AMD Execution Thread [2024]

[Kaotik]

https://www.techpowerup.com/319241/amd-zen-5c-ccds-made-on-more-advanced-3-nm-node-than-zen-5

AMD is reportedly building its upcoming "Zen 5" and "Zen 5c" CPU Core Dies (CCDs) on two different foundry nodes, a report by Chinese publication UDN, claims. The Zen 5 CCD powering the upcoming Ryzen "Granite Ridge" desktop processors, "Fire Range" mobile processors, and EPYC "Turin" server processors, will be reportedly built on the 4 nm EUV foundry node, a slightly more advanced node than the current 5 nm EUV the company is building "Zen 4" CCDs on. The "Zen 5c" CCD, or the chiplet with purely "Zen 5c" cores in a high density configuration; on the other hand, will be built on an even more advanced 3 nm EUV foundry node, the report says. Both CCDs will go into mass production in Q2-2024, with product launches expected across the second half of the year.

Assuming they continue the same pattern as with Zen 4/4c it's logical - Zen 5c CCD would have relatively less cache and thus benefits more from smaller node.
SRAM stops scaling at around 3-5nm class. TSMCs original N3B had still some scaling left in it, but the process overall had apparently some issues since TSMC has replaced it already with N3E which has exact same SRAM size as their N5 had
(Also we should all just stop talking about nanometers since they're literally just made up based on nothing. Last bit of relations "nanometers" had with the actual process size ended when we moved to FinFETs)

 

[del42sa]

https://wccftech.com/amd-medusa-ryzen-client-cpus-feature-zen-6-cpu-rdna-5-integrated-gpu-cores/

https://videocardz.com/newz/amd-medusa-zen6-client-cpus-rumored-to-feature-integrated-rdna5-graphics

* Zen 6 CPU Architecture (Expected 2nm)
* Utilizing Morpheus CPU Cores
* RDNA 5 GPU Architecture (Integrated)
* 2.5D Chip Interconnect (For CCD/IOD)
* Launch timeframe around late 2025 / early 2026

https://twitter.com/i/web/status/1759662011842240948

next APU "Medusa" use ZEN6 core with RDNA5 GPU

 

[fehu]

https://videocardz.com/newz/amd-medusa-zen6-client-cpus-rumored-to-feature-integrated-rdna5-graphics

Interesting points:
- no rdna4 apu
- zen6 with the new fanout interconnect
- maybe in late 2025

(edit: someway I missed the previous post )

 

[Seanspeed]

GPU architecture for these iGPU's probably doesn't make all that much difference really when it's still just gonna be a tiny handful of CU's.

 

[Kaotik]

GPU architecture for these iGPU's probably doesn't make all that much difference really when it's still just gonna be a tiny handful of CU's.

I don't see any reason why it would be more than the current 2 CUs, it's just meant to get image on your screen. And even though some sites are already calling for RDNA5 chiplets, that definitely isn't going to happen on dt parts. Mobile (aka APUs) maybe if they have to resort to chiplets at some point (outliers like Strix Halo aside, it could be chiplet I suppose), but 2 CUs or even few more can easily sit in IOD.

 

[Seanspeed]

I don't see any reason why it would be more than the current 2 CUs, it's just meant to get image on your screen. And even though some sites are already calling for RDNA5 chiplets, that definitely isn't going to happen on dt parts. Mobile (aka APUs) maybe if they have to resort to chiplets at some point (outliers like Strix Halo aside, it could be chiplet I suppose), but 2 CUs or even few more can easily sit in IOD.

Yep, pretty much. This could continue to be RDNA2 architecture for a while and it probably wouldn't make any difference to 99% of use cases. And if they're gonna keep putting it on the IOD, then you cant take advantage of some huge density increase to pack more CU's in, not that they'd likely want to anyways if they could choose instead to make the IOD smaller.

