AMD Execution Thread [2023]

[del42sa]

The usual large caveat that this is coming from Moore's Law Is Dead, but it's always been the obvious path for AMD, just a question of when they would finally take the plunge. A 256-bit LPDDR5 bus would finally alleviate (mostly) the APU bandwidth bottleneck.

A 40CU, 256-bit bus APU with IF cache is far more exciting to me than the next $1500+ GPU frankly, albeit of course depends on price - but powerful APU's could go a long way towards making PC's more price-competitive with consoles . The question is just if there's a market for them, but starting mobile first is definitely where there's a far greater chance for adoption rather than a 250 + watt desktop chip.

If this comes, would also love to see these sold like NUC kits outside of notebooks. You could have a system smaller than an Xbox Series S with 6700+xt performance, perfect secondary PC for the TV.

it´s coming from other leaker source too

https://twitter.com/x/status/1649661794128916480

but I agree, take it with a grain of salt , especially from MLID

PS: I would like to know, what difference is between RDNA3 and RDNA 3,5 (RDNA3+) as it´s only exlusively to APU´s iGPU

 

[TopSpoiler]

Samsung’s 4nm Chip Business Gets Small Boost As AMD Now Rumored To Shift Manufacturing Processes

Samsung may have address problems with its 4nm node, as AMD is now rumored to shift to this manufacturing process

wccftech.com

 

[DavidGraham]

AMD sales are down 64% for Clients in Q1 (down 1.4 billion $), Data Center is also flat, with no Q/Q growth at all.

https://twitter.com/x/status/1653519077275475968

 

[del42sa]

https://videocardz.com/newz/amd-con...uarter-rx-7950xt-spotted-in-rocm-pull-request

NAVI31 refresh ?

 

[yuri]

NAVI31 refresh ?

RDNA 4 is scheduled for 2024. This means the expected Navi 31 lifespan is ~1.5 years. So a "x50" upgrade is to be expected in a similar way the Navi 21 got one.

 

[techuse]

AMD could have beaten the 4090 at rasterization if they were willing to push the power high enough. They wouldn't come close in RT. That would require an entire rethink of their architecture. At the performance level of a 4090 RT is a huge selling feature. Their GPU would be DOA in the market. They still don't have a worthwhile DLSS 2 competitor. FSR 2 does not look good in the vast majority of rendering scenarios. This results in a market where AMD has to sell at a significantly lower price to be worth a purchase.

 

[del42sa]

 

[LordEC911]

AMD sales are down 64% for Clients in Q1 (down 1.4 billion $), Data Center is also flat, with no Q/Q growth at all.

https://twitter.com/x/status/1653519077275475968

Could have been a lot worse. The results honestly aren't that bad when you consider 1H '22 was an atypical high.
The real question is... is this the bottom or is Q2 even lower. AMD's outlook is flat, expecting a small bump in Q2 revenue due to seasonality... but a 5% miss would be bad.
Hopefully they moved a lot of mobile units in Q2, since this is why they shifted their focus. Hope Q3 and Q4 they can bounce back and break past $6b revenue again.

https://videocardz.com/newz/amd-con...uarter-rx-7950xt-spotted-in-rocm-pull-request

NAVI31 refresh ?

Makes sense, hopefully it is more than the 5-10% that the 6950XT had.
I wondered if that is why we haven't see Navi32, silly season had me speculating it was too close to 7900XT but this really isn't the place for that.

 

[del42sa]

AMD openSIL Planned to Replace AGESA Firmware in Client and Server in 2026

 

[DavidGraham]

The Next Platform is postulating about MI300 GPU performance based on a previous AMD statement.

