add arrow-down arrow-left arrow-right arrow-up authorcheckmark clipboard combo comment delete discord dots drag-handle dropdown-arrow errorfacebook history inbox instagram issuelink lock markup-bbcode markup-html markup-pcpp markup-cyclingbuilder markup-plain-text markup-reddit menu pin radio-button save search settings share star-empty star-full star-half switch successtag twitch twitter user warningwattage weight youtube

AMD’s New GPU - Opinions?

JustAnotherBuilderBro
  • 13 months ago

Hi there,

As many of you know CES 2019 just finished. During CES 2019, AMD took to the stage and announced their new GPU, Radeon VII, based of their 2nd gen Vega architecture. What does everyone think of it? I’d like to know everyone’s opinions.

Comments

  • 13 months ago
  • 3 points

It's a cut down compute product again with major power draw.

Apart from that the number of people getting confirmation from day one that AMD is selling these below cost and there will not be aftermarket models or supplies for launch is not good if true.

  • 13 months ago
  • 1 point

I'm going to assume that this was designed for the crypto market, and once that died they figured that their would be less egg on their face if they produced 5000 cards and then moved on to (hopefully someday) produce Navi. Or perhaps they wanted to at least keep some hype up so produced a more or less market segmented instinct (NRE might have been low if the chip came from an Instinct), even if they are selling the product for a loss).

Unless Polaris, Vega, and Navi are far more different that I've heard, they should be able to reuse a lot of the tech that went into Vega VII in making both Navi and a 7nm Ryzen ALU.

One thing that I haven't heard much of (mostly thanks to all the issues of the cost of 16GB of HMB2) is that this chip uses 300+mm2 (331mm2 is the bit number I saw) when AMD is carefully limiting zen2 chiplets to 80mm2. If they could somehow affordably make 331mm2 chips, that would mostly refute AMD's chiplet strategy for zen2 (and might imply a larger and integrated zen2+ chip). Perhaps they were willing to risk more on the Instinct GPUs (as they had less to lose and needed to close the gap more than their CPU line) than on the CPU line.

I'm guessing that if the Instinct size has decent yields (without requiring the tricks that GPUs can get away with that CPUs can't), that chiplets will only be used if AMD finds the ability to simply provide less silicon for market segmentation rather than disabling the chip. This might matter more with TSMC than a somewhat captive fab.

[gripe: Had to change my old-school FORTRAN (and Python) "raise to the power with double asterisks" to C-style carrots"]

  • 13 months ago
  • 2 points

Unless Polaris, Vega, and Navi are far more different that I've heard, they should be able to reuse a lot of the tech that went into Vega VII in making both Navi and a 7nm Ryzen ALU.

From rumors there is no Navi after AMD canned the design last year. They are using Vega architecture with a reduction of the ACE count on a monolithic chip so it can be produced and packaged conventionally.

Navi's architecture went back to the drawing board to be revamped into their Next Generation Architecture.

One thing that I haven't heard much of (mostly thanks to all the issues of the cost of 16GB of HMB2) is that this chip uses 300+mm2 (331mm2 is the bit number I saw) when AMD is carefully limiting zen2 chiplets to 80mm2. If they could somehow affordably make 331mm2 chips, that would mostly refute AMD's chiplet strategy for zen2 (and might imply a larger and integrated zen2+ chip). Perhaps they were willing to risk more on the Instinct GPUs (as they had less to lose and needed to close the gap more than their CPU line) than on the CPU line.

Multiple sources have confirmed that AMD is selling these below cost.

RTG cannot continue to hemorage market share for another year and remain viable in the discrete market.

Zen 2 is limited in size also because they cannot make it any larger and still use the AM4 package with how large the I/O package is and they need for space to assemble everything.

I'm guessing that if the Instinct size has decent yields (without requiring the tricks that GPUs can get away with that CPUs can't), that chiplets will only be used if AMD finds the ability to simply provide less silicon for market segmentation rather than disabling the chip. This might matter more with TSMC than a somewhat captive fab.

Chiplets and GPU are dead for consumer use, AMD ended that with Navi.

The innate latency cripples performance and most modern software views it as multiple GPU in Crossfire/SLI instead of a single chip. Maybe in 5-10 years it might be viable but currently the market just isn't ready.

  • 13 months ago
  • 1 point

Zen 2 is almost certainly defined by the needs of Epyc, and the I/O chip was designed to fit AM4 with the Zen2. I'd still expect them to design a chiplet containing a GPU for an ALU: with the current design they can build an ALU out of a CPU+GPU chiplet (only requires a new GPU chiplet), a whole new CPU+GPU chiplet (requires just as much design, and doesn't do anything the current line doesn't do), or build an entire chip out of 7nm (same issues, but even more expensive). I'd bet on a monolithic GPU chiplet that fits in the AM4 space.

