AMD - Zen chitchat

[kalniel]

Forgive the double post, but now I've had a chance to read the patent. In essence, not much new hardware needed at all, just a fixed function ray intersection engine, which is added into the texture pipeline.

Shaders (flexible function) are used to schedule stuff. They send instructions to the texture processor which is fairly fixed function - it performs BVH node look up (it's designed to be fast at look ups, that's a large part of texture processing) and passes that to the new fixed function intersection engine that's now part of the texture pipeline. Intersection stuff happens and results are passed back to the shaders which schedule what needs to happen next.

The BVH lookup/intersection is the slow bit in RT at the moment, so making that fixed function where it can be, and utilising fast lookup processors, is how AMD have got RT in there with little additional cost.

One concern might be falling back to shaders for the intersection stuff if you're not using functionally that's covered in the fixed function stuff. You're also now sharing texture processing and buffers/cache as posters have mentioned above, so if they're under strain then RT performance will take a hit too (or vice versa). Are AMD beefing up the texture processor/caches? I couldn't see it obviously.

[kompukare]

General purpose and useable not just for RT makes a lot of sense.
Note that there is some talk of ML in there, but in these days with phone SOCs have such a large percentage of their die dedicated to AI silicon that consoles don't is actually kind of surprising. I know both MS and Sony went for AMD and AMD are not that big in AI, but I'm sure they could have licenced some neural design thing.
Would a big differentiator between consoles and PCs (which devs would hate), but AI in games is usually the weakest feature so some kind of NPC AI framework would be nice.

[CAT-THE-FIFTH]

Anandtech have published the whole slide stack now:
https://www.anandtech.com/show/15994...cture-600pm-pt

The CPU runs at 3.8GHZ and the SOC uses PCI-E 4.0! There are some slides dedicated to texture compression,etc.

[kalniel]

General purpose and useable not just for RT makes a lot of sense.

Which isn't the route AMD have gone down really. They've gone fixed function for any kind of performance. There *is* flexibility, but at the cost of just doing it in shaders, which is effectively unaccelerated RT which GPUs have been able to do for years (but don't, because it's too slow).

Nvidia have separate hardware units for both the BVH look up and the intersection, I can't find anything about how flexible they are. If they're fixed function then it's basically the same as AMD, only AMD are using existing texture processors for the first bit and only adding on the second unit. Use of shaders for the scheduling is the same.

[CAT-THE-FIFTH]

Which isn't the route AMD have gone down really. They've gone fixed function for any kind of performance. There *is* flexibility, but at the cost of just doing it in shaders, which is effectively unaccelerated RT which GPUs have been able to do for years (but don't, because it's too slow).

Nvidia have separate hardware units for both the BVH look up and the intersection, I can't find anything about how flexible they are. If they're fixed function then it's basically the same as AMD, only AMD are using existing texture processors for the first bit and only adding on the second unit. Use of shaders for the scheduling is the same.

AMD's approach cuts down on die size I suspect. It also means in non-RT workloads more of the GPU isn't sitting idle.

Edit!!

If you read the slide,there is a lot of talk about die size optimisation,and how the features don't add massively to die area.

[Terbinator]

Sort of related, XSS/Lockhart CU count:
https://twitter.com/tomwarren/status...272141825?s=20

20CUs @ 1.55GHz.

Edit: I think Tweak mean less than 250W there, rather than more than.

$299 feels expensive for XSX w/o a disc drive if it is aiming for a certain customer. We know for Sony average LTV of a customer is $1600 spend on games - a store margin which increases with digital.

[kalniel]

Anandtech have published the whole slide stack now:
https://www.anandtech.com/show/15994...cture-600pm-pt

The CPU runs at 3.8GHZ and the SOC uses PCI-E 4.0! There are some slides dedicated to texture compression,etc.

High power requirements, needing them to design gaps around power delivery. But on the other hand, driver/firmware integration sounds a lot more efficient than PC.

[CAT-THE-FIFTH]

High power requirements, needing them to design gaps around power delivery. But on the other hand, driver/firmware integration sounds a lot more efficient than PC.

I am surprised the clockspeed is so high,especially as its a fixed clockspeed. Sure there is less L3 cache than the desktop CPUs,but the XBox has much more memory bandwidth too. Apparently the GPU portion of the SOC consumes under 140W which isn't bad for 52CUs running at around 1.8GHZ!

[kalniel]

I am surprised the clockspeed is so high,especially as its a fixed clockspeed. Sure there is less L3 cache than the desktop CPUs,but the XBox has much more memory bandwidth too

Massively high bandwidth!

[Terbinator]

Anandtech have published the whole slide stack now:
https://www.anandtech.com/show/15994...cture-600pm-pt

The CPU runs at 3.8GHZ and the SOC uses PCI-E 4.0! There are some slides dedicated to texture compression,etc.

Just coming back to this, interesting that they're only referring to the 3.8GHz speed now and also in the Q&A afterwards, whereas initially it was 3.8GHz for 8 core mode, 3.6GHz for 8c/16t.

Edit: NVM, slide 5 does say the speed different.

[CAT-THE-FIFTH]

Some interesting discussion on AT forums about the RDNA2 RT implementation(the post from uzzi38):
https://forums.anandtech.com/threads...#post-40255994

Massively high bandwidth!

It would be interesting to see if someone could run some CPU benchmarks compared to Renoir! The XBox does use a locked down version of Windows 10!

Just coming back to this, interesting that they're only referring to the 3.8GHz speed now and also in the Q&A afterwards, whereas initially it was 3.8GHz for 8 core mode, 3.6GHz for 8c/16t.

Edit: NVM, slide 5 does say the speed different.

So it seems the older news was correct.




Apparently what AMD fits in quite well with the inline RT of DX12 Ultimate.

[CAT-THE-FIFTH]

https://twitter.com/i/web/status/1299037620677287936

Navi driver PPTable
SMU_11_0_7_PMSETTING_POWER_LIMIT_QUIET
SMU_11_0_7_PMSETTING_POWER_LIMIT_BALANCE
SMU_11_0_7_PMSETTING_POWER_LIMIT_TURBO
SMU_11_0_7_PMSETTING_POWER_LIMIT_RAGE Rage mode

Navi21

PBO for GPUs??

[watercooled]

Naming a flashback to the ATI Rage series?

[CAT-THE-FIFTH]

Naming a flashback to the ATI Rage series?

I think the first two modes are a form of cTDP,and the second two a form of PBO for GPUs.

[kalniel]

I think the first two modes are a form of cTDP,and the second two a form of PBO for GPUs.

Don't AMD cards already have a power limit? Thought it was something you can alter in driver software. This just looks like preset power limits for 'EZ overclocking' or whatever they want to call it.

[CAT-THE-FIFTH]

Don't AMD cards already have a power limit? Thought it was something you can alter in driver software. This just looks like preset power limits for 'EZ overclocking' or whatever they want to call it.

Maybe,but it could also make AIB partner models more distinctive if they have better cooling.