MI300 vs H100

FP32: 96 vs 67
FP64: 96 vs 33
TF32 Matrix: 383 vs 495 (766 vs 990 with Sparsity)
FP16 Matrix: 766 vs 990 (1532 vs 1990 with Sparsity)
BF16 Matrix: 766 vs 990 (1532 vs 1990 with Sparsity)
FP8 Matrix: 1532 vs 1980 (3064 vs 3960 with Sparsity)
INT8 Matrix: 1532 vs 1980 (3064 vs 3960 with Sparsity)
INT4 Matrix: 3064 vs 3960 (6128 vs 7920 with Sparsity)

Measurements by AMD Performance Labs June 4, 2022, on current specification and/or estimation for estimated delivered FP8 floating point performance with structure sparsity supported for AMD Instinct MI300 vs. MI250X FP16 (306.4 estimated delivered TFLOPS based on 80% of peak theoretical floating-point performance). MI300 performance based on preliminary estimates and expectations. Final performance may vary.

AI Hype Will Drive Datacenter GPU Prices Sky High

UPDATED Like many HPC and AI system builders, we are impatient to see what the “Antares” Instinct MI300A hybrid CPU-GPU system on chip from AMD might look

www.nextplatform.com

 

[del42sa]

The Next Platform is postulating about MI300 GPU performance based on a previous AMD statement.

MI300 vs H100

FP32: 96 vs 67
FP64: 96 vs 33
TF32 Matrix: 383 vs 495 (766 vs 990 with Sparsity)
FP16 Matrix: 766 vs 990 (1532 vs 1990 with Sparsity)
BF16 Matrix: 766 vs 990 (1532 vs 1990 with Sparsity)
FP8 Matrix: 1532 vs 1980 (3064 vs 3960 with Sparsity)
INT8 Matrix: 1532 vs 1980 (3064 vs 3960 with Sparsity)
INT4 Matrix: 3064 vs 3960 (6128 vs 7920 with Sparsity)

AI Hype Will Drive Datacenter GPU Prices Sky High

UPDATED Like many HPC and AI system builders, we are impatient to see what the “Antares” Instinct MI300A hybrid CPU-GPU system on chip from AMD might look

www.nextplatform.com

they seem pretty close , will see in reality

 

[nAo]

How big is MI300?

 

[Kaotik]

How big is MI300?

About this size


It's apparently 9x'5nm' chiplets on top of 4x'6nm' chiplets + 128GB worth of HBM3.
At least default config includes 24 Zen4 cores. Amount of CDNA3 CUs is unknown. IIRC they also said exact chiplet config could be tailored to specific needs.

 

[del42sa]

https://twitter.com/x/status/1656783651554709506

The total market share of AMD in x86 processors, including IoT, embedded and consoles, achieved 34.6% in Q1 2023, according to Mercury Research, versus 31.3% in Q4 2022 and 27.7% a year ago. Intel has 65.4%

 

[Deleted member 2197]

AMD's FSR 3 to remain Exclusive to Company's Graphics Cards

In the wake of launching its RDNA 3 graphics cards - the Radeon RX 7000 series, AMD concurrently revealed the advancement of their FidelityFX Super Resolution (FSR) technology to the third iteration, ...

www.guru3d.com

Contrary to the frame production by NVIDIA DLSS 3's Frame Generator, an exclusive feature limited to GeForce RTX 40 GPUs, FSR 3 is projected to create up to four interpolated frames for each original frame. Moreover, it is anticipated that FSR 3 will be embedded within the drivers, thereby confining its operation to AMD's Radeon GPUs, excluding Intel and NVIDIA GPUs. This move, seemingly contrary to the technology's open-source nature unlike DLSS, is awaiting AMD's official confirmation.

 

[DavidGraham]

A driver side implementation of frame generation means no motion vectors which means very bad quality.

 

[Kaotik]

AMD's FSR 3 to remain Exclusive to Company's Graphics Cards

In the wake of launching its RDNA 3 graphics cards - the Radeon RX 7000 series, AMD concurrently revealed the advancement of their FidelityFX Super Resolution (FSR) technology to the third iteration, ...

www.guru3d.com

A driver side implementation of frame generation means no motion vectors which means very bad quality.

