AMD - Zen chitchat

[CAT-THE-FIFTH]

Caching for the wide slow pipe that is HBM must be, well, interesting so I presume they will have built lessons learnt from Fury into the cache of Vega.



Last couple of years? The GTX460 is some time back now, doesn't time fly!
Still, the 1080ti is a reworked Titan, not a consumer part, and whilst you can find compute tasks where the Vega is faster it generally loses.

https://www.anandtech.com/bench/product/1937?vs=1939

You should know better dude - when Nvidia made Maxwell,they kept on Kepler for commercial markets in the form of the GK210.

I honestly can't believe you have forgotten that. Now,look at how AMD started loosing performance/watt in gaming in all markets once they started doing that. It was the same with Fermi vs Terascale. Nvidia used a general purpose design against a stripped out design made for gaming. ATI/AMD won back then,and now the roles are reversed its the same for Nvidia.

The GP102 is a consumer part - Vega is used in certain AI stuff where the GP102 isn't,and you might want to look at the Radeon Frontier cards and Instinct cards. Its the same GPU used in the Instinct series:

https://instinct.radeon.com/en/6-dee...deon-instinct/
https://wccftech.com/amd-radeon-inst...-accelerators/

The GP102 lacks proper FP16 support,which is used in certain aspects of machine learning. This is why Vega has it - look at the larger GP100. Most of the extra die area is taken up by dedicated FP16 hardware,and it is significantly bigger than the GP102.



If you look at both the GP100 and GP102 they have the same basic number of compute units,but the GP100 adds dedicated FP16 units too. Remember,HBM2 also means less transistors taken up on the physical chip by the memory controller,so Nvidia has spent billions of transistors to add on FP16 support.

Despite this,the GP100 has 25% more transistors and is 610MM2 against 471MM2. Now are you honestly telling me with Vega having enhanced FP16 support,the ability to address an SSD directly,etc its not using a fair amount of extra transistors?? It is,so one has to ask how much better Vega would have been without all that extra stuff with regards to gaming effiency.

AMD uses a general purpose design to target multiple markets. Another set of functionality AMD has,which Nvidia lacks - the ability to address an SSD directly even in Polaris. These all take up transistors and causes increased power consumption. It adds zero to gaming performance,but increases GPU size and power consumption.

I am not sure why people are so surprised that AMD is loosing the race in performance/mm2 and performance/watt for gaming - I said this YEARS ago,that unless AMD starts having proper dedicated gaming lines,that they were not to going to be able to compete. Nvidia realised this after Fermi.

[DanceswithUnix]

Nvidia realised this after Fermi.

The GTX460 was the first Nvidia part to sacrifice FP64 throughput to make it a dedicated gaming part, so Nvidia knew it long enough to include it in the Fermi generation.

Forgot about Vega being able to do fp16, probably because there are already lots of dedicated processors out there which can do the job better of AI than any graphics card so I tend to gloss over it. I suppose for game AI it might be useful, but that makes fp16 a gaming feature.

I go back to: the 1080ti generally beats Vega64 in compute tasks. Yes there are better Nvidia platforms out there, but the 1080ti is configured with the same memory bandwidth as well as having the same die size so makes for the most interesting comparison.

[CAT-THE-FIFTH]

I guess I'm used to seeing 560mm^2 Titan chips at 28nm and the P100 at 610mm^2, but you are right AMD are just under performing per square mm if you compare with the 1080ti.

The GTX460 was the first Nvidia part to sacrifice FP64 throughput to make it a dedicated gaming part, so Nvidia knew it long enough to include it in the Fermi generation.

Forgot about Vega being able to do fp16, probably because there are already lots of dedicated processors out there which can do the job better of AI than any graphics card so I tend to gloss over it. I suppose for game AI it might be useful, but that makes fp16 a gaming feature.

I go back to: the 1080ti generally beats Vega64 in compute tasks. Yes there are better Nvidia platforms out there, but the 1080ti is configured with the same memory bandwidth as well as having the same die size so makes for the most interesting comparison.

I think you cannot fathom how much a change Maxwell was for Nvidia - Kepler was their best high end commercial GPU until the GP100.

You read that right - the the GK210 was used over even the GM210. This is how stripped out Maxwell was for certain usage patterns.

You made a statement not considering all variables,which many enthusiasts make - I have done it too. Just dismissing FP16 support is ridiculous when it clearly is adding to the size of the GPU.

Hardly any game or normal application uses it.

The FP16 support alone on Pascal needed at least 3 BILLION transistors to implement,and Vega adds it as a feature right down the stack(apparently). Its the same with Radeon SSD functionality. All this adds transistors.

Using your metrics,are we to say Pascal has extremely poor performance in deep learning just based off the GP102 or GP104?? You can't say that.

