AMD - Zen chitchat

[watercooled]

Last time I checked, mining was generally considerably MORE efficient on Polaris, even vs Pascal. OFC it depends on which algorithm is used though.

That's often been the case with AMD's GPUs - in some applications they can be considerably faster and/or more efficient than an Nvidia counterpart which beats it in gaming tasks. They usually have higher theoretical performance e.g. Vega 64 at ~12665TFLOPS peak vs Titan XP (yeah really) at ~12149.7TFLOPS peak for single precision which makes sense given they have similar die size. Just of course that doesn't translate into how every application will perform, and unfortunately for AMD, gaming is an area where they've typically fallen behind Nvidia in terms of FPS per FLOPS.

When Bitcoin and then Litecoin GPU mining was a thing, AFAIK they were predominantly core-bound so AMD's peak performance again won out massively over Nvidia, which at the time had substantially lower peak performance for comparable products. Now with the current resource-wasters, the algorithms are more memory bound; Polaris 10 matches the 1070 at ~256GB/s, and with Vega 64 you get ~483GB/s. Nvidia's cards do reach high bandwidths with GDDR5X but (again, last I heard) that's less well suited to mining so you end up with e.g. the 1070 actually outperforming the substantially faster 1080.

[Millennium]

Apparently the 2400G's Vega cores can be overclocked to 1.5-1.6GHz from the 1.1 ish base. This places it firmly in RX550 territory.

Now considering that the RX550 is £80+, and the 2400G can be had from like Amazon Europe for around £141, it's a phenomenal value for a budget 1080p gamer for example. Absolutely amazing. I'm so excited I'm considering getting one for my brother as an upgrade, though I am unsure what his R3 1200 will now fetch lol.

[CAT-THE-FIFTH]

Apparently the 2400G's Vega cores can be overclocked to 1.5-1.6GHz from the 1.1 ish base. This places it firmly in RX550 territory.

Now considering that the RX550 is £80+, and the 2400G can be had from like Amazon Europe for around £141, it's a phenomenal value for a budget 1080p gamer for example. Absolutely amazing. I'm so excited I'm considering getting one for my brother as an upgrade, though I am unsure what his R3 1200 will now fetch lol.

I did a quick build spec out of idle interest using CCL Computers,and managed to do a total Ryzen 3 2200G build,with an SSD,HDD,FreeSync monitor,Windows copy,etc for around £600 including 8GB of decent 3200MHZ DDR4.

In the current climate,its not too bad at all,and it should have enough basic grunt if you want to plonk in a faster card too.

[scaryjim]

... Even if you only select 512MB,performance is not much lower. ...

I'd imagine that's because it'll just shunt anything that doesn't fit into dedicated RAM into main system RAM anyway. For a dGPU with dedicated high bandwidth VRAM that's going to have a huge impact, but since an APU's dedicated VRAM is system ram anyway, running out of dedicated RAM really shouldn't affect performance, since it'll still be pulling the data from system RAM either way. In fact, it's probably better to use the minimum realistic amount of dedicated RAM, to avoid unnecessarily eating up RAM that would otherwise be available to the rest of the system...

... Intel says that the top Vega M SKU is competitive with the low power GTX1060 Max Q. ...

It's worth remembering that the Max-Q version is efficiency optimised, not performance optimised, so runs significantly slower (10% - 15%, iirc) than a standard (laptop) GTX 1060...

[Corky34]

The dedicate X amount of RAM to the GPU is really only for reporting purposes to work around issues in some games that weren't written to take account of unified memory architecture.

AFAIK Windows will allocate as much RAM as needed to either the CPU or GPU regardless of what you allocate in the BIOS, i guess that's why TechPowerUp didn't find much difference between 512 MB, 1 GB, and 2 GB, you could knock me down with a feather if you were able to run the Witcher 3 in 1080p on low settings with only 512MB of video RAM.

[CAT-THE-FIFTH]

I'd imagine that's because it'll just shunt anything that doesn't fit into dedicated RAM into main system RAM anyway. For a dGPU with dedicated high bandwidth VRAM that's going to have a huge impact, but since an APU's dedicated VRAM is system ram anyway, running out of dedicated RAM really shouldn't affect performance, since it'll still be pulling the data from system RAM either way. In fact, it's probably better to use the minimum realistic amount of dedicated RAM, to avoid unnecessarily eating up RAM that would otherwise be available to the rest of the system...

