AMD Execution Thread [2025]

[Cyan]

Forget past RDNA 4 leaks, RX 9070 XT GPU is faster than rumors suggested, AMD exec tells us

In an interview with PC World, Frank Azor said that past performance leaks for RDNA 4 models have undersold the frame rates we'll get with these GPUs.

www.tweaktown.com

TL;DR: AMD's Frank Azor has revealed that RDNA 4 graphics cards will deliver faster frame rates than past rumors have suggested, and that Team Red is trying its hardest to make next-gen GPUs that everyone will be satisfied with in terms of the performance they deliver for their price tags.
AMD's Frank Azor has been on a roll with RDNA 4 (and Ryzen) related revelations this week, and the latest from the executive is an assertion that the next-gen graphics cards will be faster than leaks have primed us to expect.

Read more: https://www.tweaktown.com/news/1025...rumors-suggested-amd-exec-tells-us/index.html

 

[Kaotik]

Forget past RDNA 4 leaks, RX 9070 XT GPU is faster than rumors suggested, AMD exec tells us

In an interview with PC World, Frank Azor said that past performance leaks for RDNA 4 models have undersold the frame rates we'll get with these GPUs.

www.tweaktown.com

Also according to KitGuru, PowerColor said at CES that launch/mediadrivers will be about 20 % faster than what was at CES

 

[Seanspeed]

Forget past RDNA 4 leaks, RX 9070 XT GPU is faster than rumors suggested, AMD exec tells us

In an interview with PC World, Frank Azor said that past performance leaks for RDNA 4 models have undersold the frame rates we'll get with these GPUs.

www.tweaktown.com

Frank Azor's word is utterly worthless. I mean, we shouldn't take corporate marketing claims to mean a ton in the first place, his especially just shouldn't be taken seriously.

With 64CU, GDDR6 and a rough clockspeed estimate, we can have a decent idea of how it'll perform, and I think it's extremely unlikely to be surprising many paying attention. Somewhere between 7900 and 7900XT(or 4070Ti/4070Ti S in Nvidia terms) in general performance.

I know people want the 9070XT to be priced super aggressively, but if we think to how AMD used to do their more aggressive pricing, it wasn't necessarily the top SKU that had great value, but the slightly cut down version below it. And I'd be cool with that if they bring this back. They dont need a big win at every single product level, they simply need one in every bigger tier that sticks out and is easy for people to recommend.

I think something like 9070XT - $480, with a 9070 non-XT at $400 would be a winner. I mean, they'll probably drop down to such prices within like 6-12 months anyways, but the win will be a great first impression this time around, and something that will be solidified in reviews that people will still check later on.

 

[DavidGraham]

An amusing public exchange between an AMD software engineer "Anush" and the tiny corp on X, it spread like wild fire, below are the relevant bits ..

Anush:

Exactly. Code talks, rest walk. And if you fancy something in low level GPU programming AMD is hiring to power an Open AI ecosystem
aig-shark-hiring@amd.com

the tiny corp:

(@AMD@LisaSu For $1M and two boxes, we'll get the MI300X on the next MLPerf)
Still waiting on you to take our deal. A much better bang for your buck, even at the current $2M price

Anush:

Nah. All good. We can't take shortcuts or optimize for "bang for the buck" to greatness. But happy to get you access to mi300x to showcase the power of Tinygrad on AMD hardware and the power of open source

mayfer:

you totally could. they are writing drivers from scratch for you for free. there is no world in which this isn't a worthy bet to make for the cost of a few subpar employees for a year. insane risk-reward ratio that AMD ignores entirely out of pride

Anush:

That is broadly painting AMD employees as subpar - not cool. Not everyone has a big microphone. The AMD platform is open exactly to foster the
open source innovation that Tiny can do on the platform.

the tiny corp:

lol, we got 0 docs from AMD on the 7900XTX, we worked it out ourselves. It's as open as NVIDIA in the parts that matter (would have been the same effort to do what we did for NVIDIA), we just didn't see the need to rewrite their drivers.

the tiny corp:

We aren't interested in access, we're interested in two boxes shipped to us. I'm not a cloud fan. Why would you not want to most efficiently deploy capital to improve software?

