AMD Radeon RX 7600 GPU Has Better Cache & VRAM Latency Than RX 7900 XTX

[T0@st]

Chips and Cheese published their very in-depth review of AMD's Radeon RX 7600 GPU last weekend - a team member (Jiray) took it upon themselves to actually buy the card, since a sample unit was not supplied for evaluation. The site's exploration of this graphics processing unit on an architectural level revealed a couple of positive aspects - which comes as a minor surprise since the Radeon RX 7600 received a generally lukewarm reception upon launch at the end of last month. Thanks to the Radeon RX 7600's Navi 33 XL GPU being a monolithic chip it seems to outpace—in terms of cache and memory latency performance—chiplet-based designs as featured in the vastly more powerful (and expensive) Radeon RX 7900-series cards.

Factoring in the smaller space that the RDNA 3 Navi 33 die occupies - it seems that it gains an advantage over the flagship card. Chips and Cheese reports that AMD's RX 7900 XTX takes up to 58% longer to access and pull data from its pool of Infinity Cache, when contrasted with the recently released sibling. The RX 7600 GPU exhibits 15% lower VRAM latencies compared to the RX 7900 XTX when retrieving data from the onboard GDDR6 VRAM chiplets. The review points to a greater disparity between current high-end and mid-range cards when looking back at equivalent models from the preceding generation: "The difference is especially large with RDNA 3. With RDNA 2, the RX 6900 XT had 151.57 ns of Infinity Cache latency compared to 130 ns on the RX 6600 XT, or a 16.5% latency penalty for the larger GPU." Chips and Cheese reckons that AMD's Navi 31's "chiplet configuration may be causing higher latency."



View at TechPowerUp Main Site | Source

 

[Ferrum Master]

And so what?

 

[oxrufiioxo]

Awesome!!! Shame it still sucks due to pricing.

 

[kawice]

hahaha, reminds me nvidia's article why memory bus does not matter anymore

 

[TheoneandonlyMrK]

Are these guy's tech reporter's, really, fire their arses.


I would like to announce something I found.

Point A to point B is considerably shorter then say,

Point A to B then C.

Even in the times ten or more years ago, before Vega 64 brought 2.5D GPU to the place this was largely known to be The Way.

or even J Huang , day one at 3Dfx I think it was could have told you that, wtaf.

Wtaf about the speed of light ,distances, and additional inter connection is hard to fathom here that someone thought this news.

So take f all time at all, then add 58%, and you still got f all time, that stuff will take to get to cache!?.

Someone let Dr Lisa Su know, she will be shocked.

 

[AnotherReader]

Are these guy's tech reporter's, really, fire their arses.

Chips and Cheese does micro benchmarking of CPUs and GPUs. The whole point is to measure things like execution latencies and cache latency and bandwidth. This can give a deeper understanding of why hardware behaves the way it does. Their deep dive into Intel's Arc shows why it usually performs much worse than it should.

 

[Icon Charlie]

Are these guy's tech reporter's, really, fire their arses.


I would like to announce something I found.

Point A to point B is considerably shorter then say,

Point A to B then C.

Even in the times ten or more years ago, before Vega 64 brought 2.5D GPU to the place this was largely known to be The Way.

or even J Huang , day one at 3Dfx I think it was could have told you that, wtaf.

Wtaf about the speed of light ,distances, and additional inter connection is hard to fathom here that someone thought this news.

So take f all time at all, then add 58%, and you still got f all time, that stuff will take to get to cache!?.

Someone let Dr Lisa Su know, she will be shocked.

I agree with this comment. These "Tech" reporters are awfully close if not already there to what we call... "Urinalists". Because Urinalists just love pissing into the wind.

 

[Jism]

Well duhh,

Chiplets have issues - overcoming latency for example. Testing this is like the obvious.

 

[AnotherReader]

Well duhh,

Try reading the rest of the article before dismissing their work.

 

[Jism]

Try reading the rest of the article before dismissing their work.

I'm not dismissing anyone's work, but it's generally known that chiplets due add latency rather then core2core etc.

AMD and Nvidia thus make different tradeoffs to reach the same performance level. A chiplet setup helps AMD use less die area in a leading process node than Nvidia, by putting their cache and memory controllers on separate 6 nm dies. In exchange, AMD has to pay for a more expensive packaging solution, because plain on-package traces would do poorly at handling the high bandwidth requirements of a GPU.

however the higher end card is obviously tweaked enough to overcome the latency issue. it's not like your going to measure any of that in raw performance.

 

[AnotherReader]

however the higher end card is obviously tweaked enough to overcome the latency issue. it's not like your going to measure any of that in raw performance.

GPUs are less latency sensitive than CPUs and the other cache levels haven't regressed in latency. On a memory related stall, they can just switch to a different wavefront.

 

[TheoneandonlyMrK]

Try reading the rest of the article before dismissing their work.

Not a bad shout that friend.

Good read, TPU sold them very short with they're news piece.

Though it really doesn't contain a new take, it's a in-depth look at the obvious, wrote well though, so I retract and Apologize for my harshness though the basic opinion remains.

 

[eidairaman1]

Not a bad shout that friend.

Good read, TPU sold them very short with they're news piece.

Though it really doesn't contain a new take, it's a in-depth look at the obvious, wrote well though, so I retract and Apologize for my harshness though the basic opinion remains.

I'm late to ryzen but if it's not already there it sounds like a clock generator/setting for IF may be needed, this is almost like hbm all over again...

 

[Minus Infinity]

Awesome!!! Shame it still sucks due to pricing.

It sucks on two levels. It's a rebadged 7500XT and it should be $199 as such.

 

[BoboOOZ]

Chips and Cheese does micro benchmarking of CPUs and GPUs. The whole point is to measure things like execution latencies and cache latency and bandwidth. This can give a deeper understanding of why hardware behaves the way it does. Their deep dive into Intel's Arc shows why it usually performs much worse than it should.

Yepp, these are very interesting results none the less. Basically, this means that chiplets might be great for having decent framerates in high resolutions, but might might be less interesting for low resolutions with insane framerates.
Edit: Looking at the comparison in gaming with the 4090, which is monolithic, I don't notice a tendency for the XTX to be slower (comparatively) than the 4090 in the lower resolutions, so there must be other compromises on the monolithic die, too.