You do realise that on forums,and even some reviewers(IIRC) were pushing that the APUs NEEDED 16GB of DDR4,as you NEEDED,to permanently use 2GB of system RAM for the IGPs,hence pushing up the cost,and 6GB would not be enough for Windows 10. The fact is that is utter crap,since you can in most cases use a lower pre-allocated RAM quantity and its fine. Also,TBH,not sure why all of a sudden Win10 apparently needs at least 8GB to run which is kind of weird!

It's worth remembering that the Max-Q version is efficiency optimised, not performance optimised, so runs significantly slower (10% - 15%, iirc) than a standard (laptop) GTX 1060...

That is why I said low power. But when has an RX480 been able to compete in that regard and get similar performance?? If anything the Intel numbers says its slightly faster. It will be interesting to see how it pans out TBH.

If AMD with 1536 shaders,which are also probably not clocking very high in reality,can match a downclocked GTX1060,that is an improvement over Polaris. I think part of it is the design is very unusual,ie,64 ROPS,so there is probably a bigger change under the hood it way things are arranged over earlier AMD GPUs in a similar class.

[Xlucine]

Well AMD will need to do something as Nvidia will be releasing something this year,and if they do Polaris 10 is going to get hammered in anything other than mining sales.

Rumours say 2080 will launch, and will probably come with a 2070. X60 cards normally follow later, so AMD still has time to tweak the V32* cards

* Whatever happened to those rumours of a V32 launch late last year?

[CAT-THE-FIFTH]

Rumours say 2080 will launch, and will probably come with a 2070. X60 cards normally follow later, so AMD still has time to tweak the V32* cards

* Whatever happened to those rumours of a V32 launch late last year?

If they launch the cards before the summer,it would follow that the 60 series would be out around summer time.

We know AMD already has a midrange part in Vega M,but the issue is whether HBM2 makes it viable sadly!

[scaryjim]

... when has an RX480 been able to compete in that regard and get similar performance??...

The RX 580 in the ASUS Strix ridiculous AMD laptop (can't remember it's official model number!) was roughly performance equivalent to a standard laptop GTX 1060. That would stick it 10% - 15% faster than the Max Q, which rather nicely implies than a 2048 shader Polaris at similar clocks would match a Max Q 1060.

We know that Vega is more power efficient than Polaris and clocks higher for a given voltage, and since there's less shaders it would naturally have a lower TDP giving more thermal room to bump the clocks closer to the limit. And the Kaby-G GPU is running HBM2 rather than an energy-sapping GDDR5 memory bus. So a 1536 shader Vega matching a 2048 shader Polaris in the same TDP...?

As it is, I agree that there's probably some special architectural sauce going on there - it probably won't make up the difference on clock speed alone. But it's certainly not as surprising as you seemed to be suggesting; to me it's perfectly credible that they could make up the difference in clock speed alone, given what we know about the various configurations already out there.

[CAT-THE-FIFTH]

The RX 580 in the ASUS Strix ridiculous AMD laptop (can't remember it's official model number!) was roughly performance equivalent to a standard laptop GTX 1060. That would stick it 10% - 15% faster than the Max Q, which rather nicely implies than a 2048 shader Polaris at similar clocks would match a Max Q 1060.

We know that Vega is more power efficient than Polaris and clocks higher for a given voltage, and since there's less shaders it would naturally have a lower TDP giving more thermal room to bump the clocks closer to the limit. And the Kaby-G GPU is running HBM2 rather than an energy-sapping GDDR5 memory bus. So a 1536 shader Vega matching a 2048 shader Polaris in the same TDP...?

As it is, I agree that there's probably some special architectural sauce going on there - it probably won't make up the difference on clock speed alone. But it's certainly not as surprising as you seemed to be suggesting; to me it's perfectly credible that they could make up the difference in clock speed alone, given what we know about the various configurations already out there.

It was the 4GB model,and AFAIK it got very hot in the chassis,and tends to indicate it was possibly worse in performance/watt,but since there are so few powerful AMD dGPU laptops,its hard to do directly compare things. But more importantly its a 1536 shader matching a downclocked GTX1060,and and the bigger difference seems to be 64 against 32 ROPS,and that has never happened on such a low CU count AMD GPU. So the internal layout is definitely different.