Anush:

Is it the most efficient way to deploy capital? I have about 8 companies claiming they can do the same (not on X) and I can't ship them each 2 boxes since they are not a cloud fan.

the tiny corp:

We aren't making a claim. We already did it for the 7900XTX. git clone https://github.com/tinygrad/tinygrad.git
rmmod amdgpupython3
test/test_tiny.py
Show me something else close.

Anush:

Mad respect and kudos. But for now "cloud access" is still the best option to get you access to see what is possible.

the tiny corp:

If this isn't worth 200k to AMD to send us a few boxes, I don't know what to tell you. Cultural test to see if AMD can invest in software, if not why should I waste my development time on MI300X

Joe:

Anush: "Code talks, rest walk"
Tinygrad: *shows code*
Anush: That's nice but I don't *actually* want to invest in code, just to appear to

Anush:

Anush: "Provides access.."
Tinygrad: "No but send the boxes to me.. since I don't like cloud"
Anush: "Nah can't do. But still mad respect"

Mike64:

I will say that while NVIDIA GPUs also crash in certain cases where you have to do a physical reset, AMD is really excellent at that. Let me remind you of this debacle: https://github.com/tinygrad/7900xtx/blob/master/crash/driver.py
If you don't instantly understand why physical access is needed here, I don't know what to tell you. I could go around crashing nvidia gpus on cloud providers and no amount of infrastructure that likely exists there will be able to recover from that, likely requiring manually rebooting of the machine when the machine is discovered as crashed. People will be assigned pods and no idea why their GPUs don't work. Unless you are also willing to give out IPMI access, this is worthless and you know why.

Anush:

Yes I offered full baremetal via ipmi and bmc. This is how all developers access it

the tiny corp:

You missed my point. We aren't going to spend development resources on something AMD isn't willing to invest $200k in. We aren't going to set up other systems to manage running our CI in the AMD cloud.Our work isn't worth $200k to you. Messaged received.

Anush:

Your work is respected and valuable to invest in. Not sure what "systems to manage running your CI in the AMD cloud" entails but happy to work through it. Offer stands for full baremetal access (you can do any low level dev work) and if we hit some issues which require physical access we can arrange.

the tiny corp:

No thanks, offer that to the other 8 companies. Good luck!

https://twitter.com/x/status/1879571057012793480

 

[Granath]

I prefere this post

https://twitter.com/x/status/1880507210217750550

GH sounds to me like narcistic maniac.

 

[TopSpoiler]

Yeah, I agree that tinycorp is doing interesting work, but there is no reason why AMD should give away hardware for free for a 3rd party software stack other than ROCm. Tinycorp can buy hardware if they want by spending some of their $5M funding.

 

[trinibwoy]

Oh dear, he got his ass handed to him.

 

[Granath]

https://twitter.com/x/status/1880679135875362839

 

[Seanspeed]

GH sounds to me like narcistic maniac.

He absolutely is. And begging for free hardware over Twitter is not a great look(nor was begging Elon for a job over Twitter...).

At the same time, he's got a point. It really does seem like it should be worth the tiny bit of cost for AMD to just send them some hardware. Though any such dealings should all have been done through private channels.

 

[Granath]

He absolutely is. And begging for free hardware over Twitter is not a great look(nor was begging Elon for a job over Twitter...).

At the same time, he's got a point. It really does seem like it should be worth the tiny bit of cost for AMD to just send them some hardware. Though any such dealings should all have been done through private channels.

He got offer of free acces to cloud VM. That should be enough.

 

[Florin]

He got offer of free acces to cloud VM. That should be enough.

Not just a cloud VM, IPMI and BMC access means full control of the actual underlying hardware. It should be totally sufficient for any kind of development work.

 

[Cappuccino]

He absolutely is. And begging for free hardware over Twitter is not a great look(nor was begging Elon for a job over Twitter...).

At the same time, he's got a point. It really does seem like it should be worth the tiny bit of cost for AMD to just send them some hardware. Though any such dealings should all have been done through private channels.

I don't think any company of this size would give out free hardware to people best known for whining on twitter (yes I know who Hotz is). In fact I don't understand why AMD engages with him at all (and why random individual contributors are fighting with him on twitter? Unless this guy is in a major position of authority but it sounds like he's just a software developer).

 

[arandomguy]

I don't think any company of this size would give out free hardware to people best known for whining on twitter (yes I know who Hotz is)

Is this really any different than hardware reviewers? When some reviewers don't get supported they go to "whine on twitter" as well and hope for outrage interia to favor them.

