Question Is using "Infinity Cache" for Navi 2X a good move?

[TheRookie]

AMD's previous flirting with exotic memory has been a disaster thanks to high prices and limited availability (think HBM2 in Vega and HBM in Fiji).

With Navi 2X using plain old GDDR6, AMD can get it from anywhere and inexpensively.

NVIDIA is stuck with getting relative expensive GDDR6X memory from Micron.

On the other hand, Infinity Cache has likely ballooned the size of the die and TSMC's 7nm is capacity constraint.

 

[igor_kavinski]

They have good experience with such caches from their Xbox partnership. Seems like a cost effective way to deliver great performance. Maybe they might provide a special driver to the media to turn off the infinity cache so its effect may be ascertained.

 

[Qwertilot]

Yes, but the XB stopped using it this generation of course.

I'd maybe echo the mild caution from the main AT preview. It's a different way to do things so have to test it.

That the consoles didn't use it is an indication that it definitely isn't an absolutely simple win.

Technically, it must obviously be theoretically possible to overload the cache - when their low raw bandwidth will presumably hurt them - but obviously no idea whether that'll ever happen in practice.

Support wise long term it really depends where they're going. If the cache ends up just in these big chips there's a long term worry about it on the historical lines.

If the other rdna2 chips all use something like it then no worry at all.

If it's a step towards future tiled GPU's then it'll obviously get built on & supported into the future.

 

[Dribble]

One would expect the cache to work worse the more memory is used, so if you have a game using 4gb of memory it's gonna work better then in a game using 16gb. It's also highly dependent on how the game accesses memory, it needs to fit in the cache, hence to use it best you'd need to optimise the game for the cache size. That's much easier to do when you are developing for a specific console then for a PC, and it's not like they can say "oh the consoles use it too so they will make games work with it" because they don't.

 

[Zoal]

The SRAM will scale down alot in the transition to the 5nm
The relatively small bus will help with laptop inclusion, APUs will benefit from internal memory bandwidth issues
It'll help alleviate bandwidth constraints with future multi-application/multi-gpu processing, RT-related requests
And could be a precursor to a GPU-chiplet design
(And made good use of Zen architects and their cache-design philosophy)

 

[Mopetar]

Support wise long term it really depends where they're going. If the cache ends up just in these big chips there's a long term worry about it on the historical lines.

Based on the die shots we've seen that have it labeled, it's pretty likely the whole lineup will have it, just in proportion to the number of memory controllers on the die.

We'll probably see 96 and 64 MB of infinity cache for the smaller Navi dies that corresponds with the 192 and 128 bit memory controllers.

 

[psolord]

One would expect the cache to work worse the more memory is used, so if you have a game using 4gb of memory it's gonna work better then in a game using 16gb. It's also highly dependent on how the game accesses memory, it needs to fit in the cache, hence to use it best you'd need to optimise the game for the cache size. That's much easier to do when you are developing for a specific console then for a PC, and it's not like they can say "oh the consoles use it too so they will make games work with it" because they don't.

The consoles do not have infinity cache? May I ask your source? Just asking, not doubting! thanks

 

[Kenmitch]

The only way out of the box is to think outside of the box. I'd imagine it would have been quickly scrapped if it wasn't the best solution to the problem. Maybe it's just a rev 1 thing which will be further implemented in future products?

 

[NTMBK]

They have good experience with such caches from their Xbox partnership. Seems like a cost effective way to deliver great performance. Maybe they might provide a special driver to the media to turn off the infinity cache so its effect may be ascertained.

The Xbox One didn't use the ESRAM as a cache- it was an explicitly managed memory pool.

 

[NTMBK]

Something to bear in mind... a single 4 byte per pixel 4K render target (i.e. a floating point buffer, or a BGRA8 one) is ~32MB. If you want to use a few different buffers (G buffers, Z buffers, whatever you renderer needs) then that's the whole cache used up right there. It's going to be interesting to see how it plays out.

 

[Tup3x]

Something to bear in mind... a single 4 byte per pixel 4K render target (i.e. a floating point buffer, or a BGRA8 one) is ~32MB. If you want to use a few different buffers (G buffers, Z buffers, whatever you renderer needs) then that's the whole cache used up right there. It's going to be interesting to see how it plays out.

What I'm interested to see is if it might backfire in FPS consistency. Average FPS numbers don't tell the whole story. In any case, interesting month ahead.

 

[CastleBravo]

I think "infinity cache" has added IPC/performance benefits beyond compensating for low vram bandwidth, so it might not cost much if any die space compared to a hypothetical non-"infinity cache" die that would need not only a larger memory controller, but also additional CUs in order to match Navi21's performance.

 

[itsmydamnation]

What I'm interested to see is if it might backfire in FPS consistency. Average FPS numbers don't tell the whole story. In any case, interesting month ahead.

why,

You could say the same thing about any other resource , if you didnt get good hit rates in the existing L0/L1/L2 today your performance would be trash as the throughput of the aggregate of the CU's a clock is far more then what can delivered "traditional "memory, let alone texturing and Raster. The the whole fear of getting good caching in existing cache structures but bad in a 128mb LLC is a bit over the top.

Im sure they have different policy options and controls as well.

 

[Bouowmx]

This generation of graphics architectures definitely had interesting choices. AMD RDNA 2 with very large caches vs NVIDIA Ampere (GeForce) with double FP32 throughput (without doubling associated caches, memory, etc.).

