How much better will Pascal/Arctic Islands be over GCN1.2 and Maxwell?

[angrymoose]

A lot of guys have been asking "Should I wait for Pascal or Arctic Islands, or buy a current generation Flagship (390X, 980, Fury X, 980 Ti)?"

 

It's a tough question to answer. It comes down to a lot of factors. Most importantly, what resolution are you running and what framerates are you looking to attain? Are you planning to CF or SLI, or do you want to run a single card for maximum reliability? Pascal and Arctic Islands are going to be the first process node shrink in about 4.5 years when it's finally released, and GPU's will feature new architecture and HBM2 memory. Needless to say, 2016 should be an exciting year in graphics cards.

 

I made a spreadsheet to show how former GPU flagships from nvidia and AMD have grown by generation. While we can't assume that Pascal and Arctic Islands will make the same jump as former process node shrinks, I think we can pretty safely predict that the jumps will be larger than non-process node shrink generations.

 

The graphs show pretty clearly how important a process node shrink is, and it's most well defined by looking at the AMD graph. nvidia's graph is pretty clear as well, but they've released more intermediary "Flagship" cards (280 > 285, 780 > 780 Ti, 980 > 980 Ti), which makes it slightly less pronounced (but it's still there). nvidia also has one outlier (the GTX 280), which jumped ahead by quite a bit without a node shrink, but I don't see it as much of a factor.

 

Here's the spreadsheet (if you're a redditor, you may have seen this already):

 

https://docs.google.com/spreadsheets/d/1awOqzOXrnhwgcUP7ORB4vzbtdjXwQ-2t6xiwRiYkENo/edit#gid=0

 

• Why did you take 4 hours out of your morning to build this spreadsheet?
  

I made it to explain with history the reason for the hype surrounding the forthcoming nvidia (Pascal) and AMD (Greenland) GPU's, and because an earthquake woke up my fiancee who later woke me up at 3am to tell me it was confirmed to be an earthquake, and I've been up drinking coffee since.

• "But we won't see a full fledged massive die flagship until the end of the process node's cycle"
  

WHO CARES! The biggest improvement will be to the first available flagship chip on the node, as it always has been.

• "If you look at it this way, there's always going to be a better upgrade around the corner"
  

This is admittedly true, but it doesn't tell the whole story. A node shrink is historically the biggest jump in performance, and within this spreadsheet is the statistics that prove it.

• "AMD is trash" "nVidia fanboys are sheep"
  

Don't be a dummy. Buy whatever is the best bang for the buck for you. Are you waiting for Pascal? Your best bet would be to wait for both to arrive so you can make your best purchase. If one turns out to be faster than the other, you will likely see the slower one drop in price, followed by the faster one to stay competitive. Have a look at what's happening between the Fury X and 980 Ti right now.

• "Why didn't you include Titan cards?"
  

Because they're $999 new. I limited this to single GPU 'Flagship' cards. I'm pretty sure every card on here except the Fury X is available from partners with aftermarket coolers and custom PCB's.

• "But my card can overclock way higher"
  

I built this spreadsheet with reference cards running reference clocks so that you can compare Apples to Apples within the ATi/AMD or nVidia ecosystem. Chasing down which cards were better overclockers simply takes way too much time and really has no bearing on what the future cards will be.

• Why didn't you include any benchmarks/charts/comparisons/real info?!
  

You have to click on the different sheets along the bottom. The first page is just a spec comparison.

 

If you're Linus, Luke or any of the LTT nerds (staff): Joe the janitor says hello.

[don_svetlio]

This question can only be ignored as there is NO answer.

[Sauron]

We don't have a magical link with nvidia and amd that delivers us information that is not available in any other place. The cards haven't even been presented yet, we can't know how they will perform.

[angrymoose]

We don't have a magical link with nvidia and amd that delivers us information that is not available in any other place. The cards haven't even been presented yet, we can't know how they will perform.

The information isn't there, but predictions can be made, and most likely very accurately.

 

For example, I'm betting the next Intel CPU will probably have an IPC improvement of 3-5%. That's not hard to see based on history.

[Sauron]

The information isn't there, but predictions can be made, and most likely very accurately.

