AMD Execution Thread [2023]

Status
Not open for further replies.

Good God what a fuckup. How did this make it even past the suggestion phase, let alone into a shipping driver?
 
Good God what a fuckup. How did this make it even past the suggestion phase, let alone into a shipping driver?
I am guessing AMD had no choice but to do this in this manner. Anti-Lag+ is so late in the game, and asking devs to directly integrate it into their games is going to take a very long time, most devs won't care anyway as AMD's marketshare is so low, and the tech is only available on RDNA3 GPUs anyway (even lower share). Contrast that with Reflex which has a 3 year head start, is supported on pretty much all NVIDIA GPUs and is featured in most AAA competitive titles as well as dozens of other titles. AMD had no choice but to do a driver side integration, with all the risks involved.
 
I am guessing AMD had no choice but to do this in this manner. Anti-Lag+ is so late in the game, and asking devs to directly integrate it into their games is going to take a very long time, most devs won't care anyway as AMD's marketshare is so low, and the tech is only available on RDNA3 GPUs anyway (even lower share). Contrast that with Reflex which has a 3 year head start, is supported on pretty much all NVIDIA GPUs and is featured in most AAA competitive titles as well as dozens of other titles. AMD had no choice but to do a driver side integration, with all the risks involved.
all pr is good pr? :runaway:
 
I am guessing AMD had no choice but to do this in this manner. Anti-Lag+ is so late in the game, and asking devs to directly integrate it into their games is going to take a very long time, most devs won't care anyway as AMD's marketshare is so low, and the tech is only available on RDNA3 GPUs anyway (even lower share). Contrast that with Reflex which has a 3 year head start, is supported on pretty much all NVIDIA GPUs and is featured in most AAA competitive titles as well as dozens of other titles. AMD had no choice but to do a driver side integration, with all the risks involved.

But it doesn't even sound like any consultation was done with Valve and others - my point is surely they had to have some understanding of the possibility this would trigger anticheat?
 
Threadripper and Threadripper pro, standard HEDT/workstation segmentation and feature fare - lots more io, memory (e.g. 4 vs 8 channels) etc at very high prices (workstatin prices TBD):

AnandTechCoresBase
Freq
Turbo
Freq
PCIe
(Gen 5)
Cache
(L3)
TDPDRAM
(RDIMM)
Price
($)
7980X64 / 1282500510048256 MB350W4 x DDR5-5200$4999
7970X32 / 643200510048128 MB350W4 x DDR5-5200$2499
7960X24 / 483200530048128 MB350W4 x DDR5-5200$1499

Anandtech

In slidedeck image 51 it says 266GB/s but the numbers don't add up - 5.2GT/s * 8 = 41.6GB/s * 8 channels = 332.8GB/s? That would mean 4156MT/s * 8 channels. Slide literally shows the working of 5.2*8*8 and the result is wrong
 
Threadripper and Threadripper pro, standard HEDT/workstation segmentation and feature fare - lots more io, memory (e.g. 4 vs 8 channels) etc at very high prices (workstatin prices TBD):

AnandTechCoresBase
Freq
Turbo
Freq
PCIe
(Gen 5)
Cache
(L3)
TDPDRAM
(RDIMM)
Price
($)
7980X64 / 1282500510048256 MB350W4 x DDR5-5200$4999
7970X32 / 643200510048128 MB350W4 x DDR5-5200$2499
7960X24 / 483200530048128 MB350W4 x DDR5-5200$1499

Anandtech

In slidedeck image 51 it says 266GB/s but the numbers don't add up - 5.2GT/s * 8 = 41.6GB/s * 8 channels = 332.8GB/s? That would mean 4156MT/s * 8 channels. Slide literally shows the working of 5.2*8*8 and the result is wrong
Gamer's nexus said AMD sent out two sets of slides. Anand likely has the wrong set.
 

Price difference like 50%. Performance on par.
Maybe it makes more sense to use settings where the dataset fits in the VRAM... Nobody is doing such kind of work on a consumer GPU.
Tomshardware looked at the improvements for inference:
We're not trying to make this the full Nvidia versus the world performance comparison, but updated testing of the RX 7900 XTX as an example tops out at around 18~19 images per minute for 512x512, and around five images per minute at 768x768. We're working on full testing of the AMD GPUs with the latest Automatic1111 DirectML branch, and we'll have an updated Stable Diffusion compendium once that's complete. Note also that Intel's Arc A770 manages 15.5 images/min at 512x512, and 4.7 images/min at 768x768.

4090 gets 75 images/s and 30 images/s...
 
Last edited:
In a rather sad news, AMD is massively downsizing their China team, with the Radeon Technology Group taking most of the hit.

The news comes from the Chinese media outlet ICsmart.cn, where they cite an individual within the company's social platform. It is disclosed that the expected layoff figure is around 450 employees, with a significant portion of it coming from the "Radeon Technologies Group" or RTG, responsible for the consumer GPU segment. While the source hasn't disclosed a specific reason behind the upcoming "layer" of wave-offs, it is being associated with the continuous decline in revenue for Team Red in all divisions, as well as in profit margins, which have been reduced to an "alarming" 97% compared to the previous year.
 
Status
Not open for further replies.
Back
Top