I'd expect a scaled up edition of whatever is in the current ALU (called "Vega" but rumors claim it is closer to Polaris) in the next ALU. And I'd expect the ALU to be far more important to AMD and it's future (i.e. notebooks) than discrete graphics. [It's been awhile since I dug through the OpenCL optimization books to know the differences between graphics chips. I'm not even sure that study included Polaris, so I'm not really sure of the differences between Vega and Polaris].

I could see an innovative design in a PS5 changing modern graphic design methods (as you suggest Navi was supposed to do). Not an option for AMD discrete cards (even nvidia will take a few generation to get RTG to be important).

  • 13 months ago
  • 2 points

I'd still expect them to design a chiplet containing a GPU for an ALU: with the current design they can build an ALU out of a CPU+GPU chiplet (only requires a new GPU chiplet), a whole new CPU+GPU chiplet (requires just as much design, and doesn't do anything the current line doesn't do), or build an entire chip out of 7nm (same issues, but even more expensive). I'd bet on a monolithic GPU chiplet that fits in the AM4 space.

They already killed that during the after launch Q&A, Ryzen isn't getting a GPU chiplet.

The APU will feature mobile silicon again built on a more conventional design to save space which is needed on mobile parts.

I could see an innovative design in a PS5 changing modern graphic design methods (as you suggest Navi was supposed to do).

Giving Infinity Fabric use costs 20-30% performance from latency alone I think they will go stacked like Intel is doing.

  • 13 months ago
  • 1 point

They gonna sell only 5000 of them so very few people in the consumer market will even have one.

It was a disappointment to be honest

  • 13 months ago
  • 1 point

Yea thats what i was thinking tbh

  • 13 months ago
  • 1 point

That 5000 limit is a rumor.

  • 13 months ago
  • 2 points

It's more of a realistic rumor as they are stating that there won't be aftermarket designs and they're selling them at a lose.

  • 13 months ago
  • 1 point

How is it realistic? And you mean at a loss.

  • 13 months ago
  • 1 point

As in they're literally losing money from selling them in the first place, why make more of them if it'll earn them no profit at all?

  • 13 months ago
  • 1 point

That remains to be seen. The cost of production is not so high that they're going to sell those at such losses. Even when covering advertising and shipping costs. They may not profit as much, but I doubt they'd be sold at actual losses.

  • 13 months ago
  • 1 point

If i recall the Vega 56 / 64 also launched at higher than MSRP and was shorted out quick

  • 13 months ago
  • 1 point

Sure was.

  • 13 months ago
  • 1 point

I don't like it at all.

  • 13 months ago
  • 1 point

Why’s that?

  • 13 months ago
  • 1 point

I'm actually excited AMD will be competing in the high end. Definitely upgrading from my Vega 64 to the Radeon VII.

  • 13 months ago
  • 1 point

It looks like the board promised last summer "in 2017". Except that was pro-user (non-gaming) only. So not only is this new Vega late, they are launching it as a "gaming" board which implies that any Navi board is even later.

As a gaming board it is said to be on a par with a 2070 sans ray-tracing and machine learning anti-aliasing (I suspect the later will be useful), but with freesync (but it looks like real 2070s will effectively have freesync soon). This makes it pretty pointless.

As a compute board (which was pushed pretty hard by AMD), it is probably can perform plenty of more operations than the equivalent nvidia board, but the nvidia CUDA ecosystem is sufficiently better that you really wouldn't bother unless you have to make a really odd custom job that really doesn't have a lot of CUDA foundation behind it (i.e. nothing machine learning or anything else popular with GPUs).

It looks like AMD engineers are still spending all their time on Ryzen (hey, it saved the company. Don't complain too much). Lets hope they get some real improvements on that.

  • 13 months ago
  • 1 point

AMD is disabling some of the compute instructions to keep people from buying these over the instinct accelerator products they are based on.

  • 13 months ago
  • 1 point

I missed any notice of what they shipped in [correction] 2018. I'm not even sure it is available outside of datacenter customers, and that might keep them worrying about cutting down the GPU too much (especially since datacenter customers would have trouble making plans around buying a significant fraction of the 5000 cards).

I'd assume that it might sell as a double precision monster, but as little else (and even then would have trouble against volta and the nvidia infrastructure).

  • 13 months ago
  • 1 point

FP64 is disabled which really limits their desirability for a cheap way out.

  • 13 months ago
  • 1 point

I really have to wonder why anyone would buy it then. Did they at least keep fp16? Nvidia seems to own the machine learning market thanks to their software base, I'd at least hope to try to use fp16 for quick and dirty graphics (like the old 24bit ATI GPUs) and hope they didn't disable that as well.

  • 13 months ago
  • 2 points

They kept FP16 but its running AMD fashion so mileage and support varies.

Supposedly they are trying to leverage it for a DLSS attempt but so far have no adoption or planned adoption.

It's the same with AMD Hybrid Ray Tracing since they open sourced it no adoption.

  • 12 months ago
  • 1 point

FP64 was supposed to be cut down to 1/8th of FP32 (not exactly "disabled", but still pretty much nerfed). They just released it at 1/4 of FP32 (the Instinct it gets the silicon from can do 1/2 of FP32).