 

[Kaotik]

Yep, pretty much. This could continue to be RDNA2 architecture for a while and it probably wouldn't make any difference to 99% of use cases. And if they're gonna keep putting it on the IOD, then you cant take advantage of some huge density increase to pack more CU's in, not that they'd likely want to anyways if they could choose instead to make the IOD smaller.

Rather they'll keep IOD on as old process as possible since memory PHYs don't scale anyway and 2 CU is enough

 

[DavidGraham]

Raja on HBM pricing, and Vega.

https://twitter.com/i/web/status/1760516439730212995

 

[del42sa]

https://wccftech.com/amd-release-next-gen-mi400-ai-gpus-2025-mi300-refresh-planned-2024/

AMD Expected To Release Next-Gen MI400 AI GPUs By 2025, MI300 Refresh Planned As Well​

 

[arandomguy]

Raja on HBM pricing, and Vega.

https://twitter.com/i/web/status/1760516439730212995

In fairness here for HBM (and actually memory in general) capacity isn't the only metric but also performance and that byte/$ for memory in general has slowed drastically I believe ever since circa 2016.

I'm guessing the performance/$ ratio of HBM has gone up, just not the byte/$.

The upcoming GDDR7 might be similar in that performance/$ is likely going to improve much more so than byte/$ (if not go down as well).

 

[Bondrewd]

I'm guessing the performance/$ ratio of HBM has gone up, just not the byte/$.

I think 8-hi stacks still increment in bits per dollar.
Anything higher yeah.

https://wccftech.com/amd-release-next-gen-mi400-ai-gpus-2025-mi300-refresh-planned-2024/

AMD Expected To Release Next-Gen MI400 AI GPUs By 2025, MI300 Refresh Planned As Well​

Water wet yeah.

 

[pharma]

AMD Working on an AI-powered FSR Upscaling Algorithm

AMD CTO Mark Papermaster confirmed that AMD is working on a new upscaling technology that leverages AI. A key technological difference between AMD FSR and competing solutions NVIDIA DLSS and Intel XeSS, has been AMD's remarkable restraint in implementing AI in any part of the upscaler's...

www.techpowerup.com

 

[DavidGraham]

dGPUs market share in Q4 2023: NVIDIA 80%, AMD 19% and Intel 1%. AMD gained 2% Q/Q and 7% Y/Y.

AMD Gained 17% While NVIDIA Gained 4.7% Discrete AIB GPU Shipments In Q4 2023, Market Up By 32%

Both AMD & NVIDIA managed to increase their AIB GPU shipments in Q4 2023, marking a good end of the year for the discrete segment.

wccftech.com

 

[Silent_Buddha]

dGPUs market share in Q4 2023: NVIDIA 80%, AMD 19% and Intel 1%. AMD gained 2% Q/Q and 7% Y/Y.

AMD Gained 17% While NVIDIA Gained 4.7% Discrete AIB GPU Shipments In Q4 2023, Market Up By 32%

Both AMD & NVIDIA managed to increase their AIB GPU shipments in Q4 2023, marking a good end of the year for the discrete segment.

wccftech.com

Good news for the dGPU market as well as NV and AMD. Increased shipments of dGPUs across the board. Much better than some previous quarters where shipments decreased.

• AMD’s quarter-to-quarter total desktop AIB unit shipments increased 17% and increased 117% from last year.
• Nvidia’s quarter-to-quarter unit shipments increased 4.7% and increased 22.3% from last year. Nvidia continues to hold a dominant market share position at 80%.

Regards,
SB

 

[LordEC911]

Good news for the dGPU market as well as NV and AMD. Increased shipments of dGPUs across the board. Much better than some previous quarters where shipments decreased.

• AMD’s quarter-to-quarter total desktop AIB unit shipments increased 17% and increased 117% from last year.
• Nvidia’s quarter-to-quarter unit shipments increased 4.7% and increased 22.3% from last year. Nvidia continues to hold a dominant market share position at 80%.

Regards,
SB

Good point, though my initial thought is how much is driven by real consumers vs certain countries being sanctioned and certain markets/companies trying to scoop up certain products. On the other hand this has been the case for the last decade or so with the first crypto boom in 2012/2013 but the recent peak demand is just crazy. I guess it sorta averages itself out in the long run though.

