250W unless something is really wrong.

Sidenote: Suzanne Plummer is now CVP of Radeon Technologies Group. From CPU to graphics, integrating some common stuff. Interesting.