Its like comparing a sports car which is entirely stripped out,which uses a slightly smaller engine than its competitor,but handles a bit better,is faster and more fuel efficient and proclaiming victory,when the competitor has a 4WD system and can actually go off road.

I mean I am not shocked especially when we saw the same with the previous high end AMD GPUs,they have always been a compromise. They attempt to address the same high end markets Nvidia does,but AMD rarely spends the transistors Nvidia does.

AMD is trying to cover all markets with far less models than Nvidia is currently using.

I mean Vega is tiny compared to the GP100 at 610MM2,let alone their latest Volta GPU which is 815MM2 in size.

So what do you expect when AMD is trying to make an "economy sized" one size fits all GPU to not only attempt to fight similar sized Nvidia gaming GPUs in gaming,but also 610~815MM2 behemoths.

Something has got to give,and that is usually power consumption and gaming effiency.

Why do you think AMD cards are clocked past their optimal range so often? Its happened with EVERY AMD high end card since the HD7970GE.

Plus also you have forgotten its not even the same company making the GPUs - its also GF vs TSMC,and every indication still hints TSMC 16NM is still better,let alone the fact Samsung 14NM is still probably better than GF 14NM. Yet,look at the Pascal parts made on Samsung 14NM - they have the worst effiency of all Pascal parts. It was the same with Apple - their chips made by TSMC had better performance/watt which ultimately means better performance too.

AMD will continue to loss the performance/watt race and performance/mm2 race when it comes to gaming,unless they seriously change tactics.

[scaryjim]

... I go back to: the 1080ti generally beats Vega64 in compute tasks. ...

But that's largely because current compute tasks are FP32-oriented, where GP102 has peak throughput. Not only is it massively shorted on FP16 throughput, it's FP64 ratio is half of Vega's too. It's an almost entirely FP32 focussed card, which fits with current software/benchmarks.

The whole point of AMD sticking high FP16 throughput in Vega is to try to move the software world forward. FP16 *is* currently available in game engines, but very few make use of it, and nvidia's lack of native FP16 support isn't going to encourage them to start. They're also pushing the professional cards as AI accelerators, sure, but that's a consequence of only having (being able to afford?) a single silicon variant - they need to try to sell it to as many markets as possible.

As CAT says, it's a strategy that means they aren't competing directly with nvidia's gaming-focused cards. They could gain a big advantage if game devs start using FP16, and the capability is there in modern graphics APIs, but while nvidia hold the greater market share and don't provide good native FP16 support, what game dev is going to incorporate it?

As an aside, Vega's FP64 compute is actually well down against GP100. I assume that's a direct result of trying to produce a balanced silicon to address both gaming and professional/compute markets, but it means they're ceding a - potentially small but profitable - chunk of the market to nvidia. But the existence of both GP100 and GP102 indicates very clearly that nvidia now think split silicon lines is the way to address the professional and gaming markets at the top end...

[CAT-THE-FIFTH]

Also another thing - do people honestly think AMD would use expensive HBM2,just for a gaming GPU? Nvidia itself uses HBM2,so it can cram as much stuff like FP16 units,etc into its GPUs,since the space taken up by a normal memory controllers is far more. Its blatantly obvious with AMD making a big deal of its FP16 capabilities in Vega,and using HBM2 so they can cram as much stuff into the GPU,the focus despite all their PR bumpf is not gaming(poor Volta indeed!),is making a "cheaper" alternative to the larger more expensive Nvidia GPUs for machine learning. Yep,you read it right - a "cheaper" alternative.

I mean seriously - look at the performance difference between Polaris and Vega64,gaming performance/mm2 actually got worse.



Polaris 10 is less than half the size of Vega 10. Even if you add another 10% for better drivers,its the same problem.

It clearly shows you what AMD actually did with Vega 10 - make an economy sized GPU for AI stuff so they can take some sales from Nvidia. Then they decided to clock it as high as they could for desktop,and that was that.

The thing is though we should have gleamed this much earlier,since AMD were talking more about the Radeon Instinct series and less about the gaming aspects of Vega for quite a while.

The sad thing is they kind of backed themselves into this with their marketing campaign,which probably didn't help either in some ways.

Personally I am kind of worried though - Volta/Ampere is out next year,and I really don't know AMD will be able compete,especially if they have fixed DX12/Vulkan performance.

I honestly hope the Vega refresh or Navi get released next year,otherwise AMD are going to have big issues when it comes to gaming sales.

If the GTX1060 replacement gets close to GTX1070 level performance,AMD is in a very tight place indeed.

[DanceswithUnix]

If the GTX1060 replacement gets close to GTX1070 level performance,AMD is in a very tight place indeed.

I think they already are, or at least from where I am sitting.

I bought a VR Headset, and am glad that I did as although my GPU is below the minimum spec I am enjoying playing Elite once more. That means I am in the market for a new card.