 

For example, I'm betting the next Intel CPU will probably have an IPC improvement of 3-5%. That's not hard to see based on history.

 

So what would you have predicted for haswell, given that the previous architectural change gave upwards of 40% more performance than previous chips?

 

You can't rely on these predictions, they are nothing more than guesses until we have the numbers.

[xAcid9]

If both AMD and Nvidia decide not to be a douchebag and market the card as super efficient or best performance/watt, their next flagship will be at least 80% as powerful as their current one.

HD7970(Tick + Tock) is almost twice as powerful as HD6970 though not at launch.

[angrymoose]

So what would you have predicted for haswell, given that the previous architectural change gave upwards of 40% more performance than previous chips?

 

You can't rely on these predictions, they are nothing more than guesses until we have the numbers.

 

http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

 

Anandtech shows that:

• Sandy Bridge to Ivy Bridge: Average ~5.8% Up
• Ivy Bridge to Haswell: Average ~11.2% Up
• Haswell to Broadwell: Average ~3.3% Up
• Broadwell to Skylake (DDR3): Average ~2.4% Up
• Broadwell to Skylake (DDR4): Average ~2.7% Up
• Haswell to Skylake (DDR3): Average ~5.7% Up.

Predictions aren't meant to be 100% accurate, they're intended to be an estimate. If you read the charts, you'll see that the average prediction for process node shrinks for modern GPU's is 40%-60% in comparison to non-node shrink generation increases of 10%-36%. Do you really think this is out of line?

[don_svetlio]

J

 

http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

 

Anandtech shows that:

• Sandy Bridge to Ivy Bridge: Average ~5.8% Up
• Ivy Bridge to Haswell: Average ~11.2% Up
• Haswell to Broadwell: Average ~3.3% Up
• Broadwell to Skylake (DDR3): Average ~2.4% Up
• Broadwell to Skylake (DDR4): Average ~2.7% Up
• Haswell to Skylake (DDR3): Average ~5.7% Up.

Predictions aren't meant to be 100% accurate, they're intended to be an estimate. If you read the charts, you'll see that the average prediction for process node shrinks for modern GPU's is 40%-60% in comparison to non-node shrink generation increases of 10%-36%. Do you really think this is out of line?

Jump from Sandy to Ivy was around 15% and temps were reduced by 25%

[Sauron]

http://www.anandtech.com/show/9483/intel-skylake-review-6700k-6600k-ddr4-ddr3-ipc-6th-generation/9

 

Anandtech shows that:

• Sandy Bridge to Ivy Bridge: Average ~5.8% Up
• Ivy Bridge to Haswell: Average ~11.2% Up
• Haswell to Broadwell: Average ~3.3% Up
• Broadwell to Skylake (DDR3): Average ~2.4% Up
• Broadwell to Skylake (DDR4): Average ~2.7% Up
• Haswell to Skylake (DDR3): Average ~5.7% Up.

Predictions aren't meant to be 100% accurate, they're intended to be an estimate. If you read the charts, you'll see that the average prediction for process node shrinks for modern GPU's is 40%-60% in comparison to non-node shrink generation increases of 10%-36%. Do you really think this is out of line?

 

again, you cannot know. Maybe intel will give us 6 cores on the consumer platform. Maybe there's something in their plans we don't know about.

 

As for gpus, that's as random as it gets. Some generations are mostly rebrands, some are major improvements, some are architectural improvements with little performance gain (for example maxwell before the 980ti, which took 8 months to surface).

[angrymoose]

"You cannot know"

Is correct.

"You cannot roughly predict"

Is not.

[ChrisxIxCross]

We do have Nvidia's CES press conference to look forward to unfortunately I'm afraid it'll be only about their new Tegra releases/ boring car stuff.

[angrymoose]

We do have Nvidia's CES press conference to look forward to unfortunately I'm afraid it'll be only about their new Tegra releases/ boring car stuff.

I believe you're right. We may however start to hear about Pascal when nVidia's GPU Conference rolls around in April? (March?)

[ChrisxIxCross]

I believe you're right. We may however start to hear about Pascal when nVidia's GPU Conference rolls around in April? (March?)

Yes that what be GTC in the Spring which is when is they usually talk about and launch their new GPUS.