AMD Unveils More Info About Vega

[DocSwag]

Source: http://www.anandtech.com/show/11002/the-amd-vega-gpu-architecture-teaser

Quote

First and foremost, today’s detail release is a teaser, not a deep dive, or even a preview. AMD is only releasing a few details about Vega, and those are being kept at a high level. In fact it’s fair to say that there’s just enough information to answer little and raise even more questions; just what a proper teaser should be.

Why? Well part of the reason is that we’re still months off from the launch of Vega. I believe it’s fair to say that by announcing a first-half of the year launch date when we’re already in 2017 is a strong indicator that Vega will not launch until later in that window, likely some time in Q2. So we’re still a good three to five months out from the launch of Vega, which means AMD doesn’t want to (or need to) release too many details this far out. Rather they can trickle out chosen details for maximum impact.

Quote

So what is AMD looking to do with Vega? Besides aiming for the high-end of the market, AMD is looking at how the market for GPUs has changed in the last half-decade, and what they need to do to address it. Machine learning is one part of that, being a market that has practically sprung up overnight to become a big source of revenue for GPUs. This is where the previously announced Radeon Instinct will fit in.

Quote

Meet the NCU, Vega’s next-generation compute unit. As we already learned from the PlayStation 4 Pro launch and last month’s Radeon Instinct announcement, AMD has been working on adding support for packed math formats for future architectures, and this is coming to fruition in Vega.

With their latest architecture, AMD is now able to handle a pair of FP16 operations inside a single FP32 ALU. This is similar to what NVIDIA has done with their high-end Pascal GP100 GPU (and Tegra X1 SoC), which allows for potentially massive improvements in FP16 throughput. If a pair of instructions are compatible – and by compatible, vendors usually mean instruction-type identical – then those instructions can be packed together on a single FP32 ALU, increasing the number of lower-precision operations that can be performed in a single clock cycle. This is an extension of AMD’s FP16 support in GCN 1.2 & GCN 4, where the company supported FP16 data types for the memory/register space savings, but FP16 operations themselves were processed no faster than FP32 operations.

Quote

Moving on, the second thing we can infer from AMD’s slide is that a CU on Vega is still composed of 64 ALUs, as 128 FP32 ops/clock is the same rate as a classic GCN CU. Nothing here is said about how the Vega NCU is organized – if it’s still four 16-wide vector SIMDs – but we can at least reason out that the total size hasn’t changed.

Quote

That said, I do think it’s important not to read too much into this on the last point, especially as AMD has drawn this slide. It’s fairly muddled whether “higher IPC” means a general increase in IPC, or if AMD is counting their packed math formats as the aforementioned IPC gain.

Quote

But with Vega however, it looks like that limitation has finally gone away. AMD is teasing that Vega offers an improved load balancing mechanism, which pretty much directly hints that AMD can now efficiently distribute work over more than 4 engines. If so, this would represent a significant change in how the GCN architecture works under the hood, as work distribution is very much all about the “plumbing” of a GPU. Of the few details we do have here, AMD has told us that they are now capable of looking across draw calls and instances, to better split up work between the engines.

 

This in turn is a piece of the bigger picture when looking at the next improvement in Vega, which is AMD’s geometry pipeline. Overall AMD is promising a better than 2x improvement in peak geometry throughput per clock. Broadly speaking, AMD’s geometry performance in recent generations hasn’t been poor (it’s one of the areas where Polaris even further improved), but it has also hurt them at times. So this is potentially important for removing a bottleneck to squeezing more out of GCN.

Quote

Now for Vega, AMD is back again with support for the next generation of HBM technology, HBM2. In fact this is the very first thing we ever learned about Vega, going back to AMD’s roadmap from last year where it was the sole detail listed for the architecture.

Quote

We’ll suitably round-out our overview of AMD’s Vega teaser with a look at the front and back-ends of the GPU architecture. While AMD has clearly put quite a bit of effort into the shader core, shader engines, and memory, they have not ignored the rasterizers at the front-end or the ROPs at the back-end. In fact this could be one of the most important changes to the architecture from an efficiency standpoint.

 

Back in August, our pal David Kanter discovered one of the important ingredients of the secret sauce that is NVIDIA’s efficiency optimizations. As it turns out, NVIDIA has been doing tile based rasterization and binning since Maxwell, and that this was likely one of the big reasons Maxwell’s efficiency increased by so much. Though NVIDIA still refuses to comment on the matter, from what we can ascertain, breaking up a scene into tiles has allowed NVIDIA to keep a lot more traffic on-chip, which saves memory bandwidth, but also cuts down on very expensive accesses to VRAM.