 

[Cappuccino]

Is this really any different than hardware reviewers? When some reviewers don't get supported they go to "whine on twitter" as well and hope for outrage interia to favor them.

You misunderstood. I am saying George Hotz does little more than whine on twitter. Hardware reviewers review hardware, that is their 'product' and it ends up being marketing for the manufacturer anyways.

 

[Seanspeed]

I don't think any company of this size would give out free hardware to people best known for whining on twitter (yes I know who Hotz is). In fact I don't understand why AMD engages with him at all (and why random individual contributors are fighting with him on twitter? Unless this guy is in a major position of authority but it sounds like he's just a software developer).

Well yea I agree in that I think AMD is reasonable for not giving into a Twitter whine, but at the same time, it probably wouldn't be a bad idea to have had this conversation behind closed doors with them sending him some hardware. If AMD wants to become more relevant in AI and with open standards, they're gonna have to chip at it from as many angles as they can. And Hotz will hardly be the only petty egomaniac they deal with in the tech world.

 

[Cappuccino]

Well yea I agree in that I think AMD is reasonable for not giving into a Twitter whine, but at the same time, it probably wouldn't be a bad idea to have had this conversation behind closed doors with them sending him some hardware. If AMD wants to become more relevant in AI and with open standards, they're gonna have to chip at it from as many angles as they can. And Hotz will hardly be the only petty egomaniac they deal with in the tech world.

Dealing with people as unprofessional as Hotz would probably outweigh any good he would do.

If AMD wants to make inroads here they need to hire people to do it, outsourcing the creation of your platform to twitter addicts is not a recipe for success.

 

[Flappy Pannus]

Well we have a launch date for the 9000 series! It's...March

https://twitter.com/x/status/1881435903358628047

 

[Kaotik]

Well we have a launch date for the 9000 series! It's...March

https://twitter.com/x/status/1881435903358628047

Yep, it's been discussed here

 

[Sega_Model_4]

Not just a cloud VM, IPMI and BMC access means full control of the actual underlying hardware. It should be totally sufficient for any kind of development work.

I find it difficult to develop something so low level remotely. What if the kernel crashes all hardware?

I don't think any company of this size would give out free hardware to people best known for whining on twitter (yes I know who Hotz is). In fact I don't understand why AMD engages with him at all (and why random individual contributors are fighting with him on twitter? Unless this guy is in a major position of authority but it sounds like he's just a software developer).

About 2 years ago, George realized that Radeon GPUs could be great value for money as they were cheaper and very performant. But he noticed that AMD's software had a lot of bugs and decided to try to fix it, since much of the code was open source.
As he was very vocal, this caught AMD's attention, with Lisa Su stating that she would support it. However, that was just a promise, so George decided to ask for two things: that they open source the MSE and that they send them some MI300 so that he could develop everything more closely and test it directly on the hardware.

https://twitter.com/x/status/1669848494637735936

https://twitter.com/x/status/1871640579043647918

Lisa Su Says The "Team Is On It" After Tweet About Open-Source AMD GPU Firmware - Phoronix

www.phoronix.com

I think the biggest point of conflict is that George thinks AMD's software is bad and that they should throw it all away and replace it with something new. AMD, on the other hand, wants to maintain ROCm.
I personally think that AMD doesn't really know what it wants. Nvidia tends to support everything that is a startup and if necessary, they even give them the hardware, because the return is much greater than the cost of that hardware.

https://twitter.com/x/status/1879034330284192050

George stated that when he started his other company, Comma.AI, Jensen visited the company and gave them 4 Titan X.

Ilya Sutskever, one of the founders of OpenAI, received some GPUs from Nvidia too when he worked at AlexNet.

https://twitter.com/x/status/1868027046506410225

I think it’s cultural for each company. Jensen is a fanatic, he always takes a risk on something. AMD, on the other hand, adopts an IBM style, after creating something good on the market, they settle down and seek to make as much profit as possible.

 

[Kaotik]

George stated that when he started his other company, Comma.AI, Jensen visited the company and gave them 4 Titan X.

Ilya Sutskever, one of the founders of OpenAI, received some GPUs from Nvidia too when he worked at AlexNet.

You do realize there's a difference between couple consumer or even pro cards and two full datacenter servers + $1M, right?
Had Holz requested the former from AMD he no doubt would have gotten what he wanted and maybe more.