RX 480 Decision to have 2304 vs 2560 SP?

[chris_kick]

Hi Guys,

One thing that has bugged me since the release of the RX 480 is the fact that it has 2304 SP instead of a more even number like 2048 of the R9 2/380x or 2560 of the R9 2/390.

Obviously we all hoped for a R9 380 type of card where the full die would be released as the R9 380X. (RX 480X?)

While the RX 480 isn't in the same price point as the GTX 1060 3/6GB many consumers and reviewers pit these cards together - with the RX 480 coming of worse.

On most benchmarks the GTX 1060 is ahead by a small margin (6-10%) except for dx12 titles where they trade wins. If AMD/RTG had instead placed 2560 SP, the card would have beaten the GTX 1060 or reached much closer parity in many benchmarks.

An example of this is the average summary GPU Userbenchmark where it shows the GTX 1060 6GB leading by 6-10%.
A version of the RX 480 with 2560SP (11% more SP) would have likely made up this ground.

My personal thoughts

AMD/RTG under-estimated the clock speeds that NVIDIA could obtain on Pascall in workable TDP on TSMC's 16nm. [Compared to Maxwell, the clocks surprised everyone reviewers, consumers, AMD]

AMD/RTG over-estimated their clock speeds that they could obtain in their workable TDP on GF's 14nm.
Upon obtaining leaks and the release of the Pascall series, they moved to up clock their cards to the limit.
[Stock/boost clocks of 1266mhz are almost at peak that the chips can handle, almost no overclocking room, reducing clocks down heavily reduces TDP]

How much influence did Sony with the PS4 Pro have with determining chip sizes VS AMD/RTG's vision?

Summary thoughts

If AMD/RTG chose 2560SP instead of 2304SP at the same clockspeeds of 1266mhz or similar;

The RX 480 would be at parity with the GTX 1060 6GB- heavily boosting sales, market value, brand, reviews

The RX 480 would also be at parity with the R9 390X, boosting morale that the company isn't releasing a card slower than their 2 gen old R9 390X - A thought I know many of us had upon seeing the Reviews

It would also create more options with the RX 470 and a theoretical RX 465.
(This entire discussion could also be had on the RX 460 being 860Sp instead of full 1024SP and losing the GTX 1050 and 1050ti)

Also, the PS4 Pro would be that tiny bit more powerful.

While I think the RX 480 is a great card with a great price point, I am interested in your thoughts of their SP selection.

 

[Dave2150]

Even if the RX480 had 3000 shaders and was 10% ahead of the 1060, the 1060 would still outsell it.

Why? NVIDIA. Their mindshare is monumental. People have been buying 'Geforce' for many years, and have been led to believe that AMD Radeon = cheap, nasty, hot, bad drivers, power guzzling etc.

The majority of gamers out there simply don't know that AMD drivers have been absolutely rock solid for 10+ years for single GPU's.

It's a sad situation, which will require AMD to execute flawlessly for many years to rectify and convince the majority to swap.

 

[Ancalagon44]

First, more recent benchmarks have the RX480 and GTX1060 much closer together. AMD's drivers have improved the performance of the RX480 to the point where even in DX11, it is about as fast as a GTX1060.

Second, I think AMD knew well in advance what clocks they could achieve - or at least, they knew with 95% certainty how high they would be able to raise clocks.

I don't know know how AMD arrived at the 2304 figure, but if I had to guess, it had more to do with where they wanted to position the resulting product than anything else. In other words, their starting point was where do they want to position Polaris 10 in the market, in terms of cost, power consumption and performance. They had a set of requirements, essentially, those being that power consumption must be 150W or less, performance must equal or beat R290 performance, and cost must be $200.

With power consumption of 150W or less, there are only a certain number of ways that you can get performance that is equal to or better than an R290 while costing (the consumer) less than $200. Obviously 2304 was the best fit solution to the problem. For instance, a 2560 SP GPU coudl have been clocked slightly lower to hit the same performance, but it would have cost more, and made it more difficult for AMD to hit that $200 price point.

 

[poofyhairguy]

Honestly the 480 has snuck up on the 390X since its release. Hawaii has had trouble with recent titles compared to the 480. I wouldn't be surprised it it outright beats it.

I sold my 390X because I couldn't tell a lick of difference between it and my 480.

 

[Valantar]

The RX 480 was designed to showcase that AMD could produce a budget-friendly VR card while staying semi-competitive in terms of thermals and TDP, which was by far their biggest achilles' heel the previous generation. A 10% increase in SPs would bring with it a 10% increase in die power usage (which for an RX 480 is ~120W), pushing the card beyond both what's possible to power with a single 6-pin connector (which would look bad - "Oh, AMD's so power hungry their midrange card needs an 8-pin!") and would require more complex and expensive stock coolers. The RX 480 stock blower already struggles with its 150W, and increasing that to 160+ would not be good - requiring heatpipes or some other more expensive heat dissipation method, which would again increase costs (as would a bigger die).