Seems to be drawing quite far fetching conclusions from just 2 lines of code:
uint32 frameGenRatio : 4; ///< Frame generation ratio
uint32 paceGeneratedFrame : 1; ///< Requires pacing the generated frames

 

[digitalwanderer]

The 7600 launch was badly handled by AMD imho and I'm trying to figure out why it became such a clusterfuck. It doesn't really matter since the 7600 itself sucks and is a bloody insult to the community.

The low-end GPU battle was right there for AMD to lose, and somehow they managed to.

Very disappointing launch, very disappointing product. To quote Kratos AMD, "Don't be sorry, be better"!

 

[Silent_Buddha]

The 7600 launch was badly handled by AMD imho and I'm trying to figure out why it became such a clusterfuck. It doesn't really matter since the 7600 itself sucks and is a bloody insult to the community.

The low-end GPU battle was right there for AMD to lose, and somehow they managed to.

Very disappointing launch, very disappointing product. To quote Kratos AMD, "Don't be sorry, be better"!

I'm not sure how they lost it? Sure we're still waiting for the 4060 to get benched, but so far it's this generation of video cards best perf/dollar video card with a larger uplift over the card it's replacing (6600) than the 4060 Ti was over the 3060 Ti.

So when looking at budget cards, you really want value for money (perf/dollar) and considering both AMD and NV are struggling to have meaningful performance gains at the budget level, it's exactly what someone would be looking for.

AMD Radeon RX 7600 Review - For 1080p Gamers

Today AMD is launching their Radeon RX 7600, which is targeted at gamers with 1080p Full HD screens. Our review confirms that thanks to the new RDNA3 graphics architecture, the new card is highly energy efficient.

www.techpowerup.com

Caveat, of course is if RT is super important in the budget space (I wouldn't touch any vendor's budget video card if RT was important full stop) then NV is the obvious way to go. For everything else, money is better spent on the 7600.

And considering that without a crypto boom, cards are likely to be well stocked, I certainly see AMD cards falling in price before NV cards just due to the halo effect of NV's 4090 on public perception. That should make it an even better value proposition for this generation of cards as that chart in the link shows.

I suppose a better way to think of it, is that as this generation of video cards as a whole (NV and AMD) are disappointing for anyone not in the market for a 1k+ video card, it's one of the brighter spots of a bad generation. In that a budget card is actually almost budget priced and brings decent performance for the price. Assuming sales for it are initially low, the actual market price may actually hit affordable budget pricing levels, something I doubt any NV card will do.

Regards,
SB

 

[DegustatoR]

It doesn't really matter since the 7600 itself sucks and is a bloody insult to the community.

7600 is okay, nothing "insulting" about it, some minor gains on previous SKUs in the same price range (not accounting for 2nd hand sales of course as that's a wholly different can of worms).

It is also technically interesting as an RDNA3 product which is almost a copy unit wise of an RDNA2 one - the gains there are unexpected in places, in both good and bad ways. CB.de did some comparisons here: https://www.computerbase.de/2023-05/amd-radeon-rx-7600-test/4/

but so far it's this generation of video cards best perf/dollar video card with a larger uplift over the card it's replacing (6600) than the 4060 Ti was over the 3060 Ti

The problem here is that AMD's pricing on 6000 series have deteriorated pretty far from what MSRPs they actually announced. You can get a new 6650XT ($400 launch MSRP) for less than $250 so the gains of 7600 are there only on paper not in practice.
This is a lesser problem for Nvidia as Ampere cards didn't really fall that far below their launch MSRPs in actual retail - cheapest 3060Tis are selling at ~$380.
So this comparison isn't as much in favor of 7600 over 4060Ti as one would assume from the paper specs. They are fairly close in perf/price gains to each other (4060Ti does have a bonus in DLSS3 though).