 

For Vega, AMD will be doing something similar. The architecture will add support for what AMD calls the Draw Stream Binning Rasterizer, which true to its name, will give Vega the ability to bin polygons by tile. By doing so, AMD will cut down on the amount of memory accesses by working with smaller tiles that can stay-on chip. This will also allow AMD to do a better job of culling hidden pixels, keeping them from making it to the pixel shaders and consuming resources there.

So AMD ended up giving us quite a bit of info about Vega today. Not only did we learn about the new NCUs, but AMD has also made a lot of changes in Vega including a change in the ROPs that might indicate a move to tile-based rasterization, similar to what Nvidia has been doing since Maxwell. I was personally expecting Vega to be mostly just a larger Polaris with HBM2 but it seems that Vega might instead turn out to be a very large revision to GCN. This seems pretty exciting and I'm excited to see just how efficient and powerful Vega really is.

 

Also a note: AMD showed a video of Vega running Doom at 4k on Ultra with around 70 fps on average looking by the framerate counter on the top right. You can see the video here:

I'm assuming that they're running Doom with Vulkan, not OpenGL. The Fury X gets around 55 fps average on DOOM at 4k with ultra settings on Vulkan, so this means Vega is ~25% better than the Fury X, which is a pretty good increase. What are your thoughts?

[Nena Trinity]

This is tremendously great but why so much hype about a Volta beater?

[DocSwag]

4 minutes ago, Nena360 said:

This is tremendously great but why so much hype about a Volta beater?

I don't think Vega is supposed to beat Volta. That's Navi's job. Vega is just a high end Pascal competitior.

[deviant88]

As long as i can afford it its good.

[Guest Kloaked]

Went from Graphics Core Next to Next-gen Compute Unit.

 

These names, man.

[Nena Trinity]

12 minutes ago, DocSwag said:

I don't think Vega is supposed to beat Volta. That's Navi's job. Vega is just a high end Pascal competitior.

I agree it is only because it says "poor Volta(sticker)" on the trailer... o3o

[Orcblood]

I think with AMD software catching up, this tile based Rasterizer thing, and dx 12 and volcan Amd will outshine Nvidia.

 

 

BTW I love amd control panel that is based on the QT framework. Faster and looks good. The only thing they aren't doing right is their events. This time no live stream. Last time for ZEN no framerate counter in BF1, no price for zen, and no release date. 

[VerticalDiscussions]

lol

[Guest Kloaked]

Just now, jman629 said:

I think with AMD software catching up, this tile based Rasterizer thing, and dx 12 and volcan. Amd will outshine NSH*TIA. 

Your post gave me a stroke.

[MageTank]

If the performance in doom is any inclination on how this thing performs in other titles, this is great. While it won't be enough to answer the Titan XP (or it's rumored cut-down offspring), it will be a definitive answer to the GTX 1080 and 1070. Especially if you compare the results to the previous 4k Vulkan benches:

 

Nvidia: 

 

AMD:

 

That's a solid 28% boost over the Fury X, and 11% over the GTX 1080. Again, this is just in Doom, so results will most definitely vary, but it's still good news nonetheless. We just might see some competition in the high-end before Volta after all. 

[MageTank]

2 minutes ago, ONOTech said:

Doing some very rough math, AMD gets a 20% increase using Vulkan over OpenGL in DOOM. 70 FPS/1.20 = 58.33 which puts it about 8% behind the 1080 in OpenGL. 

With the change to tile-based rasterization, that might change. Especially if AMD is done with ROP-limiting their hardware, like the Fury X. 

[captain cactus]

7 minutes ago, MageTank said:

If the performance in doom is any inclination on how this thing performs in other titles, this is great. While it won't be enough to answer the Titan XP (or it's rumored cut-down offspring), it will be a definitive answer to the GTX 1080 and 1070. Especially if you compare the results to the previous 4k Vulkan benches:

Problem here is that the 1070/1080 have been on the market for months now. People who waited for a Vega card have most likely bought a 1070/1080 right now. Hyping things up and then staying silent will only lose you customers. By the time Vega hits the stores Pascal will have been out for more than a year at this rate. There's no way Vega is going to win the market share back, not with nVidia also having Pascal refreshes planned.

 

AMD's move with the RX480 launching first might have been a smart one on paper, but the money's made in the high-end market, not the mid-range market. AMD has nothing in the high-end save for a last-gen power hungry Fury or Fury X on sale.

[Techicolors]

8 minutes ago, lots of unexplainable lag said:

Problem here is that the 1070/1080 have been on the market for months now. People who waited for a Vega card have most likely bought a 1070/1080 right now. Hyping things up and then staying silent will only lose you customers. By the time Vega hits the stores Pascal will have been out for more than a year at this rate. There's no way Vega is going to win the market share back, not with nVidia also having Pascal refreshes planned.