On the other hand, it's pretty clear that (knowing how they were pushing driver development at the same time) AMD knew the card would grow to perform better than it did at launch. While they're still not quite on par with the 1060, the 480s out there have consistently gained on the green team's offerings in DX11, while tieing or beating them in DX12 and Vulkan.

As such, I think they made exactly the right decision: making the most powerful card they could fit into a single-6-pin-powered card with a cheap stock cooler, thus showing that AMD wasn't that far behind on power efficiency while keeping their position as the bang-for-your-buck king.

 

[crisium]

Well, you shouldn't use the 290 in your comparison - that's a cut down chip. It's the 290X that's the full, and it's 2816. Why not 2560 or 3072? The 780 Ti is a full chip at 2880 - why not 3072?

I assume you view these as less "even" numbers. But I'm not sure how you determine what numbers makes sense. If you're going off powers of two, then 1024, 2048, and 4096 make sense and everything else isn't "even". Why is 2560 more "even" than 2304?

 

[Valantar]

Well, you shouldn't use the 290 in your comparison - that's a cut down chip. It's the 290X that's the full, and it's 2816. Why not 2560 or 3072? The 780 Ti is a full chip at 2880 - why not 3072?

I assume you view these as less "even" numbers. But I'm not sure how you determine what numbers makes sense. If you're going off powers of two, then 1024, 2048, and 4096 make sense and everything else isn't "even". Why is 2560 more "even" than 2304?

I guess 'cause it's closer to 2500, and ends in a zero? Also, because it would be 40 CUs?

Possible amounts of SPs based on AMD CUs:
1: 64
2: 128
3: 192
4: 256
5: 320
6: 384
7: 448
8: 512
9: 576
10: 640
11: 704
12: 768
13: 832
14: 896
15: 960
16: 1024
17: 1088
18: 1152
19: 1216
20: 1280
21: 1344
22: 1408
23: 1472
24: 1536
25: 1600
26: 1664
27: 1728
28: 1792
29: 1856
30: 1920
31: 1984
32: 2048
33: 2112
34: 2176
35: 2240
36: 2304
..... and so on.

Should they make a 25CU GPU just because it would have 1600 SPs? Well, I think targeting a performance and power number and selecting the fitting number of CUs is the better approach ...

 

[poofyhairguy]

I just can't figure out why it's taking AMD so long to put out a double Polaris with the Vega label slapped on.

 

[Valantar]

I just can't figure out why it's taking AMD so long to put out a double Polaris with the Vega label slapped on.

Because single-card dual GPU setups are ... not a good idea. Cooling is invariably bad (seriously, cooling more than 250W in a single card is nigh on impossible), boards are expensive, and so on. The cards need to be huge too, limiting them to only the biggest cases. So, what would you get? A card that's at least as expensive as two 480s, at best performing as well as two 480s (although likely worse due to power and thermal limits in a single-card configuration), with all the issues that stem from CF setups. Can't say I see much reason for this when it's dead easy to just buy two off-the-shelf 480s ...

Also, the last thing AMD wants is to sully the Vega name with a 300W behemoth of a card. AMD has enough of a reputation for being the inefficient/power hungry brand. That would only make things worse. Not to mention that it would look to many as an admission that AMD can't compete with high-end single-GPU solutions from Nvidia. They're far better off holding off on the high-end cards (while being very competitive in the mid-range) until they can get them right.

 

[poofyhairguy]

I guess you are right, but the alternative was not competing at the high-end for a whole generation of GPUs. At some level a bad offering is better than no offering just to keep an AMD card at the top of benchmark charts.

I just can't understand AMD's GPU business anymore.

 

[Valantar]

I guess you are right, but the alternative was not competing at the high-end for a whole generation of GPUs. At some level a bad offering is better than no offering just to keep an AMD card at the top of benchmark charts.

I just can't understand AMD's GPU business anymore.

I wholeheartedly disagree. The Polaris generation has been a big PR win for AMD, and not competing in the high end hasn't hurt them nearly as much as delivering a card using nearly 2x the power (300W for 2x 480 vs 180W for the 1080) for at-best matching performance - and 50% lower performance in unoptimized titles. That would be a PR faceplant of epic proportions. AMD has been very smart to not feed the haters more than needed, keeping their cool while putting their time and money into developing Vega. Let's just hope it's competitive.