Also:

https://www.notebookcheck.net/AMD-Ra....215184.0.html
https://www.notebookcheck.net/AMD-Ra....278686.0.html
https://www.notebookcheck.net/NVIDIA....224734.0.html

Look at the clockspeeds. Its not much in it,and like I said its suspected to be a 1792 shader chip. Its not even bandwidth as Intel lists around 205GB/s and the desktop RX580 has 256GB/s.

However,look at the mobile Vega IGPs,they do very well despite the limited bandwidth and the fact it is shared with a CPU too.

So it makes me wonder whether Vega,is not the disaster many are thinking,as its only the dual use large Vega we saw first,with a lot of stuff not really required for consumer workloads was bogging it down??

This is why I am quite excited to see how the top Vega M SKUs perform in a laptop with sufficient cooling.

[kompukare]

Has anyone considered how Vega in the APUs compare to Intel's Gen9 in terms of area used, transistors etc?
WikiChip.org has articles on both complete with annotated die shots (but only for the normal dual and quad core 24 EU Skylake dies with no indication of what for instance Iris Pro Graphics 580 with it's 72 EUs plus 128MiB eDRAM takes up).
Ryzen 2200G/2400G:

Taken from https://en.wikichip.org/wiki/amd/ryzen_5/2400g#Die_Shot

Intel Quadcore Skylake:

Taken from https://en.wikichip.org/wiki/intel/m...ent)#Quad-core

Some of the other specs of interest here, WikiChip also list:
Skylake Quad GT2 die is 122.3 mm²
Ryzen 5 2400G die is 209.78 mm²

So it seems Intel's cost are a lot lower, but the question of efficient use of space/transistor for the iGPU, the relevant would not be the 24 EU part but rather the very rare 72 EU part with the eDRAM.

Can't find anyone who has measured the die size of the 72 EU part nor many benchmarks. But roughly, on the 24 EU parts, the iGPU (without the other logic like the display etc.) is roughly 1/3 of the die and for the 72 EU part it would have to be 3 times as big. So 2/3 for the CPU, cache, IMC etc. and 1/3 for 24 EU Gen9; and for 72 EU it's 3/3 for Gen9 and 2/3 for the CPUs. That is 3/3 vs 5/3 or x1.66, giving around 202mm². In other words pretty similar to Raven Ridge.

Guess the biggest problem with Intel's Iris Pro is that for the DIY market only Broawell gave the option. In general it seems that Intel are not willing to sell a 200mm² CPU for the mainstream market while AMD are. And tellingly rather than trying to scale their design up, they're going to use Vega for their 'G' part and have hired AMD's ex Radeon chief.

[Xlucine]

Just holding a ruler up to the screen here, but coincidentally the GPU part of the intel chip pictured and the GPU+multimedia engine of the AMD chip are both 35% of the overall area. That's ~74mm^2 for the Vega GPU + multimedia engine, and ~43mm^2 for the intel iGPU. Fully enabled RR is a lot more cromulent than the intel iGPU, and more than doubly so (when compared to a later gen intel iGPU). This comparison is completely ignoring the issue of GPU compute silicon:fixed function multimedia silicon ratio messing with the smaller intel's perf/mm^2, of course

[Xlucine]

Actually, the wikichip article spells out which bit of the intel iGPU is the Gcompute bit (the 8 EU & sampler/tex block). Looking at the size of that bit alone (then multiplying by 3), gives some odd results:

Raven Ridge:
Vega GPU size: 65mm^2
multimedia engine: 8.7mm^2

Skylake:
Overall GPU: 43mm^2
3x 8 EU&tex blocks: 18mm^2
GPU uncore: 25mm^2

Either intel's GPU tech isn't as good on a per mm^2 basis as AMD, or their fixed function block is laughably dire, or something in my analysis is wrong (probably this option)

[CAT-THE-FIFTH]

Apparently Raven Ridge is using an improved 14NM+ node.

[kompukare]

Apparently Raven Ridge is using an improved 14NM+ node.

This is from what CB wrote?
Because I see that, but since AFAIK no other site mentioned, I ignored it.
None of the reviews mentioned any better max OC clocks, or reaching similar clocks on less voltage, so hard to see what improvement this node is meant to bring

[Corky34]

I could be misremembering but doesn't the R5 2400G retain the same TDP as the R5 1400, i would guess the change to 14nm+ had something to do with lowering the power draw, along with losing a CCX obviously, but to throw in a GPU while retaining the same TDP is pretty impressive IMO.