 

AMD's move with the RX480 launching first might have been a smart one on paper, but the money's made in the high-end market, not the mid-range market. AMD has nothing in the high-end save for a last-gen power hungry Fury or Fury X on sale.

I had a 1070 since July last year (well technically August because I've RMA'd it) but I do not mind moving to Vega if it finally enters the market and is promising. but now that I've learned CES isn't the release window, a 1070 is satisfactory of what I throw at it, least for a while. with that said I'm still not sold on tech such as G-sync, while it can make a lot of the types of games I play much more palpable. 

[TheRandomness]

Oh here's a new thing I'd like to see that AMD could give without revealing too much. Is it going to be stack like:

RX 485 (small cut down Vega), RX 490 (small Vega), RX 495 (big cut down-ish Vega), RX 500 (big Vega)? Or will it skip straight to the 500 series?

[Mira Yurizaki]

19 minutes ago, MageTank said:

With the change to tile-based rasterization, that might change. Especially if AMD is done with ROP-limiting their hardware, like the Fury X. 

This depends on what else they did with the architecture. Tile based rasterization only saves you bandwidth requirements.

[MageTank]

Just now, M.Yurizaki said:

This depends on what else they did with the architecture. Tile based rasterization only saves you bandwidth requirements.

True, but at 4k, bandwidth is a pretty big deal. I am more interested in seeing the TMU/ROP setup. For some reason, AMD has been ROP limiting their flagships lately. Anyone know if these Vega flagships use HBM or not?

[cj09beira]

5 minutes ago, TheRandomness said:

Oh here's a new thing I'd like to see that AMD could give without revealing too much. Is it going to be stack like:

RX 485 (small cut down Vega), RX 490 (small Vega), RX 495 (big cut down-ish Vega), RX 500 (big Vega)? Or will it skip straight to the 500 series?

they are going straight to 500 series i believe

[cj09beira]

1 minute ago, MageTank said:

True, but at 4k, bandwidth is a pretty big deal. I am more interested in seeing the TMU/ROP setup. For some reason, AMD has been ROP limiting their flagships lately. Anyone know if these Vega flagships use HBM or not?

its 2 stacks of hbm2 512GB/s

[MageTank]

1 minute ago, cj09beira said:

its 2 stacks of hbm2 512GB/s

So it most definitely won't be bandwidth starved, at least not without copious amounts of AA stacked on top of obscene resolutions. By then, it would be out of raw horsepower, lol. 

[Mira Yurizaki]

2 minutes ago, cj09beira said:

its 2 stacks of hbm2 512GB/s

Same as Fury X's HBM1

 

I mean, I guess 512GB/s is faster than 384GB/s Titan XP employs, but calling it "high bandwidth cache" is like calling DDR4 RAM "L4 CPU cache"

[marldorthegreat]

36 minutes ago, MageTank said:

That's a solid 28% boost over the Fury X, and 11% over the GTX 1080. Again, this is just in Doom, so results will most definitely vary, but it's still good news nonetheless. We just might see some competition in the high-end before Volta after all. 

You cant compare tests like that, you dont know what part of the game was tested, what settings were used, which cpu was used etc, Comparing benchmarks is useless. 

[MageTank]

1 minute ago, M.Yurizaki said:

Same as Fury X's HBM1

 

I mean, I guess 512GB/s is faster than 384GB/s Titan XP employs, but calling it "high bandwidth cache" is like calling DDR4 RAM "L4 CPU cache"

Hey, my overclocked DDR4 competes with Broadwells EDRAM (L4) in terms of raw bandwidth, I just don't know if I can match it's latency or not, lol. Mine is at 37.8ns. 

[MageTank]

1 minute ago, marldorthegreat said:

You cant compare tests like that, you dont know what part of the game was tested, what settings were used, which cpu was used etc, Comparing benchmarks is useless. 

True, which is why we always wait until the product is released before speaking in absolutes. My context was completely hypothetical. 

[TheRandomness]

1 minute ago, MageTank said:

So it most definitely won't be bandwidth starved, at least not without copious amounts of AA stacked on top of obscene resolutions. By then, it would be out of raw horsepower, lol. 

I vote for it going to have 4096 SPs or more, judging by the core die size appearing a bit smaller (or the same size) than the one on the Fury X combined with the 14nm process. If it has 4480 SPs, then I think it's going to be packing a punch running at around ~1100-1200Mhz

[Mira Yurizaki]

9 minutes ago, MageTank said:

Hey, my overclocked DDR4 competes with Broadwells EDRAM (L4) in terms of raw bandwidth, I just don't know if I can match it's latency or not, lol. Mine is at 37.8ns. 

According to http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/3, the eDRAM on Iris Pro on Haswell is 50GB/s one way (so 100GB/sec aggregate). The fastest RAM I see on Newegg is about 34GB/sec. o3o