1/4 FP32 is an interesting choice. It is roughly what you should get if you try to design a system that maximizes both FP32 and FP64 (probably more thanks to power issues, I suspect the transistors aren't holding things back anymore). Of course, I think actual systems designed for both FP32 and FP64 really gave 1/6 FP64 (and I'd be very happy if Navi provided such).

  • 12 months ago
  • 1 point

It might be neither getting priority since NVidia won the DXR race.

If Sony sides with NVidia as well AMD may very well have to focus their compute performance into RT cores and build the main graphics cores around throughput instead of compute.

Now that Custom Silicon was rolled into RTG they share graphics IP.

  • 13 months ago
  • 1 point

From the limited comparisons, it's on the 2080 level. 2070 actually competes against the Vega 64, though it Beats that card in most titles.

  • 13 months ago
  • 1 point

Not too disappointing, AMD have been kinda middling the last generation with their GPU's. I expect it will trade blows with the RTX 2080 on some titles while other titles it will likely approach RTX 2060/2070 territory. Not expecting anything different to Vega 64 competing against the 1070/1080/1080Ti. Now if Radeon VII was a $500 card, then fantastic, a value for money option for a fairly high tier GPU. At $700, it will be harder to justify a purchase. Now this is all guesswork. I take AMD's benchmark with a grain of salt, potentially rigged i.e. card OC to max verses stock clock Nvidia. Benchmarking is only a guide anyway, performance in a real titles are what counts. We shall see when it releases, but potentially, setting the price at $700, AMD will wreck their chances of gaining any inroads in the GPU market this gen. Only thing going for them is their GPU will feature on next gen Xbox/PS5 consoles.

  • 13 months ago
  • 1 point

Part of me is disappointed with the notion that it's priced to match an RTX 2080, and is only matching it. Rather than either being able to undercut the 2080 or surpass it, and be somewhere between a 2080/Ti, for instance.

On another, one game that I think would want that extra vRAM is Final Fantasy XV. Yes, the game is an NVIDIA leaning title with as many such technologies supported. That said, up to this point, the only GPU's that have been able to actually run the game at 4K and Very High (if not Max) settings, have all been Titans with 16GB(+) vRAM - such as the Titan V or RTX. This might be the first such GPU able to do it for less than 4-digit pricing. That might be a use case to consider this card seriously.

It's probably just me, but I was hoping for more of a direct answer to what nvidia had at the flagship level. Certainly not needing anything competing directly on the Quadro or Titan level specifically, but yeah... something between a 2080/Ti at the $700 asking price would have absolutely been a market disruptor, and force nvidia to compete on price.

  • 13 months ago
  • 1 point

Better wait for NAVI.

  • 13 months ago
  • 1 point

With AMD outright saying that the rest if the year is refresh either Navi is a refresh of Vega 56/64 or not coming till next year.

  • 13 months ago
  • 1 point

New report from https://www.google.com/amp/s/wccftech.com/four-amd-navi-gpu-variants-leaked-rumored-for-july-2019-launch/amp/

Says Navi may beon par or close to Radeon VII but cost half the price. This confused, excites and frustrates me as I want to get the VII lol

  • 13 months ago
  • 2 points

Its WCCFTech.

And if they were right 10% of the time AMD would have triple the performance per watt they currently have.

  • 13 months ago
  • 1 point

So do you believe this as a possibility or just a way if?

  • 13 months ago
  • 2 points

"Navi" has already been confirmed as a Polaris replacement.

AMD confirmed Navi and its entire design was scrapped when Raja defected to Intel as the design was unworkable with current software and we were looking at maybe a decade before it was feasible.

AMD hired nobody to work on Navi.

AMD has made multiple Vega teams and hires for Vega.

Navi from its reveal was intended to take Vega's architecture and scale it across the entire product stack.

The hope for it being amazing is people wishing for another option for anything more then budget-mainstream range where you have to choose between lower prices but hotter running power hungry cards and whatever NVidia is currently selling.

  • 13 months ago
  • -1 points

AMD failed us this year. RTX is taking over. And now that the 2060 is out. I think it’s safe to say game over

  • 12 months ago
  • 1 point

You gave up on AMD on Jan 20? Seems a bit early to declare the year a failure (although it looks like Intel will be AWOL this year on CPUs thanks to needing 10nm working for their next microarchitecture).

They can't afford to make a card everybody wants to buy that is being sold for less than cost, so don't expect a great value. There was at least one report that insisted that for individual video editing, 16G made a huge difference with editing 4k video streams and that this was the card you want. I also suspect that the 1/4 double precision floating point will find interested buyers.

Sort

add arrow-down arrow-left arrow-right arrow-up authorcheckmark clipboard combo comment delete discord dots drag-handle dropdown-arrow errorfacebook history inbox instagram issuelink lock markup-bbcode markup-html markup-pcpp markup-cyclingbuilder markup-plain-text markup-reddit menu pin radio-button save search settings share star-empty star-full star-half switch successtag twitch twitter user warningwattage weight youtube