Maybe I'm just too deep in the gaming/hardware enthusiast echo chamber but nobody really seems too excited/motivated about the GPU market right now. Other than Intel entering the market with everyone hoping they shake it up... both Nvidia and AMD have been business as usual with Nvidia pushing the pricing and AMD undercutting them.

 

[Bondrewd]

Maybe I'm just too deep in the gaming/hardware enthusiast echo chamber but nobody really seems too excited/motivated about the GPU market right now

Because your free lunch is over.
If you look at leaked Blackwell configs, stuff not 202 is tiny or none SM count bumps, which nets you vert incremental perf/$ gains.
And the market wants perf/$.

Other than Intel entering the market with everyone hoping they shake it up...

People want cheaper NV GPUs.
AMD's not driving NV pricing down aggressively anymore so the hopium is on Intel.

 

[trinibwoy]

Maybe I'm just too deep in the gaming/hardware enthusiast echo chamber but nobody really seems too excited/motivated about the GPU market right now.

I think the lack of excitement is due to diminishing returns from both hardware and software. Games look really good now and the improvements are incremental. Hardware upgrades just don’t improve the user experience as much as they did in the past. It’s also cause lots of enthusiasts are entitled whiners but I won’t get into that

 

[pharma]

AMD’s Lisa Su steps in to fix driver issues with GPUs in new TinyBox AI servers — firm calls for AMD to make its GPU firmware open source, points to issues with Radeon 7900 XTX

The intervention comes as Tiny Box publicly frets about Radeon-based platform bugs.

www.tomshardware.com

Tiny Corp’s dissatisfaction with its Radeon-based solution appeared to come to a head on Tuesday. The Tiny Box development team was initially happy to say they received updated firmware from AMD. They also congratulated the company for being responsive and making “big strides” on its drivers.

Things soon turned sour, though, as the updates from AMD apparently didn’t iron out all the important wrinkles. A follow-up Tweet on Tuesday again highlighted “serious driver issues” before pleading for open-source firmware. This became a recurring theme with tweets issued late Tuesday/early Wednesday morning. With pre-order customers already lined up and production beginning in earnest, the Tweets became increasingly desperate.

A furious thread of Tweets from the server startup opened with an incendiary blast about bugs, asking AMD to “fix their basic s*t.” Tiny Corp also complained that “we are not AMD’s QA team” before stating that the Radeon RX 7900 XTX driver is unfit for its customers.
...
Thanks to the unsubtle Tiny Corp, we can also be quite sure we will learn about the company's call with AMD later today, during which it appears the company will ask AMD to open source part of its firmware.

Tiny Corp is 70% confident that AMD will make at least some of its GPU firmware open source

Tiny Corp is kinda hopeful.

www.tomshardware.com

AI company Tiny Corp. says it is 70% confident that AMD will make some or all of its Radeon GPU firmware open source by the end of next week, according to an X post made earlier today. The somewhat hopeful prediction comes after Tiny Corp and AMD discussed the issues the former has been having with readying its upcoming RX 7900 XTX-powered TinyBox, which is intended to be used for AI work. Should AMD make the firmware open source, it would allow Tiny Corp. and others to fix bugs on their own, and potentially squeeze more performance out of AMD GPUs.

Call went pretty well.We are gating the commitment to 6x7900XTX on a public release of a roadmap to get the firmware open source. (and obviously the MLPerf training bug being fixed)We aren't open source purists, it doesn't matter to us if the HDCP stuff is open for example.… https://t.co/Zh6df2utIKMarch 6, 2024

 

[techuse]

Is Zen 5 going to be a tick or a tock?

 

[pharma]

AMD video upscaler arrives for RX 7000 GPU owners yearning for Nvidia's RTX Video

AMD users now have a video upscaler, but we aren't sure if its using AI upscaling.

www.tomshardware.com