I am initially skeptical on why a GPU would need very large cache, as the parallelism of GPUs hides memory latency. The die area used could have been used for more cores instead.

 

[DJinPrime]

What I'm interested to see is if it might backfire in FPS consistency. Average FPS numbers don't tell the whole story. In any case, interesting month ahead.

That's my thoughts too. I'm not sure if majority of people really understand what cache is/means. Cache memory have been around a long time. Everyone knows the benefits, so why aren't caches super big? There's a down side to cache besides the die size cost. When you access cache, and if the data is not in the cache, then you have to go to the main memory (or the next level of caching and so forth). So, the cost is the cache access plus the regular access on misses. As NTMBK points out, 128 MB might sound like a lot but when your VRAM is 16GB and you have all those CUs processing data, you're going to have cache misses. On the miss, you're falling back to 256 bit GDDR6. In those situation, you actually have worse performance because of the cache. Will this cause lag spikes in game?

The "infinity cache" is definitely the most interesting part about big Navi. Can't wait for the white paper. Also, CPUs have cache hit rates of 9x%, so that's why you don't see crazy big cache on CPUs, because there's not that much more to gain. I couldn't find any solid hit rates numbers for GPU.

 

[itsmydamnation]

This generation of graphics architectures definitely had interesting choices. AMD RDNA 2 with very large caches vs NVIDIA Ampere (GeForce) with double FP32 throughput (without doubling associated caches, memory, etc.).

I am initially skeptical on why a GPU would need very large cache, as the parallelism of GPUs hides memory latency. The die area used could have been used for more cores instead.

The feature of rdna cuhas been to reduce latency to increase gaming performance. Thats the enite point of wave32.

 

[Hitman928]

That's my thoughts too. I'm not sure if majority of people really understand what cache is/means. Cache memory have been around a long time. Everyone knows the benefits, so why aren't caches super big? There's a down side to cache besides the die size cost. When you access cache, and if the data is not in the cache, then you have to go to the main memory (or the next level of caching and so forth). So, the cost is the cache access plus the regular access on misses. As NTMBK points out, 128 MB might sound like a lot but when your VRAM is 16GB and you have all those CUs processing data, you're going to have cache misses. On the miss, you're falling back to 256 bit GDDR6. In those situation, you actually have worse performance because of the cache. Will this cause lag spikes in game?

The "infinity cache" is definitely the most interesting part about big Navi. Can't wait for the white paper. Also, CPUs have cache hit rates of 9x%, so that's why you don't see crazy big cache on CPUs, because there's not that much more to gain. I couldn't find any solid hit rates numbers for GPU.

AMD gave infinity cache hit rate measurement in one of their footnotes.

Measuring 4k gaming average AMD Infinity Cache hit rates of 58% across top gaming titles. . .

AMD Unveils Next-Generation PC Gaming with AMD Radeon™ RX 6000 Series – Bringing Leadership 4K Resolution Performance to AAA Gaming

– Groundbreaking AMD RDNA™ 2 gaming architecture delivers up to 2X higher performance1 and up to 54 percent higher performance-per-watt compared to AMD...

www.globenewswire.com

 

[maddie]

This generation of graphics architectures definitely had interesting choices. AMD RDNA 2 with very large caches vs NVIDIA Ampere (GeForce) with double FP32 throughput (without doubling associated caches, memory, etc.).

I am initially skeptical on why a GPU would need very large cache, as the parallelism of GPUs hides memory latency. The die area used could have been used for more cores instead.

Try power consumption savings instead of focusing on latency as the only parameter.

 

[EliteRetard]

I think AMD was sort of forced to go with an oddball memory choice like this.
GDDR6X is a Nvidia joint venture, and they likely had exclusive rights to it.

I doubt AMD would have even known about GDDR6X development when making these GPUs.
For this performance class of GPU, they need a ton of bandwidth and GDDR isn't keeping up.
They either had to go with an extra wide memory bus, or come up with another solution.
There's a reason we haven't seen a 384/448/512bit bus in a long time.

 

[Guru]

This is a ultra fast ram essentially, its essentially sram, but better. I mean these type of technologies when implemented well can give insane advantages. Considering they have 128mb of the cache, its going to be extremely fast and powerful.

 

[AnandThenMan]

Why did AMD basically invent HBM, then end up with bog standard GDDR6? And isn't Nvidia about the only one using HBM now?

 

[EliteRetard]

Why did AMD basically invent HBM, then end up with bog standard GDDR6? And isn't Nvidia about the only one using HBM now?

HBM didn't really pan out for consumer GPUs, I think primarily cost issues.

Especially for the large and high speed stacks needed.

The cost of the interposer etc.

 

[Qwertilot]

AMD are presumably very likely to use HBM on their compute focused chip - CDNA I think they've called it?

 

[igor_kavinski]

Still, at least if they are charging $999 for a high end card, put HBM on there to justify the price premium. But I guess if nVidia doesn't feel the need to do it then AMD has no reason to decrease their profit margins. We need more competition. Wish Intel would join in already!

 

[EliteRetard]

Still, at least if they are charging $999 for a high end card, put HBM on there to justify the price premium. But I guess if nVidia doesn't feel the need to do it then AMD has no reason to decrease their profit margins. We need more competition. Wish Intel would join in already!

If AMD had used HBM, the 6900XT would have cost $1500.
16GB of that stuff would have added like $300 to the BOM.
It also would have restricted the 6800s to HBM as well.
You can't just switch GDDR for HBM on the same die.