• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

AMD RDNA3 unveiling event

Soldato
Joined
28 May 2007
Posts
10,078

Bare in mind i got mine for £450 with 3 games so £400 after game sales. It's a Red Devil as well so proper bargain at the time compared to this day and age. I also got fine wine and undervolt :cry: :cry: :cry: :cry: :cry:. You guys should be applauding me, i ain't the problem supporting all the bs prices. it's damn torture trying not to waste my money on an upgrade. Sitting here with two consoles that have more gpu power than my PC. This should not be a thing.
 
Last edited:
  • Haha
Reactions: TNA
Associate
Joined
11 Aug 2005
Posts
414
Location
Cambridge
from what i see amd will have to offload more of the batching and scheduling work in software..
I haven't seen anything that points to this. Yes, AMD changed their processing resources this generation, much like they do most generations, but there's no reason they need to offload significantly more work into software - they just need to optimise the existing stuff, much like they (and Nvidia) do most generations. The only change I can tell is the devs/drivers probably ought to try and up the parallellisation to better take advantage of the dual FP32 units we've now got (but this is similar to the work they had to do to take advantage of Nvidia's equivalent dual FP32 in ampere).


this needs a hardware revision like increasing the number of stages, expanding instruction buffers and that sort of thing..
It's done in the same revision of hardware as the changes to the resources, you don't wait for the next revision. Much of the changes to caches/registers/schedulers are there for that reason.


amd has been using a simd architecture and relies on vectorization to manage workloads and utilization unlike nvidia which further adds to complexity.. its like they have extended cpu design philosophies to the gpu
All GPUs manage workloads/utilisation in similar ways - AMD, Intel and Nvidia all use SIMD - and have done for many generations: you group stuff together and perform operations on groups of things at the same time. What they do differently, and what changes between generations, is how long those operations take and what things you can do at the same time.

CPUs on the other hand are traditionally not very SIMD - it's very much a GPU design philosophy, but of course, we now have greater SIMD on CPU these days.
 
Last edited:
Associate
Joined
19 Sep 2022
Posts
543
Location
Pyongyang
I don't see how Nvidia has to rely on extensive vectorization. Their alus are special purpose. For example Nvidia's cuda cores only do FMA in FP32 or int32 data formats in a given clock cycle. FMA by definition is an simd operation but i am talking about more extensive packing of data and instructions. Amd on the other hand uses a more general purpose alu that can do many things other than FMA.

Parallelization in vector form is a combinatorially complex problem. Unlike Nvidia which uses relatively dumber special purpose units for parallelization amd might need an altogether different level of sophistication. For now it looks like software fallback.

And on your second point, that's a speculation that i have is that amd didn't rightsize the hardware for the intended throughput
 
Last edited:
Associate
Joined
19 Sep 2022
Posts
543
Location
Pyongyang
I haven't seen anything that points to this. Yes, AMD changed their processing resources this generation, much like they do most generations, but there's no reason they need to offload significantly more work into software - they just need to optimise the existing stuff, much like they (and Nvidia) do most generations. The only change I can tell is the devs/drivers probably ought to try and up the parallellisation to better take advantage of the dual FP32 units we've now got (but this is similar to the work they had to do to take advantage of Nvidia's equivalent dual FP32 in ampere).



It's done in the same revision of hardware as the changes to the resources, you don't wait for the next revision. Much of the changes to caches/registers/schedulers are there for that reason.



All GPUs manage workloads/utilisation in similar ways - AMD, Intel and Nvidia all use SIMD - and have done for many generations: you group stuff together and perform operations on groups of things at the same time. What they do differently, and what changes between generations, is how long those operations take and what things you can do at the same time.

CPUs on the other hand are traditionally not very SIMD - it's very much a GPU design philosophy, but of course, we now have greater SIMD on CPU these days.
Here's a short article I found. It's too short to fully address amd's approach but still..


The thing abt CPUs is i don't know the history but CPUs support very long vector operations like actually comparing 2 large vectors and then returning another large vector. Scheduling a cpu is more complicated than scheduling a GPU and Nvidia's dumb approach has a distinct advantage because their schedulers can be equally dumb
 
Associate
Joined
3 May 2021
Posts
1,232
Location
Italy
Considering the 6600 is a PCIe 4.0 only X8, is not that great for users with PCIe 3.0. motherboards. so... yeah, much "love" from AMD towards its customers.
The impact of PCIe 3.0 is quite limited unlike the 6500XT TBH. Remember that the target for this card is mostly people coming from the Polaris/Pascal cards, which will find it a reasonable improvement for the money.
Like I said, not exciting but if you're in need for a card NOW and your budget is limited... Those cards are the best pick from a disappointing price/performance basket.
 
Soldato
Joined
28 Sep 2014
Posts
3,464
Location
Scotland
Soldato
Joined
6 Feb 2019
Posts
17,831

Community Driven MorePowerTool For AMD GPUs Will Not Support RDNA 3 GPUs Due To Hard-Lock, Users To Pay For Power Limits & Features​


I'm not familiar with this software it didn't exist the last time I had a Radeon but if my understanding is correct this was basically the core overclocking software for previous RDNA cards right and without it users are limited to what AMD gives them which is extremely limited options?

After reading the article, the part where they say "users will have to pay for power limit and features" - sounds like a weird translation from German. What they mean is you're locked to how the RDNA3 card runs out of the box so if you want a card with better headroom and more power you have to buy one that has it out of the box, allowing AIBs to sell more premium models

I hope Nvidia doesn't copy this and lockdown tools like afterburner to force users to buy premium models
 
Last edited:
Soldato
Joined
18 Oct 2002
Posts
4,340
I hope Nvidia doesn't copy this and lockdown tools like afterburner to force users to buy premium models
It's been like this with Nvidia since the 1080ti (lack of voltage controls, power limit lower than it should be) but with the 4090 its even worse as the custom model's don't offer much beyond the FE.

Hence, EVGA quitting the scene.
 
Back
Top Bottom