RX580 is only 50% faster than my current card, and that isn't enough to part with £250.
Vega 56 would be tempting at £400, but isn't tempting as it is not in the shops.
Vega 64 can be found, but only at upwards of £550 which seems rather steep.

I would much prefer an AMD card because of the open source Linux drivers, but if I had to buy a card right now it would have to be a 1070 ti. Frankly the main reason I haven't is that the whole range is now quite old, so given the expected 5 year support cycle for drivers I would only get about 3.5 years of Linux driver support before it drops into the legacy driver skip where I would usually switch to open source drivers which in this case don't exist and probably never will.

[Xlucine]

There was stock available for vega not too long ago - I bet black friday cleared a lot of the old stock, so now we've got a temporary lull while custom boards work their way through production. Note that the powercolour V64 (with the enormous cooler) is due on scan in just 4 days. It's £90 more than a decent 1080, and given stock V64 beats aftermarket 1080's it seems to fit where you'd expect given the funny price : performance ratio at the high end.

The lack of V56 so far is bad, I'm still hopeful that models will follow the (higher margin) aftermarket V64's

[DanceswithUnix]

The massive three slot format of custom Vega is a bit off-putting as well.

[MLyons]

The massive three slot format of custom Vega is a bit off-putting as well.

It's not all three slot is it? I'll have to look in the benchmarking area tomorrow.

[DanceswithUnix]

It's not all three slot is it? I'll have to look in the benchmarking area tomorrow.

The few pictures I have seen have been. On the plus side, I presume such a beast can stay on boost for longer. On the downside, you are paying for more metal and heatpipes and tbh I am more interested in a quiet card than wringing every last MHz out of the silicon at the expense of heat. If money was no object I would already have one of the water cooled cards.

[Xlucine]

XFX should be 2-slot:
http://hexus.net/tech/news/graphics/...design-photos/

Powercolour have past history (3 different links) as providing graphics cards with OTT coolers that also produce a racket, so the powercolour V64 is not likely to be the best of the bunch. Copper and aluminium aren't that expensive, and a big heatsink means you can shed similar heat with less noise if you're careful (fr'example, the asus V64)

[CAT-THE-FIFTH]

Interesting if true:

http://www.bitsandchips.it/52-englis...ready-for-2018

Raven Ridge is still MIA in Desktop (AM4 Socket) market, but it seems AMD will commercialize a Raven Ridge 2 APU in 2018 (Maybe a stopgap until Picasso, or a rushed Picasso).

We don’t know a lot, now AMD is very reserved (Even my sources have difficult to say something!), but we know that AMD is working on Raven Ridge 2 APU. It will be probably produced on 12nm of GlobalFoundries, and it will be an enhanced version of current Raven Ridge APU (Raven Ridge will be used in Embedded and Server markets, too. Just like Ryzen is used in Enterprise market via the EPYC SKUs).

It seems AMD is trying to produce a CPU (Pinnacle Ridge) and an APU (Raven Ridge 2) able to be clocked higher, dedicated to the Consumer market, waiting for Ryzen @ 7nm (2019). In our opinion, it’s a good strategy in order to put pressure on Intel and to commercialize better products dedicated to gaming, VR and content creation (Threadripper) markets.

[scaryjim]

Isn't this just "confirming" what we already know about AMD's roadmap (i.e. that there will be 12nm parts next year, followed by 7nm parts the year after)?

The only question is whether we'll ever see 14nm RR on desktop. Assuming 12nm will help increase both CPU and IGP clocks, desktop APUs have a lot to gain from the higher performance node. A desktop 14nm Raven Ridge would be up against that hard clock limit, and it would be coming less than 12 months before a 12nm desktop part should be feasible. In fact, it wouldn't surprise me at all to see 14nm RR in mobile-only parts and 12nm RR in desktop-only.

[CAT-THE-FIFTH]

Second generation Ryzen is confirmed for Q1/Q2 2018:

https://www.reddit.com/r/Amd/comment...aunch_on_12nm/



Also there is a Ryzen 3 mobile SKU and a mobile Ryzen Gaming SKU being launched.

[scaryjim]

And no mention of a desktop Ryzen APU on that slide, I note....

My best guess would be Ryzen Gaming APUs being 35W mobile parts with bumped clockspeeds, particularly for the IGP.

Wonder if the Ryzen 3 mobile APU is going to be the 2 core die. Fingers crossed it is, it comes in reasonably priced laptops, and it hits retail early in Q1 next year ... not sure how much longer my current laptop is going to keep going...

EDIT: of course, that slide only shows "premium" desktop and mobile; I guess that makes "premium" Ryzen 3 less likely to be dual core, and also possibly explains the lack of desktop APUs (which I guess wouldn't be described as "premium" desktop...)

[CAT-THE-FIFTH]

AMD does need to use defective Ryzen 5 APUs which don't have functional SMT - so I suspect 4C/4T and perhaps 6 CUs active for a total of 384 shaders??