Welcome, Guest
Username: Password: Remember me

TOPIC: Fiji, Polaris, and Vega: Idea for an investigation

Fiji, Polaris, and Vega: Idea for an investigation 2 months 2 weeks ago #15770

  • IceDreamer
  • IceDreamer's Avatar
  • Offline
  • Lurker
  • Posts: 1
  • Karma: 0
Hi everyone! I found the GN Youtube channel recently, and have been incredibly impressed by the attention to detail, scientific approach, and neutrality shown. Props to Steve and the crew!

Though we don't know the final form Vega will take yet, it's safe to say it's not looking like a brilliant advance, especially looking at power consumption. My thoughts about this turned away from NVidia, and towards AMD's own GPUs, and wondering if they've really moved.

Question: Was Vega a complete waste of time? Should AMD simply have expanded the Polaris architecture to more cores?

First some specs...

Tonga R9 380X - 2048 GCN 2.0 cores
Fiji Fury X - 4096 GCN 3.0 cores
Polaris RX570 - 2048 GCN 4.0 cores
Polaris RX580 - 2304 GCN 4.0 cores
Vega RX Vega - 4096 GCN 5.0 cores

Obviously these guys filled different price points and ran at different frequencies. However, they are similar enough for us to build a dataset.

1 - Find your most reliable, repeatable benchmark. Get a number, averaged over 5 runs per test, for 5 or 6 different locked frequencies (Try and avoid turbo/throttle) for the RX570. This allows you to form a curve to extrapolate theoretical "Benchmark result at frequency X".

2 - Repeat this process for the R9 380X and Fury X, at each frequency. This allows you to form a mathematical relationship between core count and benchmark.

3 - It's not completely solid and there's definite uncertainty, but you can now extrapolate a benchmark result for "2048 GCN 4.0 at frequency X" and "4096 GCN 4.0 at frequency X". At this point you could bring in the RX580 with its 2304 GCN 4.0 cores to sanity check the relationship between different core counts of the same architecture.

4 - Run your benchmark using RX Vega once it's released. Compare the IRL results with the extrapolated results of a hypothetical 4096 GCN 4.0 running at RX Vega's frequency, and at frequencies we know to be attainable with GCN 4.0.

Was it worth building Vega, or a waste of time? Should a simple high-clock Fury have performed better? Is there any noticeable architectural improvement from Polaris to Vega? Should AMD simply have focused on making a bigger Polaris die with GDDR5?
The administrator has disabled public write access.

Fiji, Polaris, and Vega: Idea for an investigation 2 months 2 weeks ago #15771

  • i7Baby
  • i7Baby's Avatar
  • Offline
  • Moderator
  • Posts: 590
  • Thank you received: 87
  • Karma: 18
Was Vega a complete waste of time? - Its a bit early to tell. I'm waiting for reviews and pricing. My interest is in how good it does compute - for rendering. For the time being, I'm happy with my R9 Nanos. You can't overclock them - they get too hot and come to a stop. But they game the same as a gtx980 anyway.

Should AMD simply have expanded the Polaris architecture to more cores? - I don't know if they were able to do this. Maybe they were at max cores with the GPU as is.
R7-1700 @3900/1.375V, H110iGT, ASRock x370 Taichi, GSkill 2x8GB 3200/14 @3466/14 1.38V, Samsung 250gb EVO (W10), 2TB Barracuda, 2 x R9 Nano, EVGA G2 750, Enthoo Pro, LG 34UM88. Aizo kb, M6580, HyperX Cloud
Last Edit: 2 months 2 weeks ago by i7Baby.
The administrator has disabled public write access.

Fiji, Polaris, and Vega: Idea for an investigation 2 months 2 weeks ago #15817

  • i7Baby
  • i7Baby's Avatar
  • Offline
  • Moderator
  • Posts: 590
  • Thank you received: 87
  • Karma: 18
R7-1700 @3900/1.375V, H110iGT, ASRock x370 Taichi, GSkill 2x8GB 3200/14 @3466/14 1.38V, Samsung 250gb EVO (W10), 2TB Barracuda, 2 x R9 Nano, EVGA G2 750, Enthoo Pro, LG 34UM88. Aizo kb, M6580, HyperX Cloud
The administrator has disabled public write access.
Moderators: 2ndPlayer, i7Baby
Time to create page: 0.087 seconds
Powered by Kunena Forum

We moderate comments on a ~24~48 hour cycle. There will be some delay after submitting a comment.

Advertisement:

  VigLink badge